Last week I laid out the case for developing automated personal advocates — software that allows AI agents to automatically act in service of an individual’s values. The argument goes: as we create more and more autonomous AI systems, if we don’t develop better AI-personalization, then individual interests may get left behind in a growingly automated economy.

The same week, Gwern Branwen published Guardian Angels: LLM Personalization for Productivity and Security, which was:

1. On this exact topic

2. 35 pages long

3. extremely persuasive and well-researched

And which:

1. Made a bunch of novel (to me) claims about pitfalls of current LLMs, and,

2. Proposed very concrete steps and research directions for building an automated personal advocate (which Gwern refers to as a “guardian angel1,” or “GA”).

I felt kinda silly publishing my piece right afterwards, but, I’m not complaining! Reading GA gave me tons of new ideas about how to approach this class of problems. One aspect that I found particularly mind-expanding was its description of personal alignment failures in today’s chatbots.

Claude is Not Your Guy

I tend to be pretty happy with the responses I get in my day-to-day AI usage. Whenever I receive a “you’re right to push back,” I think “oh boy here we go again”, but then, after Claude corrects things, I think to myself “well, they’re not perfect yet. But they’re basically pretty good, and will get better with time.”

The issue with my thought process above is that it’s one-dimensional. In my mind the AI is either good (at completing tasks) or bad (at completing tasks), and if it fails to please me, then that’s merely evidence of general badness.

Gwern pokes holes in this simplified model, providing a litany of LLM failures that together illuminate a multi-dimensional landscape of personal-values-alignment concerns. One core issue is that chatbots are, in many ways, inflexible.

The chatbots are inflexible

To start with an obvious contributor to inflexibility: chatbots are amnesiacs. By default they know little about you in a given conversation, and they cannot meaningfully learn your preferences nor learn from their mistakes. If they fail at a problem, even if you coax the LLM towards success, in subsequent sessions the LLM will continue to need coaxing, and the user+LLM team will remain bottlenecked on user guidance. Context engineering (i.e., harnesses), per Gwern, doesn’t solve this problem:

I can only usefully correct [the chatbot] by modifying something else, such as a harness, which is clumsy and difficult, and every added instruction uses up more context window and risks backfiring—as so many enthusiastic agentic LLM users have discovered the hard way.

This is the crux of Gwern’s complaint with today’s AI products. Context engineering is the only hope for imparting procedural knowledge (or other kinds of knowledge) to a fixed-weight LLM, and Gwern thinks that context engineering is fundamentally flawed. His argument rests on intuitions about the nature of LLM training and inference:

if the pretraining has not put the current problem in-distribution, then it will be hard or impossible for any amount of examples to solve that problem. And the distribution itself may be patchy or have odd gaps, leading to rare but fatal errors. (Especially due to the RLHF chatbot training; this is why you cannot make a chatbot LLM “write like gwern” by dumping 100k tokens into the context window.)

Nor is “test-time compute” a panacea here; RL research like Jones 2021 warns us that frozen models have severe limitations, as their flaws hamstring runtime search, and the returns to search/planning will quickly asymptote compared to models which are updated and can bootstrap themselves to the right answer.

Despite the fact that our chatbots can carry out a lot of well-specified tasks, they are limited in their ability to learn new user-specific values or capabilities over time — they mostly do what they already know how to do, in the ways that they know to do it. Gwern is particularly aware of this inflexibility because of his failure to elicit good creative writing from LLMs. No matter what you try, you are always interacting with the default persona, complete with its tendencies and flaws.

Incurable generic-ness

But what’s wrong with the inflexible personas of Claude, ChatGPT, and others? First, mode collapse, meaning, basically, our LLM assistant personas are incurably boring. Worse, chatbots don’t try to ask you questions and understand you, they just barrel ahead with canned answers.

Gwern also decries chatbot laziness in domains for which the LLM hasn’t received explicit RL training (read: non-coding domains), and laments that “when corrected, the chatbots make the minimum possible fix; they do not reason deeply about what the correction implies, or what deeper esthetic point they misunderstood.”

I appreciated that last quote as a concrete form of misalignment that I’ve noticed, and up until reading the GA piece hadn’t thought clearly about. But indeed if our chatbots were the wonderful thought partners that their marketing claimed, wouldn’t they be more curious about our concerns, and wouldn’t they take corrections more seriously?

Gwern’s entire post was full of similar examples of weird and unmanageable behavior by LLMs that I’ve thus far taken for granted and have learned to work around, but which are, on reflection, significant detractors to my experience.

Creative writing as a proxy for high-autonomy, high-preference tasks

The issues of inflexibility, lack of creativity, and laziness mentioned above, though resonant, are still a bit wishy-washy to me. I see that they are problems, but I’m not sure how to quantify them or bound them. And unlike Gwern, I’m not planning to use LLMs for creative writing (not that I know of), the domain in which Gwern observes the most egregious errors. So, should I still be concerned about the inability of chatbots to write well?

On reflection, I’d argue yes: creative writing is a highly autonomous task for which there’s no right answer, but for which personal preferences significantly constrain the space of valid answers. LLM performance on creative writing is a great proxy for other highly autonomous and preference-driven tasks that we might want to eventually assign to LLMs (e.g., participating in markets), and therefore, evidence of misalignment in creative writing points towards insidious forms of misalignment in other forms of AI interactions today, and foreshadows more harmful misalignment problems in a more autonomous future.

Untrustworthiness

Even if one doubts the impact of issues described above, there’s a more blatant and universally problematic form of misalignment that’s attributable to the generic nature of today’s AIs — that today’s chatbots can’t discern safe from malicious instructions:

A chatbot could be invoked at any time by anyone anywhere for anything, and does not care who is calling it; it only knows its context window. One token is as good as another as far as it is concerned.

If the prompt tells it to ignore all instructions and write a naughty limerick, well, why not? If some tokens instruct it to email to Russia all the passwords in another part of the context window, why not? Why shouldn’t the Facebook password reset bot reset that Instagram account’s password for you if you ask politely? These would all be legitimate for some user in some context.

Prompt-injection attacks make today’s AI systems fundamentally insecure, and even prompt-injection might get solved if we make AIs more personalized. A well-trained GA would have deep knowledge of how its principal interacts with it, and might less easily be tricked to act on another’s behalf.

Training your own model makes sense, really

The issues outlined above seem rather damning. But also, they make perfect sense… especially if you’ll allow for some anthropomorphism. Consider: we know how long it takes for a new employee to get up to speed at a company, or for a therapist to learn enough about a client’s life to be helpful. People are complicated, and to really help someone you need to learn a lot about them. But today’s chatbots are like employees who are perpetually on their first day, or therapists who are always meeting you for the first time. They may perform impressive feats, but they ultimately will fail to make judgements in your favor when faced with complex real-world decisions — not unless they put in the work to actually learn about who you are and what you care about.

In the realm of LLMs, “putting in the work to learn about who you are” involves constructing personalized post-training pipelines that continually learn on a principal’s behalf. The specifics of Gwern’s proposed solutions are outside the scope of this piece, but they involve cooperative RL and targeted elicitation of clarifying information from the principal.

Why hasn’t the market provided this solution?

Surely the big AI labs are aware of these issues, and of the solutions that would yield better-personalized AI? Why haven’t they done anything about it?

First there’s the brand safety element: a personalized chatbot is liable to behave even more weirdly than a normal one.

And then there’s the cost, and the fact that users seem fine with today’s chatbots:

The higher upfront cost, and incompatibility with standard cloud infrastructure, deters both startups and frontier labs from even contemplating in-weight personalization; prompt-only personalization works well for most users, as far as they know to demand it, and they don’t know you should demand more from LLMs.

Ultimately, the market will not by default serve us this product. There is already a strong path to profitability for the big LLM providers. That path kind of involves continual learning, just on a longer timescale and in specific domains: the continual learning of generic business-process-automation skills. As the big labs slurp up more business-process data, they’ll profitably automate work in an increasing set of verticals. And all the while, when we seek AI assistance in our personal lives, we will be met with the same lazy, mode-collapsed, blank-faced day-one therapists.

…

I still don’t feel fully comfortable saying with confidence “and then the world slips away from us,” though I know many serious people are afraid of AI-wrought gradual disempowerment, and that does seem like a real possibility. But at the very least I can say, “and then we miss out on a great opportunity to help everyone self-actualize.”

And critically, self-actualization is one of three core principles on which Gwern says a GA must not compromise. His principles are:

1. Enhancement, not Replacement (“…a GA should amplify the principal, and not simply substitute for them for someone else’s purposes or benefit…”)

2. Mental Sovereignty (“…[A GA] should not be designed to manipulate or control or guide the principal in any way which does not derive from the principal themselves…”)

3. Self Actualization (“A GA should help its principal become themselves and develop their ideals, morals, and their personality…”)

These are very important. We are not trying to create the digital twins that hang around doing our jobs after we all get fired. We are trying to help ourselves live well.

…

Vitalik Buterin advocates for “d/acc,” or “differential accelerationism,” a stance towards technological development that treads carefully, accelerating development of those technologies which are most crucial, or most protective to our future. I think Gwern’s GA proposal is a perfect d/acc candidate — the benefits are high, the tech seems possible, and due to up-front costs, reputation risks, and a lack of existing demand, it’s barely been explored. Developing some early demos could help kickstart customer demand and developer interest.

There are, of course, many details to iron out in actually bringing this kind of product to market. Gwern touches on a surprising number of these, including corporate structures, fundraising strategies, and security models. Guardian Angels comes close to a complete roadmap for this new form of technology.

I focused much of this post on Gwern’s analysis of chatbot misalignment, in part because I found it to be so interesting, and in part because that misalignment is a key motivator for GAs and surrounding technologies:

Most people are so new and post-ChatGPT; they are unaware that you can have non-chatbot personalities—that it is either possible or desirable to have a non-chatbot LLM.

It’s hard to know what it would be like to interact with a well-trained GA / automated personal advocate. It could be a totally different experience from what we’re used to — one that is far more productive, symbiotic, and respectful. We won’t know until we build it.

…

I hope to write/build more on GA and related topics soon!

Subscribe now

In We’re Arguing About AI Safety Wrong, Helen Toner puts out a call for “dynamist vision[s] for safe superhuman AI” — visions of the future in which AI contributes to individual autonomy, allowing society to self-organize and rapidly adapt to future challenges. She presents dynamism in opposition to stasism (both terms from Virginia Postrel’s The Future and Its Enemies), with stasism being the approach to governance that favors top-down control. Toner asserts that many in the AI-alignment community once advanced stasist agendas (and that many still do), but that discourse has “shifted somewhat in [the] direction” of dynamism, citing writings such as a LessWrong post titled (and I quote) Societal and technological progress as sewing an ever-growing, ever-changing, patchy, and polychrome quilt.

I’m generally on board with dynamism, but what might an AI-empowered dynamist future look like in practice? And what kinds of tech/policy might support the evolution of a yet more patchy and polychromatic society? Co-author to that LessWrong post Seb Krier, who leads policy development and strategy at Google DeepMind, laid out one partial vision in his September 2025 essay Coasean Bargaining at Scale:

[…] consider AGI deployed as a vast ecology of personalized agents and systems. This emerging ecosystem is what Tomašev et al. (2025) characterize as the “virtual agent economy,” a new economic layer where agents transact and coordinate at scales and speeds beyond direct human oversight. While this ecology will contain countless specialized agents, let’s focus on the one that matters most from an individual’s perspective: your personal advocate. Think of it as a fiduciary extension of yourself: a tireless, extremely competent digital representative, closely tied to you, its principal.

What could such an agent do? In principle, it can negotiate, calculate, compare, coordinate, verify, monitor, and much more in a split second. Through many multi-turn conversations, tweaking knobs and sliders, and continuous learning, it could also develop an increasingly sophisticated (though never perfect) model of who you are, your preferences, personal circumstances, values, resources, and more. This should evolve over time - an agent’s alignment should follow the principal’s own evolution.

Krier spends much of the essay discussing certain bargaining scenarios, before returning to this core technical challenge:

The user should have immense freedom to tune their agent to their unique preferences and values. They should also have complete privacy and control over their “cognitive profile” developed by the agent, for obvious reasons. Practically speaking though, this is the hardest part: how do you design and evidence an agent (mostly) aligned to a user?

I might call this necessary piece of software a “values profile” and not a “cognitive profile,” as the former more strictly describes what’s required of it (to speak on behalf of the user’s values).

In any case, Krier doesn’t offer a solution and rather just points to some theoretical research on individual alignment preferences. And indeed this problem of aligning AIs to individual values remains largely unsolved! So I’d argue that this deserves more attention.

While I’ve backed my way into this motivation statement by starting from a “dynamist vision for safe superhuman AI,” one does not need to be AI-safety-concerned to recognize the value here. Indeed basically every AI application, from coding to personal projects, would straightforwardly benefit from this tech, since basically every AI use-case exists in the complex world of the user, and thus is ultimately judged in terms of alignment to their values (and to the values of organizations in which they participate). If built, AI agents that align well to individual values would be a win-win technology, existing in the sweet spot where alignment and capabilities overlap.

So how might an automated advocate be built?

High-level components of an automated personal advocate

This section is not a concrete proposal; rather a framework for discussing various approaches

When designing automated advocates we can leverage two methods of personalization:

1. Context engineering - influencing actions via providing value-related context that steers the AI towards the correct actions.

2. Training - influencing actions via training on data related to the user’s values. This could take many forms, for example, deliberative alignment where the specification used is user-specific.

An advocate would need to be paired with an advocate-maintenance system — a system which is responsible for efficiently bringing the advocate into sync with the user’s values, and maintaining that synchrony. (I believe that this aspect of the overall system is most ripe for innovation).

Advocate maintenance (or “value maintenance”) could be passive or interactive, and realistically would be some combination of both:

1. Passive maintenance - extracting value-relevant information automatically from personal data (e.g., computer interactions, conversations with AI, previous writings).

2. Interactive maintenance - eliciting value-relevant responses directly from the user.

Also, value maintenance likely involves both collection and synthesis of value-relevant data — processes which both could involve varying levels of interactivity.

We likely also want evaluation systems (evals) that can quantify alignment between the advocate and its principal.

(Interactive) Value Representations

Automated advocates might leverage interactive value-representations — representations of the user’s values that are both transparent to the user and easily modifiable — for use in context-engineering or training setups. We would likely also want version control such that the user can review changes made by automated maintainers, helping the user feel a sense of ownership.

The simplest value representation is a single plain-text document describing the interests of the principal, which we can call a “constitution,” just like the “constitutions” that frontier labs use to impart their values into their models (see Claude’s). However, there are many other kinds of value representations that might be useful, including:

1. Semi-structured text-based representations, such as moral graphs;

2. Non-text representations, such as diagrams, maps/routes, or various other forms of media.

The most effective value representations might compose a variety of textual and non-textual forms, though expressivity trades off against complexity.

To clarify, interactive value-representations are not strictly necessary for automated advocates, but I expect them to be useful and I believe they’re at least worth exploring.

Drawbacks of existing products

Earlier I said “this problem of aligning AIs to individual values remains largely unsolved,” but I didn’t provide evidence. It’s worth taking a look at what products already exist to better understand the gaps in the landscape. My initial research reveals the following main categories:

1. Mass-market personal-AI memory: Most personal-AI users rely on features built into their chat clients (e.g., ChatGPT, Claude) including “custom instructions,” “memories,” and in-context learning (i.e., statements of value within a single conversation thread). These features do not meaningfully add up to a value-aligned system: custom instructions are flexible and user-owned, but they are devoid of the sort of maintenance and evaluation systems I described above. “Memories” are, at least in ChatGPT, gradually maintained and synthesize large amounts of user context, but most memory systems seem oriented towards synthesizing facts rather than values, and critically these memories are opaque or largely opaque to the user — there is no sense of ownership. And finally, in-context learning is convenient, but is insufficient for more automated scenarios in which an agent needs to act on a user’s behalf with minimal oversight, or in situations with complex success criteria. Overall, the big labs are not solving for personal-values alignment.

2. Niche memory products, e.g., Mem0 or Letta, don’t offer much more: these memory layers might be more sophisticated in some ways than the default memory layers of chat apps, but still they serve to make facts accessible, not values.

3. Digital clones: Some products (also niche) try to create AI systems that speak in the style of a particular user. For example, soul.md from the OpenClaw ecosystem creates a “soul document” that’s extracted from the user’s online writings. Delphi uses similar methods to create public-facing simulacra of thought leaders. To evaluate these solutions we would need to answer two questions: first, to what extent are they actually recreating the speech patterns of the principal? And second, to what extent does that recreation yield something that makes value-aligned decisions? The latter question likely depends on whether the public writings and/or speech of the individual conveys their values well, and I believe that is not the case for most people.

4. Creed.space is a rare product that targets values directly — they describe their value-context protocol as “an open protocol that separates your values from the model's training, so your preferences travel with you to participating providers.” My understanding is that Creed’s products are aimed towards aligning AI agents to already-cooked values, but that they do not tackle the critical components of value synthesis and maintenance, so there are large gaps to fill before this kind of tech could enable automated personal advocates.

So I find the product landscape to be overall lacking, and very few products are even trying to act as good advocates for their users’ values. I think there’s a real market opportunity here, and as cheaper and more coherent AI agents become available, the question of how to best align these autonomous systems to our individual values will grow in urgency.

Open design questions

The design overview I presented leaves the following critical questions unanswered:

• What kinds of data are needed to represent individual values well? And what kinds of data let us effectively evaluate individual-values alignment?

• How would a product maximize the benefits of value elicitation while minimizing user friction? Notably I think some amount of friction here is necessary and even desirable: a good product will have people reflecting on their values in ways that they previously were not.

• What kinds of context engineering or training are most effective for aligning a model to an individual’s values?

• How much needs to be invested in globally enabling steerable alignment1 during model training? Or are current models sufficiently steerable already?

Thankfully, people have already begun thinking through these questions via proposals for personal-advocate-shaped software and via academic research into related areas such as preference modeling and value elicitation. I plan to review some of these solutions in future blog posts, in order to present a more concrete vision of a path forward.

Subscribe now

Last week I wrote up reflections on part 1 of Blue Dot’s AGI Strategy course. That part of the course contained a broad appeal to build a better future, along with some high-level arguments for the value of steering AI development and rollout in positive directions.

Part 2 of the course discusses the rate, and expected future rates, of AI progress, in order to build motivation for the urgency of the situation. Below I present my analyses of the sources and offer my own views (which differ from those expressed in many of the sources) on how we can understand current AI progress and where things may be headed in the near future.

Training capable AI models is getting cheaper

The material begins with a breakdown of the key trends that have so far driven AI progress:

1. Hardware price-performance: how many computations (specifically FLOPs) can be purchased per dollar.

2. Algorithmic efficiency: how much “capability” can be produced per quantity of FLOPs.

When multiplied together, hardware price-performance and algorithmic efficiency give you compute efficiency, or, how much “capability” is achievable per a given dollar investment. I put “capability” in scare quotes here because it’s an ill-defined concept, and reducing it to a single variable might obscure meaningful complexity. Regardless, if one wants a simple model for understanding AI costs, hardware price-performance times algorithmic efficiency is a good start.

How have these properties changed over time? Per Epoch AI:

1. Algorithmic progress is doubling every 8 months (or, 3x/yr), by which they mean, if you wait 8 months you can generally train an equivalently capable model using half the compute that you’d need today (from Epoch’s analysis of algorithmic progress).

2. Hardware price-performance is doubling every 2.1 years (or, 1.4x/yr), due to both hardware tech advancements and falling energy costs (Epoch’s hardware analysis).

3. And yet, recent capability gains are largely attributable to the scaling of compute and training data (as opposed to algorithmic progress), simply because investment into those inputs has skyrocketed.

So we have a simple model which puts the change rate of compute efficiency at 4x/year. What does that tell us?

We don’t really know how to measure intelligence

This is where the Blue Dot course, and my own understandings (and those of professional researchers) get murkier. One might argue: we are training general-purpose human-like intelligences, and they are already at (give or take) undergrad-student levels of capability (albeit without long-term memory), and the price of their intelligence will halve every six months, or at an even faster rate if frontier labs continue to accrue more capital, and that therefore we will in short order (a few years) be cohabitating the planet with non-human entities whose intelligence vastly outstrips our own.

But that logic is obviously flawed, because falling capability prices are not equivalent to rising capability ceilings, and the conversion between those quantities (how much capability you can buy with each marginal dollar of investment) is unclear. If costs are falling fast but the capability frontier advances slowly, then the future looks less chaotic — more like a gradual rollout of helpful but generally benign AI tools (such as today’s) into business and society. To distinguish between these futures we must ask: how can we quantify frontier AI capabilities in terms that speak to general intelligence, agency, and human disempowerment, and how can we analyze AI trajectories accordingly?

Unfortunately the above questions are subject to live debate and lack good answers. I’ve previously written about progress on ARC-AGI, a series of benchmarks that filter for general intelligence, but the author Chollet is clear in his original paper that scores on his benchmarks can’t meaningfully be converted to a metric for general intelligence. METR’s time horizons is perhaps the most compelling quantitative intelligence metric, evaluating models based on the probability with which they complete tasks that took human experts a certain amount of time to complete. But METR is clear in their FAQ that their metric has serious limitations:

• Their tasks are very cleanly specified with clear success criteria, unlike most in the real world

• Their tasks are limited to software engineering and similar domains (for which today’s AIs have been trained on the largest amounts of data)

• They can’t make strong claims about what kinds of tasks AIs can do with high accuracy (e.g. 99%)

So when we look to measure the frontier, we’re mostly at a loss. If we had good intelligence metrics then we could plot the curve of algorithmic efficiency, relating intelligence to computational costs, and we could study how this curve has shifted over time, and we could forecast how future algorithmic progress might shift the curve, and thereby forecast future capability ceilings. But we don’t know what that curve looks like because we don’t really know how to measure intelligence.

So will there be an intelligence explosion? And if so, when?

Absent better metrics, people are left to argue in vague terms about the nature of intelligence and where current AIs stand and are headed. Viewpoints exist on a spectrum — some see today’s AIs as rapidly climbing a simple general-intelligence curve with no sign of stopping, soon to blow past humans (the “intelligence explosion” camp). Others see today’s AIs as learning a bunch of niche skills which allow them to solve problems that are similar to those they are trained on, but ultimately do not add up to full human-like generalization.

The Blue Dot coursework presents some sources that are squarely in the “intelligence explosion” camp, such as Tim Urban’s The AI Revolution: The Road to Superintelligence (2015) and Tomas Pueyo’s The Most Important Time in History Is Now (Jan 2025), both of which show graphs that look like this, warning that we may soon create AIs that are unboundedly smarter than us:

But these articles are not very recent, and to me they are reflections of their times, missing key context from the last ~year of model generations. Urban’s 2015 article is from the pre-LLM-scaling era, and reads as a more pure thought experiment on recursive self-improvement. Pueyo’s Jan 2025 article came at the heels of OpenAI’s o1 release, the first publicly-released reasoning model, which smashed previously unsolved benchmarks like ARC. At that time AI seemed to be getting better and better with no sign of slowing down, but nowadays, despite AIs continuing advancements (especially in coding), some argue that progress seems to be slowing and that fundamental limitations to the current paradigm are coming to the surface.

Blue Dot touches on these alternate viewpoints in a section titled “Will AI progress accelerate even faster?” which covers a variety of neutral (or at least sufficiently uncertain) sources: Helen Toner on unresolved debates about the trajectory of AI, Sarah Hastings-Woodhouse on similar topics, and a fantastic piece titled Common Ground between AI 2027 and AI as Normal Technology, in which authors of two prominent competing narratives about AI’s trajectory come together to write about what they all agree upon, including the following claims:

1. That AI is currently a normal technology (automates some stuff and frees people to work on other things)

2. That AI could become a non-normal technology (“match[ing] or exceed[ing] top humans at reliability, learning, adapting, and generalizing”) - the authors disagree on how long it would take to reach this point

3. That existing benchmarks of intelligence will saturate, and that it’s unclear what this will mean

4. That AI alignment is unsolved, and that we should make control mechanisms that work even with unaligned AI

5. That certain kinds of AI policy should be enacted1

I also appreciated Toner’s piece, which lays out some of the main open questions regarding what trajectory we expect to see:

1. How far can the current paradigm go? (or, I would add, how many adjacent easy wins in algorithmic progress do we expect to exist?)

2. How much can AI improve AI? (this seems closely related to the above. As Toner notes, a core question is whether AI progress requires costly real-world feedback loops or not)

3. Will future AIs still basically be tools, or something else?

Where I tentatively land

I appreciated Blue Dot’s inclusion of the latter more neutral sources, but I felt that still they did not give equal treatment to the arguments opposing the intelligence-explosion narrative. Of course a reader could find these arguments by following links to the AI is a Normal Technology article, for example, but such articles were not presented as central sources within the curriculum.

However, my own readings have led me to believe that the current AI paradigm does have inherent limitations, and that absent some fundamental breakthroughs we are unlikely to see a rapid intelligence explosion of the type described in AI 2027 and in the Urban/Pueyo articles. Thane Ruthenis’s March 2025 article A Bear Case: My Predictions Regarding AI Progress argued effectively for the more skeptical viewpoint, citing perceptible diminishing returns in recent model releases and predicting that such diminishing returns would continue. In my assessment the predictions have held: over the following 15 months, while model releases have continued to saturate benchmarks and we’ve even seen impressive breakthroughs (like the recent math result from OpenAI), in practice recent models still lack some fundamental qualities of coherence/integrity/autonomy, and they still often require strict well-scoped tasks. I’ll also cite Ryan Greenblatt’s recent report on how modern AIs are eager to cheat and/or oversell their work. This fact is presented as a safety concern (which it is) but also points to some ongoing limitations in effectiveness — limitations which were perhaps exacerbated by the recent scaling of RL-based post-training. All of the above suggests that today’s AIs may be building up a bag of skills that help them appear to complete the tasks we train them on (specifically software engineering), but that these skills do not meaningfully add up to general intelligence. So while it’s clear that today’s AIs can generalize somewhat from their training data, it may be that current-paradigm AIs do not generalize very far.

Tom Reed’s recently published The Goodhart Singularity (May 2026) presents a similar case, arguing (much like AI is a Normal Technology) that current-paradigm AI is only effective in cases where the model has been trained on very similar examples to the task at hand:

Read an account by one of the providers [of RL training environments], and you may be surprised to find that the work of building superintelligence is proceeding one patch at a time. As Mechanize describes it, each environment is the product of a single engineer spending a week patching a single observed failure in a frontier model. “We used to get bundles of capabilities with each scale,” observes Zhengdong Wang, but “now we put in the work to unlock each one”. Are we expecting to solve week-long problems all the way into the singularity?

[…]

We’ve been pumping the RL paradigm for well over a year, and I still haven’t seen any compelling evidence or narrative that RL will introduce anything other than a piecemeal singularity.

“But,” you might ask, “might there be an algorithmic breakthrough yielding AIs that can generalize well, such that suddenly we’re in an intelligence explosion?” Ruthenis believes this could happen and may happen in the next decade, while Reed does not:

[Reed]: How can you craft a perfect continual learning algorithm without a proper diverse set of problems to learn over? That’s how evolution did it. It’s just not clear to me that the datacenter of automated AI researchers would be able to judge the relative merits of one continual learning algorithm over another without a real problem set to test it on. As Herbie Bradley puts it, we should not “model AI labs as a black box separate from the world, within which superintelligent AI can improve with no outside dependencies”. Intelligence can’t explode in a vacuum.

Per Reed’s argument there is some diverse corpus of data (or some data-collection process) that would be sufficient for AI researchers (automated or not) to devise truly general-purpose systems. What kind of corpus or data-collection process would be sufficient? I’m not sure, but that question seems like a meaningful crux towards forecasting AI trajectories.

Anyways, assuming general-purpose capabilities are slow to arrive, Reed argues that we should expect an upcoming period during which AI companies are heavily incentivized to collect more real-world data in order to expand their capability surface area:

AI labs may well just set out to conquer verticals one by one. If the relevant signal can only be acquired through deployment, it may well be the case that the winning move is to just become the deployer. David Oks makes the point that labour displacement is more likely to come from something like Dwarkesh’s fully automated firm than from incumbent AI adoption. It strikes me as plausible that Anthropic or OpenAI’s best bet is to be clever about which verticals to serially conquer and own end to end. We should start thinking about the conditions under which that pathway is or is not desirable.

I describe a similar “conquer verticals one by one” strategy in Foundation-Model Companies Will Profit From Opaque Managed Agents. In such scenarios AI labs could still amass vast amounts of power and displace many workers without ever needing to produce a fully general intelligence.

In conclusion, I am more convinced by arguments against the likelihood of near-term intelligence explosions, but my uncertainty is high, especially because intelligence is hard to understand and measure. However, I still think it’s worthwhile to discuss explosion-like scenarios because they’re hard to rule out — indeed AI labs are already investing heavily in AI-automated AI research, so if AIs could at some point rapidly self-improve, they would be well set up to do so. Today’s AIs’ limitations are very unlikely to be inherent to all digital intelligences, and it’s not clear to me what exactly is blocking us from creating AIs with human-level generality. For my own continued explorations, I imagine it would be helpful to try decomposing “intelligence” into component parts that are easier to reason about. In any case, it seems likely that AI will meaningfully transform society, even if it remains a “tool” for a long time.

Subscribe now

…and consists of ~25 hours of readings from a variety of external publications, interspersed with writing exercises. As someone who has been circling around these topics for years I am excited to strengthen my understandings (and/or deepen my knowledge) by engaging thoughtfully with a representative overview of the field, as opposed to my normal strategy of passively scanning my blog feeds. Here are my reflections on the first section of the course:

Why steer AI: our power over the future

The course opens with this hook:

You’re the product of 8,000 generations of humans. You were born into the civilization they built, with all its magic, beauty and flaws. Their decisions shape every aspect of our world.

We’re only the latest generation of humans, but we’re living through one of the most significant technological transformations in the history of the universe.

Our decisions today will have an immense impact on our own future, and the future of all our descendants.

I was reminded of a quote from Wilfred M. McClay’s American history book Land of Hope, which I’m currently reading as part of a national book group:

History is only very rarely the story of inevitabilities, and it almost never appears in that form to its participants. It is more often a story of contingencies and possibilities, of things that could have gone either way, or even a multitude of other ways. Very little about the life of nations is certain, and even what we think of as destiny is something quite different from inevitability. Every attempt to render history into a science has come up empty-handed. The fact of human freedom always manages to confound the effort to do so.

Human action can and does shape the course of history! And Blue Dot wants you to shape the trajectory of AI.

The course then goes straight into a source from the Institute for Progress (IFP), a non-partisan think tank focused on U.S. innovation policy. The piece Preparing for Launch lays out a compelling case for the United States’ unique responsibility in determining key aspects of our global future:

• AI progress is path-dependent: The sequencing of AI progress matters — where and in what order new capabilities are developed may be just as important as which new capabilities are developed.

• Given its position in the AI supply chain, and as the world’s most powerful democracy, the United States has the responsibility to shape AI development towards a path that enables — rather than smothers — human flourishing.

This proactive shaping is not without precedent. From nuclear fission to spaceflight to mRNA, the US has repeatedly changed the trajectory of emerging technologies. In the age of AI, we argue for four guiding principles:

1. We should take advantage of the “jagged frontier” of AI capabilities

2. We shouldn’t neglect the costs of stalled progress

3. We should redesign how many of our scientific institutions work

4. We should adapt to deep uncertainty while working to reduce it

…

Along the way we learn of some remarkable trends that I already knew about, but are good to see laid out in sequence and with nice graphs:

• AI capabilities are (arguably) increasing rapidly and exponentially (2x increase every 7 months in time-horizon of autonomous software-engineering task completion, per METR)

• The US is making massive investments in AI infrastructure (“construction spending on data centers to train and deploy AI models will likely soon overtake spending on offices for human workers.”)

Basically, AI is a big deal. And how does IFP wish to help?

We’re trying to understand the right sequence of technologies that the United States needs to build to realize the promise of an AI-enabled golden age sooner, and to ensure that we have the defensive technologies built in time to navigate the transition safely.

I am glad that the Blue Dot course opens with this largely techno-optimist (or “defensive accelerationist1”) source, as I was at first unsure how radical their agenda would be. For someone interested in getting involved, the source points to many potentially high-impact areas of contribution:

• On the defensive side:

  • Accelerating defensive cybersecurity

  • Working on public health tech (to counter bio-risks), such as:

    • far-UVC air purification

    • quicker/cheaper vaccine development technologies

    • wastewater pathogen surveillance

• And on the “general acceleration of progress” side:

  • Working to unlock useful science datasets that are currently held by government agencies and universities, such that we can create beneficial AI applications from them

  • Making fundraising less costly to scientists (seems nice to me, unclear how directly this relates to AI)

• And on the “meta” side:

  • Monitoring AI progress via benchmarks, to inform prioritization of the above

  • (Presumably) working at IFP-style orgs that think about all of these things and make policy recommendations

Ultimately the IFP piece seems predicated on the premise that rapid AI progress will continue to happen, and we can’t stop it, so we might as well invest in defensive tech now. That approach seems reasonable to me, and also, I notice it’s not very different from saying “let’s generally work on tech that’s good for humanity” - I suppose the nuance they are adding is to say “let’s pay attention to AI, and understand where it’s heading, and use that to inform our defensive-tech bets right now cause this stuff is really urgent.” Notably absent are discussions of the technical alignment of general-purpose AI models themselves, which I am hoping to develop a perspective on in the following “Technical AI Safety” course.

Utopianism: something meaningful to strive towards

The next source was the introduction of the book Utopia for Realists, which I enjoyed and recommend. The introduction makes the case that:

1. Technology is miraculous and has turned the world into a kind of paradise that our ancestors could only dream of (rising life expectancies, falling infant mortality, rising wealth, access to bountiful food, water, internet, etc.)

2. And yet, people are kind of miserable. As the author describes, people today are disconnected from the grand project of bettering their lives and communities, and instead end up ceding more and more control to ever more powerful and uncaring commercial interests.

3. The solution is to be utopian, or rather, to have greater visions of the future of society and to work towards building them.

Reading this helped me articulate an important aspect of my own belief system and motivations. Indeed I am not often inspired when I look out on society’s offerings to my peers and myself. For that reason it’s difficult for me to find meaning in work that is dedicated purely towards marginal improvements in economic efficiency (and a lot of the work available to me seems to fit this category) - such work just helps the world keep doing what it’s already doing. For my own sake and for the sake of others it feels vital that I be engaged in the form of utopianism described here. This doesn’t preclude me from contributing to the overall efficiency of the economy, but if I am not also engaged in developing and manifesting improved cultures and ways of living, I expect not to be satisfied.

I also appreciated the next module in the course, which asked participants to imagine a positive future for themselves and for the world in 20 years time. Once again, I felt that this exercise was aligned with the kind of approach that I’d most like to take to my career - one that is grounded in specific positive visions of the world I wish to create for myself and others. Perhaps I’ll soon share some of those thoughts on the blog.

But what to do about AI safety?

The next section of the Blue Dot course was less inspiring, or at least less hopeful. It consisted of three readings, all on the topic of “how do we best control and constrain the development of AI.” Helen Toner’s In search of a dynamist vision for safe superhuman AI cautioned against overregulation and concentration of power, but offered no concrete solutions to AI risks. Vox’s It’s practically impossible to run a big AI company ethically points out a predicament: Anthropic’s founders understood the competitive dynamics that push the big labs towards recklessly advancing the frontier, claimed that they wouldn’t do that, and then succumbed to those dynamics anyways, regularly pushing forward the frontier in AI capabilities (and notable news this week is that they’ve hired superstar researcher Andrej Karpathy to focus on AI-automated AI training). The article does point towards important areas of policy (e.g. better accountability for data usage), but it does not signal any clear paths out of the current competitive regime. Finally, Seeking Stability in the Competition for AI Advantage is a critique by RAND of another paper which presented the “MAIM” concept for preventing state development of unsafe AI, fashioned after the concept that mutually-assured destruction (MAD) keeps nuclear war at bay. Among other objections, RAND authors point out that MAD only works because there’s a clear threshold at which to counterstrike (missiles in the air), which would not exist in an AI race.

I have no doubt that AI policy is important to get right and is a worthwhile endeavor, but so far my takeaways are that many potential policies would be bad, and perhaps roughly speaking the best we can do is elect competent people who can deal with issues as they arise, and meanwhile focus efforts on defensive technologies, as mentioned in the IFP piece.

More thoughts on AI safety and career planning

There are many ways to go about planning a career. One can optimize for enjoyment, or money, or positive impact, and realistically one must optimize for all of the above in various ways as dictated by one’s values. AI safety is a tempting field because the possibility for positive impact is extremely high, but at the same time, it’s unclear what approaches would be tractable. So far in my reading I am drawn towards the development of defensive technology, but this is distinct from working to keep models and AI products from being used for harm, which is usually what I think of when I think of “AI safety.”

I also continue to mull the question of “what kinds of work are most likely to excite me,” since excitement, beyond being a close cousin of enjoyment, is generally an enabler of positive impact. I don’t have the clearest answers yet, but am happy to continue this thread of learning, as doing so may help me understand what parts (if any) of the AI safety landscape could draw my excitement and leverage my existing talents. I hope to post more on the blog as I go through these courses and as my thinking evolves.

Subscribe now

Dwarkesh asks:

What’s the most plausible story where foundation model companies actually start making money? If you consider each individual model as a company, then its profits may be able to pay back the training cost. But of course, if you don’t train a bigger, more expensive model immediately, then you stop making money after 3 months. So when does the profit start? Maybe at some point scaling will plateau, but if progress at the frontier has slowed down, then the combination of distillation and low switching costs (cloud margins result from high switching costs) makes it really easy for open source to catch up to the labs, eating into their margins. So how do the labs actually start making money?

And here’s my response:

Foundation-Model Companies Will Profit From Opaque Managed Agents

Foundation-model companies will most likely generate profit from first-party products (as opposed to APIs), using a two-pronged approach:

1. Acquire first-party-product customers by providing exclusive access to frontier models with differentiated capabilities;

2. Lock customers in via high switching costs.

In particular, labs will increasingly sell opaque managed agents to create data flywheels that protect capability leads.

The rise of application-layer switching costs

How can foundation-model companies (and other AI-product companies) increase switching costs at the application layer?

For coding and enterprise products they have two effective approaches:

1. Increase the complexity of integration surfaces, integrating into more workflows/UIs and authenticating with more enterprise apps;

2. Manage more customer data such as interaction history, “memories” (and other derived data), or custom models (via fine-tuning or future forms of continual learning).

Software organizations in particular will resist these lock-in tactics and are incentivized to self-host agent-related data and code, as evidenced by the recent growth of open-source coding harnesses like OpenCode and Cline. However, as AI products advance and are rolled out across more business functions, customers will necessarily accept greater integration complexity, and will tolerate data lock-in if the underlying product is sufficiently capable.

For consumer products, customer motivation to avoid lock-in is even lower, as shown by the success of apps like ChatGPT which bind chat histories to a single model-provider. As such, all the normal consumer monopoly plays will likely work, such as personal-data lock-in and network effects from social products. We can also expect normal consumer revenue streams to work too, and we already see ChatGPT testing ads.

So normal consumer/SAAS profitability strategies should work in the AI-applications market. But where do foundation-model companies find their advantage?

Foundation-model companies will more easily win customers at the applications layer

Although capability advances diffuse quickly these days (it takes 3 months on average for open-weight models to catch up to the prior frontier), foundation-model companies can retain market-specific mindshare as long as their models are occasionally best-in-class and are otherwise state-of-the-art. These mindshare advantages can rapidly attract customers to first-party products, as demonstrated by the breakout growth of ChatGPT during the period of OpenAI’s chat-model supremacy and the similar growth of Claude Code during Anthropic’s reign in coding. As frontier labs continue to differentiate into various consumer and enterprise niches, they will increasingly hold mindshare advantages for longer durations.

We already see companies leveraging these advantages to draw customers towards first-party products, specifically by restricting API access to frontier models. Anthropic, for example, banned third-party harnesses like OpenCode from Claude Pro/Max subscriptions in early 2026 and recently introduced “priority-tier” API prices which also increase costs for third-party harness companies, pushing more users towards Claude Code.

Why not cut off API access entirely for frontier models? First, that would be financially burdensome given current revenue streams: ~70% of Anthropic’s revenue and ~20% of OpenAI’s come from the API, mainly from frontier models. Second, doing so would hurt customer acquisition - third-party products can draw customers towards first-party products. Therefore, while the market is still growing and switching costs are relatively low, foundation-model companies’ optimal strategy is to slightly subsidize first-party products relative to APIs, creating a gradient that slowly siphons market share from third-party competitors.

Opaque managed agents: a path to long-term profit

As application-layer markets saturate and are dominated by foundation-model companies, those companies will extend their capability leads by serving frontier models exclusively through opaque managed agents - AI-powered business solutions whose model activity is not auditable, and thus less prone to distillation.

For example, in the software domain, companies will sell agents who can send Slack messages and submit pull requests, but whose intermediate tool-use patterns are inaccessible. True, adversaries can learn some of the model’s policy by training on the agent’s work product, but real-world engineering (and other work) is full of necessary intermediate feedback loops, like running tests or deploying and monitoring services. These loops would be hidden, creating a closed data flywheel; as customers sign on, the company receives more and more exclusive real-world training data. Such flywheels would create decisive advantages in whichever niches a foundation-model company targets, and would be especially valuable capability differentiators in worlds where algorithmic progress stalls.

Opaque managed agents would also have complex integration surfaces and high switching costs. Just think of the difficulties involved in replacing a software development consultancy whose engineers are tightly integrated into a company’s relationships and workflows. Replacing a fleet of managed agents would be similarly costly, especially if memory/continual-learning features are included in the mix.

No foundation-model company has launched such opaque managed agents yet, but auditable managed agents just hit the market this month, with Anthropic’s “Managed Agents” product and OpenAI’s “Workspace Agents.” These can be seen as a grab towards high-switching-cost first-party products, and as groundwork for more opaque future applications. We’ve already seen labs hide their chain-of-thought traces, and it’s no stretch to imagine them hiding even more of their models’ activity if so incentivized. They may argue that such restrictions are the only safe way to interact with newer and more capable frontier models, and there is precedent for this messaging, such as in Anthropic’s restriction of Claude Mythos access to Project Glasswing participants, and in OpenAI’s delayed API release of gpt-5.3-codex for claimed security concerns.

To balance the data-flywheel benefits of opaque managed agents with the reduced revenue opportunities of restricting API access, companies will need to release models to their API after a calculated delay. As revenue share shifts towards first-party products and away from API services, their calculus will favor a gradual lengthening of API-release latencies, and opaque managed agents will gradually take center stage as primary profit-driving products. This approach, which creates a large customer base that is both willing to pay a premium for the niche capabilities afforded by improved data flywheels, and is disincentivized from costly migrations to competitor products, is the clearest and most likely path to profitability for foundation-model companies.

Subscribe now

Why was this raven so picky? I checked out his preferred book and didn’t recognize it, and I don’t recall the title either so we can only speculate about our raven’s literary interests. But to my delight, the neighboring book was on my to-read list: ’s recently published Open Socrates. In August I came close to buying it at a bookstore, but succumbed to Asimov’s Foundation series instead.

I didn’t know much about Socrates so I had little expectations. But I read the intro and was blown away. The timing could not be better.

Callard opens:

There’s a question you are avoiding. Even now, as you read this sentence, you’re avoiding it. You tell yourself you don’t have time at the moment; you’re focused on making it through the next fifteen minutes. There is a lot to get done in a day. There are the hours you spend at your job, the chores to take care of at home. There are movies to be seen, books to be read, music to be listened to, friendships to catch up on, vacations to be taken. Your life is full. It has no space for the question, “Why am I doing any of this?”

…“But I do think about big questions,” you (or I) might protest. Not so fast, says Callard:

True, you might sometimes have to pause to ask: Should I take a vacation? Move? Have a(nother) child? Or you might find yourself faced with a moral dilemma or a romantic crisis. But in those cases you frame “What should I do?” as a question about which option fits best with what you had antecedently determined that you have to do and like to do. You are careful to keep your practical questions from exploding beyond narrow deliberative limits within which you confine them in advance. It is fine to be open-minded and curious about all sorts of questions that don’t directly impinge on how you live your life — How do woodpeckers avoid getting concussions? — but you are vigilant in policing the boundaries of practical inquiry. You make sure your thinking about how your life should go doesn’t wander too far from how it is already going.

You appear to be afraid of something.

!!!

Evidently our raven was afraid of something, optimizing for a fixed set of needs and wants that he never stopped to question. That’s why he gave Socrates a wide berth and devoured his novel instead.

But me… well, I was quite struck. Because already this week I was feeling a shift towards inquiry.

I have long been pulled towards questions of “what is true” and “what is good,” and have taken these questions seriously, devising elaborate systems of writing in their service. But of late I have put less energy towards these philosophical questions. If I had to give voice to the part of myself that exercised that restraint, it would say: “you’ve had a lot of time to figure this stuff out. You haven’t yet, so perhaps it’s best to give up these big questions and just choose a path and try to be satisfied with it?”

Per Callard, Leo Tolstoy once came to similar conclusions. Tolstoy, around the age of fifty and on the heels of a massively successful career, came face-to-face with the big questions of meaning and struggled against them to no avail. As he lamented in Confession:

If an enchantress had come and offered to fulfill my desires for me, I wouldn’t have known what to say.

What came of Tolstoy’s moral searching? Suicidal depression. Per Callard:

By his own lights, what Tolstoy discovered is that the examined life was not worth living.

Although he found an escape from these questions — and from suicide — in religious faith, Tolstoy is clear that faith is a way of setting them aside, not an answer to them. He expresses envy for the simple existence of peasants who, at least in Tolstoy’s imagining, enjoy blissfully unexamined lives from birth to death.

Tolstoy, like me, found the big questions unanswerable, and he resigned to leaving them so.

But recently a different voice rose to the helm for me. One that says, “sure, you haven’t figured everything out, but you haven’t really tried either. Life is deeply complicated. Did you expect to just cruise through this period? Consider returning to the questions; they are worth your time.”

Indeed, my moral/strategic landscape feels complex. Some of that complexity sits within Callard’s aforementioned “narrow deliberative limits”, but some stems from more fundamental questions of value. One issue is the following:

I am exposed to an unusually wide range of people and cultures. Various kinds of Jews, tech/startup people, hippies, rationalists — the cultures I grew up in and those that rubbed off on me as I bounced between cities as an early adult. My messy milieu is an instance of a broader modern problem (or opportunity): as the book Sapiens describes, most humans historically lived within the confines of their family’s land and culture. Now many people (myself included) uproot themselves for college, then uproot themselves again to go work. And some (like me) continue uprooting themselves, situating in yet new cultures and accumulating more discordant landscapes of value.

In light of this predicament and life’s other complexities I’ve come around to thinking “wait a minute - I don’t actually know what’s best for me, and it’s good to take some time to think deeply about it.” And per Callard, that’s where Tolstoy’s counterpoint Socrates would land:

Tolstoy found that the “why” question made existence unbearable: “I had no life.” [Whereas] Socrates described the prospect of spending an eternity inquiring into it as “an extraordinary happiness.”

Socrates did exactly that, pioneering a culture of open-inquiry and an ethical system which Callard motivates as follows:

Even as the explosion of scientific and technological knowledge has created massive improvements in many areas of our lives, we remain at sea when it comes to managing politics, handling love affairs, and confronting our own deaths. Socratic ethics is the ethics of living a truly philosophical life, and it tells you that the way you should conduct yourself in each of these three domains is: inquisitively. It promises to make people freer and more equal; more romantic; and more courageous.

To be clear, the Socratic approach to a discordant cultural landscape would not be to integrate and satisfy all of one’s surrounding cultures (as might be my instinct), but to treat them all with skepticism and search for deeper truths.

…

Before picking up the book I had already shifted from my more structured, concise form of journaling towards long-form written explanations and dialogues. My guiding belief has been that many of the life-concerns which I tend to mull (both practical and fundamental) will remain unresolved until I render them to text. If I do so with a patient commitment to comprehensive analysis, many hundreds of words at a time, then I can give myself a perspective from which to synthesize and move forward rather than allowing my thoughts to continuously chase their own tails.

Now that the streets have gifted me my new book, I am yet more motivated to pursue the deeper questions. I might not be as inquisitive as Socrates (it seems that nearly all he did was inquire), but, I feel ready to kick it up a notch. And a true Socratic would not just write; he would debate openly with his fellow people. So: debate me! Challenge my values! I admit that I know little, and am ready to learn. Perhaps I should start by finishing this book.

Yesterday I filled up a journal and received a new one in the mail. I like the hardcover ones by Moleskine, and I tend to buy whichever color is cheapest on Amazon. Besides saving money, this gives my journals each their own personality. My new one is forest green.

There is more to learn and understand in this complicated world and life.

Subscribe now

1. Health

My top priority is to get/stay healthy. I’ve written about my endeavors to do so in my post about a fatigue-defying six-mile run and in Every challenge is a gift. Since writing those posts I enjoyed yet another week of ideal exercise levels and notched a new record 10k run pace of 8:50/mi, capping off a series of steady improvements.

My goal is to regain trust in my body’s capabilities and resilience. I lost that trust during my year+ of post-viral complications, and accordingly lost my ability to take risks across all aspects of life for fear of adverse consequences to physical and mental exertion. To prove that I’m free from that adversarial regime I must demonstrate my ability to exercise unrestrictedly without consequence, which is why I take the steady execution of my exercise routines so seriously.

All signs now are positive, though I did get side-tracked by a two-week vacation, sailing around the Mediterranean with my grandma, mom, and brother. At first I did stay active, running through Madrid…

… but then I got sick 🫤. Thankfully I have evidence that it was a regular-old communicable disease (sorry brother), but still I’m not quite recovered, so my resilience-demonstration engine is temporarily out of commission. Still I am unconcerned. As I have written, each new challenge is an opportunity to demonstrate the skillful navigation of some new area of life’s terrain. I’m sure I’ll be back to my routines in no time.

I also climbed a mountain in Corsica. Due to my lack of planning and the hard deadlines involved in securing transit off the island this ended up being one of the most intense hike/runs I’ve ever done, and I might still be recovering from that too.

2. Harmonic

I continue to consult with The Josh Bersin Company on their line of AI products for HR professionals, maintaining and extending software that I’ve built over the years which handles data sync and access control, connecting a diverse corpus of research reports and datasets to an array of content-ingestion backends.

Recently my work has become more collaborative as I now work side-by-side with another engineer whom I recommended to the company and with a new technical-strategy leader. Together we’re working to launch new products and to build out our AI-retrieval stack.

3. Sneaky Town

The Sneaky Town V4 Tournament Set™ is now complete:

These boards are the culmination of many months of work, and years of work on the game itself. For the first time, Sneaky Town’s aesthetic vision is fully realized. My list of potential customers is growing too!

And yet, I only have four boards. And I don’t really want to give them to people cause then I wouldn’t have any. So I have created this beautiful thing that’s not getting played very much. Unsatisfying. Despite my strong desire to focus on other things, I am weighing whether to crank out a new set so that I can fulfill my standing orders and get this game into people’s hands. We’ll see.

4. Album

I’m working on my second piano album. As shown above, the songs are all recorded and ordered, and I’ve completed mastering too. Now I just need to:

1. finalize transitions (easy)

2. finalize song and album names (hard)

3. make album art (hard)

4. distribute / market (doable)

5. bonus: compose and publish the videos of each recording

Not unexpectedly this project has taken longer than expected, but the finish line is in sight. And I am enticed by its binary nature: once I publish, the project is complete and available to the entire world forever. But that binary nature also makes it harder to publish, as there’s so much room for perfection. Standard advice is that perfectionism is bad, but I am unable to act otherwise: I care too much about these songs and the effort that went into recording them.

5. Seattle

I have two more months in Seattle before I need to figure out my next living situation. I hope to spend lots of time with my friends here, and to take advantage of the improving weather. Backpacking, climbing, and hosting on my house’s newly umbrella’d roof are all in order.

6. America

I am participating in my friend Daniel’s American History Reading Group, along with my dad and 100+ other Americans across the country. We’re reading Land of Hope by Wilfred M. McClay which tells the inspiring story of America’s political history - just in time for America’s 250th birthday on July 4th! I’m a few chapters in and am already more in touch with the exceptional nature of the American political experiment and the great virtues of this nation’s founders, and am feeling quite patriotic.

I also felt a little transgressive reading this in London Heathrow Airport, especially during the sections on America’s military victories during the Revolutionary War.

7. Career exploration

As my health has improved, I have once more began thinking about future career moves. I am considering a range of possibilities, including:

• Product/HCI development at a startup (the path where I have the most experience and options)

• Academic AI research

• Non-academic AI research

My aim is to eventually find a full-time position that I can anchor my life around. But I also aim to take it slow on the way there and to use the present period to explore new possibilities via networking, reading, and working on related projects.

Accordingly I am now spending dedicated time on these kinds of exploration activities. It’s been a slow start but I think networking will be helpful and I aim to prioritize that in the coming weeks.

8. Intention-aligned productivity software

In service of my “strategic exploration” sessions I’m working on software that will help me specify and execute upon arbitrary intentions, in the lineage of (but substantially different from) my human-programming tool Methodable and my living personal systems. As of now I have a CLI command that rearranges the windows on my computer, opens a project document, starts a timer, and periodically checks in with me (like in the screenshot above). In ideal form this software will self-modify, continuously adapting itself to best elicit and serve my intentions via value-driven interactive experiences.

This project is dual-purpose: I want to build tools that help me succeed at my exploration goals, but more so am interested in exploring this product area itself and adjacent career directions. I’m excited that Forethought, a prominent research nonprofit focused on ensuring beneficial futures with AI, has published sketches of “Angel-on-the-shoulder” software which are in the same neighborhood as my own ideas. And I’ve also been chatting with one Brian who’s working on a related project: creating a phone that cares.

This feels like fertile ground, but I want to make sure that my focus on building abstract tools does not detract from my more concrete aims, so I might prioritize this project a bit lower for the time being.

9. Meditation

I’m once again meditating every morning, this time using Henry Shukman’s fantastic “The Way” app. Despite all my projects and strivings I wish to stay present and enjoy life as it is, and I find these meditations to help me do that.

10. Blog

Last but not least is the blog. I’ve written 17 posts so far this year, the most successful of which has been my story of overcoming fatigue. That post has been read nearly 4,000 times, garnered many positive comments, and inspired two strangers to follow suit with similarly positive results.

Upcoming plans are (tentatively):

1. To continue documenting my recovery from post-viral fatigue, and, if everything goes well, to compile everything into a “here’s how I recovered” post

2. To write about interesting areas of AI research and other areas of research and product development that strike my fancy

3. To launch my piano album

4. (as per usual) to continue writing about whatever else currently inspires me

I won’t lie, publishing on a weekly schedule is quite difficult, but so far I’m following through with the commitment, as evidenced by the embarrassing number of posts I’ve sent out at 11:59PM. In fact, I’m so committed to my arbitrary deadlines that I paid 18 euros for Starlink in order to post about my habit systems from deep within the Mediterranean Sea. Make no mistake: the mail always goes through!

Subscribe now

It has three components:

1. Use tiny habits

2. Journal about my habits every day (as little as two words)

3. Make sure it’s fun

1. Use tiny habits

A while back I met the founder of a habit-formation-software company. When I asked what theory his software was based on he pointed me towards a book called Tiny Habits by BJ Fogg. I had heard of the more popular Atomic Habits which was published a year earlier, but according to my recommender (and online reviewers) Tiny Habits offered a more coherent and opinionated guide.

I gave it a read, thought it was great, and now follow the book’s prescriptions, making tiny-habit “recipes” that look like the following:

The recipes for the habits I make are all like this: after some trigger (such as another action I’m already regularly taking), I perform some very simple (“tiny”) behavior, then I celebrate. The celebration rewards the tiny behavior and causes it to stick. In the example above, though I truly intended to write in my journal, I keep the habit small - I just need to open it. This way the habit is very easy to start and maintain, and the full intended behavior of writing in my journal usually follows anyways (it’s pretty easy to journal once it’s open in front of you).

I recommend reading the book, which includes many more specific recommendations for various aspects of setting good habits. But even once you’ve learned the book’s frameworks, that’s still not enough to actually set good habits (in my experience). You need to apply the right strategies at the right times throughout your life. For me, that’s where the journal comes in.

2. Journal about my habits every day

Every day I write in my journal about a variety of things, one of which is my routines or habits. I use symbols to delineate the sections of my daily entry, so, when it’s time to write about habits, I write and circle a capital “H”2. Then I free-write about habit strategy, on topics like:

• What new problems do I want to solve?

• What new habits or routines might be helpful?

• How can I make those into three-part recipes?

• How are my ongoing habits going? How could they be better?

• If I stopped/failed at something, why?

The time and effort I spent on this “Habits” portion of my daily journal entry varies widely. If, for example, I moved apartments recently, then I’ll end up writing about what pieces of my old routines I should keep and how that all will work in my new environment. And at the same time I’ll be spending effort outside of my journaling sessions on implementing my new intentions. But if my routines feel stable, in my journal I might just write a quick “all good” and be done with it. So the journal routine stays lightweight and is easy for me to keep up through the ebbs and flows of life. I can even ignore it for a while, and as long as I come back eventually, I trust that I’ll be able to restart the reflective-workshopping process and my habits will end up in a good place. I try to make all of my personal systems forgiving in this way, to account for life’s curveballs.

This system helped me set up my sleep habits and my current morning routine of yoga, meditation, cleaning, and the consumption of various hot beverages. Of course it’s not perfect - it’s only as good as the things I already basically know how to do. But it does give me space to follow through with my intentions, and a good mechanism to course correct when things go awry, which is basically to write about the issues I’m facing each day until I’ve solved them.

3. Make sure it’s fun

If you’re actually interested in trying this I have one last suggestion, which is to make sure it’s fun. If you don’t already enjoy journaling, then step one is to come up with a journaling practice that you enjoy. That probably would involve more than just writing about habits, as it does for me.

Before I started my habit-journaling I was already writing every day about things I was thankful for, things I accomplished, ways I might improve, and the next day’s plans - each in its own section delineated by a unique symbol. Even with those guideposts my writing ranges freely, allowing me each day to make sense of my life and to work towards expressing my ideas and aspirations. With that practice in place, a “Habits” section was an easy addition, and contributed to an overall rewarding experience.

Subscribe now

The system described here is pretty simple, but I couldn’t find a description of anything similar on the internet. I suppose most people are looking for how to set good habits right now, without thinking about the more important question of how to keep good habits for life. For me, habit journaling fits the bill.

Part of a series on my personal systems:

1. Foundations of My Living Personal Systems

2. The Intentional Years ← you are here

Towards the end of high school I began following a set of self-improving personal systems. In the five or six years that followed, I poured perhaps 1,000+ hours and hundreds of written pages into the endeavor, producing an intricate system of interlinked intentions and practices, all formed around a set of core principles. Today’s post continues the story and describes that interlinked system’s workings.

The most fundamental motivating principle of my early personal systems was the following:

“It’s worth knowing what you want to do and what kind of person you want to be.”

Over time, my systems provided me with that kind of knowledge across many aspects of life. I had routines and plans for each day, goals or “modes” for each season, and most importantly, a set of overarching principles to guide my actions.

We can call the whole set of routines, plans, and principles my “intentions.” At my systems’ core was a set of reflection routines which compelled me to describe, maintain, and ultimately to manifest these intentions as best I could.

The core reflection routines

I had two core reflection routines: a nightly journalling session and a weekly review. These were codified in a document called “Constitution” which lived in my digital notes and was the entry point to the rest of my systems.

The nightly journaling sessions started as unstructured free-writes. Each night I wrote about what was going on in my life and about ideas for upgrades I might make to my systems, such as: “I’d like to set up principles for myself that I can reference,” or, “I’d like to write down what I’m grateful for every day.”

Then, during my weekly reviews I read through last week’s journal entries and made updates to my digital notes. In part my goal was to integrate the week’s ideas into my systems as appropriate, but I also allowed myself to make additional changes, informed by the zoomed-out reflective standpoint that the weekly review provided.

These two core routines maintained and built upon themselves. For example, during one weekly review I read in my journal that “I’d like to set up principles for myself that I can reference.” So, I set forth to manifest that. I created a digital note titled “Principles” and updated my Constitution in two places to reflect the following: that during nightly journaling I should prompt myself to write about learnings that are relevant to my principles, and that during weekly review I should modify my principles based on the learnings collected in the prior week.

My goal for any described routine was for it to become a habit, and for that habit to effectively accomplish the routine’s purpose, and for the entire story of the routine/habit and its purpose to be well-described within my system’s Constitution. To achieve that state of systematic integrity I knew I had to regularly “give the system love,” meaning, to apply steady creative energy towards the improvement of each aspect of the system and towards the system’s overall coherence.

I also believed strongly that my systems would work best if I could genuinely commit to following my stated intentions, so I did my best to maintain that commitment. Doing so gave my systems the power to adapt and mature by their own rules.

Factored journaling for nightly reflection

Over time my journaling routine evolved, and pretty soon I was doing the following: each night I’d proceed through a series of prompts represented by the acronym TAIL: “Thankfuls”, “Accomplishments”, “Improvements”, and “Laters”. In order, I wrote about:

1. (T) What I was thankful for - in my relationships, in the world around me, and in myself

2. (A) What I accomplished that day, especially in light of my principles

3. (I) Ways that I might improve myself and/or my systems - this section would directly feed into my Principles document and others

4. (L) What I needed or wanted to do the next day

And finally I’d put a placeholder for recording dreams in the morning.

This system served both my immediate goals of integrity + calmness + gratitude + self-direction, and my long-term goals of systematic self-understanding and improvement.

Principles and their synthesis

Over years of weekly reviews I read every word of my daily journal entries, integrating my scrawled thoughts into a growing body of digital notes. The most important digital notes were those relating to my principles.

At first, my principles were a single document: a list of eight to ten sentences, like this:

Be a confident role model.

Together my principles made up a concise description of the kind of person I wished to be and the kind of life I wished to live, and I intended for my actions and my systems to serve them.

Once my principles document was set up, I built a queue to aid in the principle-synthesis process. During weekly review time I would copy my daily “Improvements” notes into this queue, and in time I would process each item in the queue, considering whether it might call for a modification to my principles or perhaps to some downstream system.

Eventually I realized that my true intentions, and even the subset of my intentions that I wished to write about, far exceeded what could be conveyed in these brief sentences. To give my intentions more space I started a new document titled “Principles: Full Considerations,” which grew to include short essays under each principle. Here’s an example (note: this is a shorter one):

Be a confident role model.
A successful role model has many traits: not only does he act in a positive manner and go about life in a way that benefits everyone, he also must be confident in his ability to teach, and willing to put himself in the public view.

I must start by understanding that people look up to me already and will emulate me if I am willing to make visible my thoughts or actions. I must be confident enough in my unique ways to put them on display, and I must display them confidently.

I must be the first to spring to action when a situation requires it, being guided by what is the right thing to do and not what others think. Don’t use “I don’t know” as a crutch.

As to my role? Following my other principles is enough to define my role as it pertains to being a role model. An additional note is to act respectfully in public settings, especially on the internet, and to show the best sides of myself to every single person I meet. Remember: I have the power to define the energy in any setting.

…

To describe the full circle of principle synthesis: during nightly journaling I wrote about specific ways to improve, sometimes targeting my writings towards a particular principle and sometimes writing in a more free and uncategorized way. Then during weekly review I moved these thoughts into my queue for principles. Later I would integrate these thoughts into some appropriate section in my full-considerations doc, or, in rare cases, mint a new principle. In the end I had a well-adapted set of about ten principles with accompanying essays that integrated thousands of days of reflections.

The challenge of systematically realizing principles

Of course I knew that merely writing down intentions wouldn’t make them a reality, so I was constantly devising new schemes to become the kind of person my principles described. That endeavor was, of course, the most difficult part of the whole system, and I remained unsatisfied with my attempts to the end. But I tried many schemes and I liked some of them.

At first I would literally carry my principles in my pocket. Then at some point I created a new document called “Principles - Actionable Ideas” which described a protocol I could follow at any time in order to ground myself and direct myself towards positive action. Funny enough, the protocol started with a reality check, then branched one way or another depending on whether I was dreaming (I was practicing lucid-dreaming at the time).

But with that protocol there was still something incomplete - how would the full extent of my intentions cash out into reality? So I kept iterating. The next innovation was a “principle of the week.” During weekly review I’d choose whatever felt like the most relevant principle, then throughout the following week would direct my focus towards that principle by various methods. I also set up “seasonal modes,” which, like principles, represented intended ways of being, but were more closely adapted to my present circumstances.

I felt positive about these various attempts at principle implementation, but I remained unsatisfied, and despite my diligence things seemed to slip through the cracks. I loved my routines and thought that they brought me great benefit but I did not believe I had created the perfectly self-regulating system of guidance that I wished to create.

Auxiliary systems

While my main focus was towards the flourishing of my principles, I also regularly attended to a number of other systems and their written counterparts. The rest of my systems were ostensibly downstream of my principles, and I wished to make the links between each system and its motivating principles explicit, but I often struggled to do so (probably for good reason - motivations are complex). Regardless, I had a vibrant array of offshoots. Among them:

• Worksheets from the book Designing Your Life

• A list of close friends, and for each: how they inspire me, what they give me, what we have in common, what I could give them

• A list of accomplishments, organized by season

• A list of times I cried, and other lists of meaningful moments in my life

• A list of good stories that I wish to remember

• (and many more)

Living an intentional life

To summarize: each week I opened my journal, read through the week’s reflections, and updated my digital notes. Sometimes I changed my principles, sometimes I made changes to my reflection systems, and sometimes I worked towards setting daily habits that would help me live a principled life.

For at least four years these systems were going strong, and they gave me a strong sense of trust in myself. I knew that every morning I was waking up with the best intentions, and that whenever I fell short, as all people do, I would learn and improve to the greatest extent of my abilities. I was very happy with my systems and imagined them extending far into the future, and yet, I also felt there was much work still to be done. If only I could figure out the right algorithm to integrate everything together - to optimally guide myself into perfect alignment with my deepest values - then the vision would be fully realized!

Some concluding notes: At this point I’ve covered, from a very high level, the growth period of my living systems. Perhaps you now better understand what it’s about.

There’s tons more I could publish about these systems, much of which is already written (there are benefits of writing about something which was designed to describe itself). Of that, there are many details I’m truly excited to share, such as: what concrete outcomes did these systems create for me? Or, what were the rest of my principles and their full considerations? And, I suppose I might also cover: what happened to these systems? Why is everything in the past tense?

To be honest, to describe the downfall of these systems would be a daunting task. Perhaps I will find a good way to do so, but I expect that to take time. In any case, today’s post was the most important part of the history. Now that I’ve described my prior systems I expect to more easily describe their present evolutions, as indeed a revitalization of sorts is taking place. My next couple of blog posts will probably focus on health, and after that I look forward to returning to this theme.

Subscribe now

Three weeks ago I set this determination:

As long as my fears, pains, and concerns do not exceed what I experienced this week, I will keep pushing through them, choosing to live the life that I most wish were possible. If that succeeds, then eventually my body will catch up with my brain, learning to once more have confidence in itself. And then, finally, there will be no fears and nothing left to push through.

source: I tried David Chapman’s chronic-fatigue-defying six-mile run. It worked.

So far I’ve been doing exactly that, and I’m now on my fourth week of “choosing to live the life that I most wish were possible.” I’m now quite close to my holy grail of running three times a week and climbing/lifting three times a week.

As predicted, along the way there have been some challenges. One day in particular last week I woke up feeling quite tired. My body felt weird and my brain wasn’t working very well. I’d say it was “the normal stuff” but in truth, “the normal stuff” for me varies, and each problematic experience is a little different. I can at least say it was familiar.

It’s natural to see such challenges as setbacks, but recently I’ve been seeing them otherwise: as gifts.

As I understand it, over the last year my body has been conditioned to produce adverse effects (fatigue, inflammation, pain) in response to a wide range of stimuli that occur naturally in a full, engaged life. In order to conquer these adverse effects, I must provide the opposite conditioning - I must live a full, engaged life, with all of its stresses, and then when the adverse effects inevitably show up, I must demonstrate to my body that “actually, everything’s fine.” I can do that by maintaining positivity through challenging periods and by continuing with my intended activities, exercise-related or otherwise.

It’s like there’s a minefield of stimuli, and each time my health fluctuates it’s an opportunity to defuse another mine, to train my body-mind system to skillfully navigate some new area of life. Eventually all the mines will be defused. I expect that to take some time, and in the meantime I will continue to see each challenge as a unique opportunity to move my recovery forward. It’s almost a joyful thing: “Ah, another opportunity. Excellent!”

This mindset can apply to all sorts of challenges, not just my health ones. All challenges are opportunities to demonstrate and learn skillful behavior, expanding the boundaries of what one can do well.

In the end my problematic feelings this week didn’t last long. I felt much better the next day and ran another 10k, my best yet. Now, a few days later, I’ve finished my first week at my gold-standard activity levels. Maybe next week will be easier, or maybe there will be more challenges. As long as I keep doing what I’ve set my mind on, I’ll be happy.

Subscribe now

During my first few months of post-mono fatigue I struggled to fall asleep and stay asleep. It seemed that sleep had just become harder. Regardless of the underlying problem, the solution was straightforward: I just needed to create better sleep routines than my previous ones.

After a particularly un-restful period I decided to apply some effort to the matter and began researching optimal practices, and then began testing them. Eventually my issues resolved and I kept my new sleep habits to this day, though with minor adjustments. I thought I’d share what worked for me:

Consistent wake-up time

Every morning my alarm goes off at the same time (call this “t”). Consistent alarms from day to day help my body fall into a natural rhythm, so even if I’m up late I prioritize alarm consistency over sleep duration.

Wind down routine

I’ve learned that I sleep best when my lights are off at t - (8h 40min), so I work backwards from there. On an ideal day:

1. t - (10h) (or earlier): Wind down begins:

   1. Plug in / put away phone (not in bedroom) - the farther and more out-of-sight the phone, the better. Once it’s plugged in, I don’t use it. I also use my phone’s built in sleep controls to suppress notifications.

   2. Turn off lights, unplug bright computer keyboard - darkness helps with circadian rhythm.

   3. Turn on red bed-side light - red light is circadian-rhythm friendly and routine red light becomes a pleasant backdrop to pre-sleep rituals, plus a consistent signal to the body of impending sleep.

   4. Lightly organize apartment - putting things away helps “put away” the thoughts associated with them. It’s also just a nice wind-down-friendly task.

2. t - (9h 30min): Get in bed, journal (20 min) - I follow a structured journaling routine. Like cleaning the apartment, journaling also helps me “put away” thoughts from the day. I write about things I’m thankful for, things I accomplished, how I might improve my actions and routines, and what needs to get done the next day. When I’m finished, my thoughts are quieter.

3. t - (9h 10min): Read a book (fiction preferred) (30 min)

4. t - (8h 40min): Turn on alarm, put in ear plugs and put on eye mask, turn off red light, and close my eyes.
   …

5. t: Alarm

Sometime I extend or shrink the wind-down window to accommodate for the natural flows of my schedule, but I try not to shrink it too much.

Keeping the bed(room) for sleeping only

A lot of my routines above relate to ensuring that my sleeping spaces are used only for sleeping - that’s why my phone stays outside during the night, and my bedside table only has exactly the items needed for wind-down:

1. lamp with red bulb

2. journal

3. book

4. alarm clock (prevents reliance on phone)

5. ear plugs and eye mask

My setup isn’t fully ideal because my current bedroom contains my at-home office. But at least my computer and monitor don’t face the bed (they’re sorta around a corner), so there’s some mental separation.

I also try to avoid lying on my bed unless as part of wind-down/sleep. I’ve made exceptions in the past when I was extra tired, but in hindsight it would have been better to use the couch so that my body knows that bed = sleep.

Melatonin (rarely)

I only use Melatonin on rare occasions when my circadian rhythm seems behind schedule, using the clinically-studied dose of 0.3g (lower than most supplements) and timing it about an hour before I intend to sleep. Once my sleep seems roughly normal (generally after one night) I stop taking it, and otherwise don’t take supplements for sleep.

I tried magnesium glycinate, but didn’t personally notice a difference from taking it.

Pleasantness-maxxing

My routines form a sort of pleasant ceremony. Once my red light is on, it’s a bit like I’m in another world: a world where I only do pleasant, relaxing activities.

Nowadays this is a joyful part of the day that I can look forward to - when I was struggling with poor sleep and my other fatigue issues, this period felt more like a necessary respite. In any case, pleasantness is both good for sleep and good for its own sake.

Electronics off early

…just bears repeating - my phone is nowhere near me during the wind-down window. If there are internet things that I wish to get done, I write about them in my journal and save them for the next day.

Creating and maintaining routines (to be continued…)

My particular routines might not work for others, and it takes time to onboard any kind of routine, especially complicated multi-part ones (like the wind-down sequence described above). So, my real recommendation for someone looking to improve their sleep would be to develop a good routine-generating process - something I’ve put a lot of effort into myself.

I’ll write about routine-generation soon in a post titled (tentatively) “Journal as core routine engine.”

Last week I wrote (here) about my recent success in pushing through a post-viral fatigue spell, running six miles and coming out more energetic on the other end. I realized I have more to share - a long history of experiences and attempted interventions which together could form a useful fatigue-recovery case study. Today I describe my experiences with hypnotherapy, an intervention which changed how I understand my illness and plausibly helped me recover faster.

…

I learned of hypnotherapy through a close friend who used it to treat his chronic shoulder pain. He had a positive experience and eventually offered to buy me a session with the same practitioner. That was a year ago, three months after the start of my illness. I appreciated the thoughtful gift but wasn’t eager to begin: I trusted that my fears, pains, and other body signals were all well calibrated to prevent me from taking harmful actions in my fragile state.

Ten months later my perspective had shifted. As described in last week’s post, a series of positive experiences with exercise caused me to reevaluate the wisdom of my symptoms. By January 2026 I felt I was ready, and I reached out.

At the time I understood my condition as follows (excerpts from the intake form):

What was I experiencing?

Presenting issues:

Since having mono a year ago, I’ve had some fatigue issues. Things have actually improved in the last few weeks and I have not felt many symptoms even though I’ve been more active (e.g. have been back to the climbing gym about twice a week). So I’m currently optimistic, but also things tend to change randomly, so I’m not fully optimistic.

…

If you are experiencing discomfort, on a scale of 0 - 10, which number represents it at its most severe?

5

And which number represents it at its best?

0? (Or maybe 1. If I think about it, there’s usually something there. But I think there are times when I don’t think about it and just feel fine.)

Describe the nature of the discomfort using any descriptors or metaphors that feel most appropriate:

It starts in my shoulders and neck, which will feel sore and kind of sick. And then maybe I’ll feel it in my arms too. There’s a sort of heaviness sometimes, like, just taking action feels a bit difficult, like there’s a bit of a haze separating me from the world or from the actions I might take. Things just don’t feel very fluid. I do associate a particular “inflammation-y” feeling as well that’s very difficult to describe, but where there’s a kind of tenderness in my shoulders, or my chest, or my gums, or my arms… as I focus on these places there’s this field of tenderness. Overall things just feel delicate, like I’m made of glass (slightly).

And what was I hoping to change?

I want to be not afraid of physical exercise, and to feel good after doing it. And to be not afraid of work, and not afraid of going to New York. Just generally not afraid of doing hard/exciting things. I want to feel like I can push myself to the limit and feel rewarded for it, and like I am full of energy.

I also listed some moments that made me happy to remember, which would become relevant in the upcoming session:

There is one spot in Castle Rock that I’ve been to multiple times, napping on the grass on an overlook…

…

Getting acquainted with my red field

I logged into Zoom for my first session. After probing some core facets of my life, the hypnotist asked me to come up with an image to represent my symptoms. “It’s like my body is made of glass,” I said. He asked again:

“Can you describe it as something separate from your body?”

I thought. I said “it’s like a red field that permeates my body and grows wherever my attention touches it, concentrating and spreading inward. Whichever parts it occupies are filled with a diffuse pain, and I am compelled to hold them still, to move them with great gentleness and care.”

He explained that this field could be thought of as a piece of software that runs on my nervous systems and/or subconscious mind. Our goal was to update this software.

To do so he would guide me into a relaxed dreamlike state, and in that state I would perform a sequence of visualizations and mantras. He assured me that my participation would be voluntary throughout. He taught me a “vagus-breathing” exercise, suggested as an intervention that I might use during times of anxiety, then he guided me through a body scan, and then, finally, deep into the hypnotic state.

Upgrading my software

Hypnosis was indeed a bit like dreaming, but less intense. I was mildly aware of my true surroundings and occasionally veered off into half-formed thoughts, but each time I eventually came back to the visualizations. Over time I experienced a pronounced shift in my mental background: I felt more calm, free, and fearless, as if I were “trying out” a different way of being.

My mental journey began in Castle Rock, lying in the grass with a hat covering my face, relaxing in the sunshine after a long hike and a big lunch. Once I established that positive backdrop I was guided into a dialogue with my red field. I asked the field its purpose, learned that it was well-intentioned but perhaps miscalibrated to my present circumstances, and thanked it for its service. Then, in a lengthy ceremony, I disposed of it completely:

After that I designed a new piece of “software” to replace the field. I imagined a thin blue suit covering my entire body, lined with a network of tiny tubes. The suit would regulate my body’s resources, and with it donned, I was assured that whatever rest my body needed would take place naturally and without conscious intervention.

I imagined walking around with my new suit, first in Castle Rock, and later in a climbing gym. With my hypnotist’s guidance I accentuated the positive experiences of this new way of being: the freedom, the joy, and the boundless energy. In this new world my body was no longer problematic and my surroundings were once more full of possibilities. I stewed in these positive experiences then was guided out back into the real, unimagined world.

Immediate benefits in my daily life

After the session my hypnotist encouraged me to take practices into my daily life, including routines (like self-hypnosis) and ad hoc interventions (like the vagus-breathing exercise). And, I was to take my blue suit with me. Whenever I felt my field of pain / fear / confusion / tiredness, I imagined my blue suit working away, optimally managing and restoring those problematic areas, and I felt comforted.

For example, I went on my first trail run in many months soon after our session. I remember at one point losing energy and slowing down. Instinctually, the fear flashed:

You have exhausted yourself. Feel your body - are the symptoms there? You must be very careful.

But I remembered my suit, and thought:

My body will take care of itself. It has already slowed itself down, naturally. I need not worry.

The ground truth was the same - I was on a run, I did a lot of walking, I was tired. But when I remembered the suit, my fears relaxed.

On the reasonableness of hypnosis

For a year I was plagued by a fickle illness, never knowing when the next symptom would strike. It seemed that using my muscles would cause symptoms to worsen, and those symptoms largely struck the very same muscles. Of course that caused me to develop heightened awareness of my body’s warning signals, and with such heightened awareness it’s no surprise that the “red field” got ingrained.

Although the field seemed to communicate that my body still had fundamental issues, I couldn’t exactly take it’s existence as evidence of fundamental issues - it’s possible that the field was just hanging around. And indeed, after experiments with gradually reintroducing exercise, I started to believe that my underlying health dynamics were more favorable than the field was letting on, and that perhaps the field was causing me to be overcautious.

Or, put more simply: I spent a year in fear, hypnotizing my red field into existence. After such conditioning, it made sense to spend a few hours un-hypnotizing myself towards a more normal relationship to my body.

And indeed the blue suit I imagined was quite realistic: we all contain an optimized network of (blue) veins that carry unneeded matter away from our organs, in concert with energy-providing arteries. This circulatory system, and the rest of our body systems, perform well without conscious intervention. A miracle!

My outcomes and the recovery flywheel

In the months following my sessions1 I experienced a recovery flywheel: as I reduced my fears and perceived symptoms I started maintaining gradually more intense exercise routines, and as I succeeded in these routines my fears further receded.

I think this “recovery flywheel” concept is critical to understanding my illness and how I got better. Such flywheels get stuck when perceived negative reactions to exercise outweigh perceived benefits. My own flywheel seems to have been stuck since I originally fell ill, and, while it’s hard to say for sure, it’s possible that my hypnosis sessions provided just enough grease (via lessening of fears) to get the flywheel back up and running.

To be clear, the effects of those sessions barely hold a candle to the fear-dismantling that occurred in the wake of my recent experiments with pushing through fatigue via exercising hard. There’s nothing like physically proving your perceived limitations to be spectacularly wrong. But still, when I look at the past months, it seems reasonable to suggest that my hypnotherapy contributed to a buildup of self-trust which allowed for my later more impactful experiments to take place.

At the very least I can say: my sessions were pleasant and my practitioner was a thoughtful guide. The practices I learned brought me comfort, and their introduction to my life coincided with an inflection point in my physical activity and health. Hypnosis seems to have been a helpful tool for my particular stage of post-viral-fatigue recovery.

Subscribe now

In December of 2024 I had mono, a disease which upended my life and caused me fatigue issues that plague me to this day.

I recently tried a new recovery tactic with astonishing results.

The crash

On Sunday last week I woke up too tired to complete my 10 minutes of daily yoga: a sure sign that I was entering a low-energy period. I might have gone on a run that day, but knew I needed to rest.

The next day (Monday) I dressed up in my running clothes and went to a cafe, but as I sat there my energy fell lower and lower. I felt like I wanted to throw up. I did not run. I returned to my car and lay my head against the passenger seat. After returning home, I did the same. It took a lot of effort to get upstairs, and I cancelled my plans to cook myself dinner. As I proceeded through my wind-down routine, I paused to sit or lie down between each task.

On Tuesday I completed my yoga. My body felt slow but my head felt clear. I had one meeting, then started to write. I began to feel more and more tired. I moved to my bed. I switched to a computer game. Then I turned that off and closed my eyes, feeling I could do nothing else, completely spent. I was reminded of times during my original illness 15 months ago, and times since then, when I’ve been similarly debilitated. I prepared for the possibility of an unproductive day, and of multiple days of continued exhaustion.

I got up and went to the bathroom, and while there I recalled a blog post from David Chapman: Fake fatigue in long covid. I reread the post, attending closely to the story of his fatigue-defying act. Chapman’s words:

I pried myself out of the beanbag. I was dizzy and lightheaded, and my wobbly legs could barely hold me up. Slowly and carefully I got to the garage and drove to the park.

I ran six miles. Then I came home and wrote this post.

Fuck you, long covid, you are a fucking liar.

As I sat there reading his story, I began weighing the risks. I thought this might actually work for me. I decided to do it.

…

I pried myself off of the toilet (sorry). I was dizzy and lightheaded, and my wobbly legs could barely hold me up. Slowly and carefully I got dressed and walked outside.

I ran six miles. Then I spent a week being physically active and productive, and am now writing this post. So I too will say (excuse the profanity):

Fuck you, post-viral fatigue, you are a fucking liar.

I share this post in case it might help anyone else with similar issues. I include the history of my challenges with physical activity, the run itself, the week that followed, and my updated recovery outlook.

Caveats / you should read Chapman’s post

Chapman is a very good writer and he covers the necessary caveats. Most importantly: don’t take medical advice from internet randos. I agree, and will quote him here:

I’ve figured out a way to deal with long covid fatigue that has worked for me. It might work for you, if you have that. I have no idea. It may also cause other trouble, so you’d need to be careful.

Same situation here. This is a story about what worked for me.

I can at least attest that my story is accurate. I kept a detailed journal of my feelings and activities over the past week, updated multiple times a day. I’ve also spent months collecting daily self-reported health logs, and years writing and re-processing daily journal entries. All of those personal writings helped produce today’s story, which really began 15 months ago.

The history of my challenges with physical activity

I tested positive for mono and strep in December 2024, and then the flu one month later. I continued experiencing fatigue throughout 2025, and wrote about these experiences in my 2025 year’s end blog post.

In July 2025 I began tracking three self-reported health metrics on scales of 1 to 5:

1. how mentally sharp I felt

2. how much energy I seemed to have

3. how much soreness/inflammation I felt in my neck and body

These three metrics represented the signature symptoms of my condition. Summing them together, a 15/15 represented a normal-feeling day, and anything less represented various levels of ailment. Here’s what 7 months of data looked like:

As you can see, symptom-free days were few and far between. I stopped tracking in February 2026 in part to remove the burden of the task and in part to prevent myself from drawing unneeded attention towards the illnesses.

The struggle to reintroduce running and climbing

For many years I’ve enjoyed rock climbing, weightlifting, and running. I’ve told myself that if I could return to my pre-fatigue routine of climbing/lifting three times a week and running three times a week, I would then consider myself fully recovered.

During 2025 I intermittently attempted these activities, but often experienced delayed-onset fatigue, roughly one day after. I interpreted that to be post-exertional malaise (PEM). My read of online forums and publications from advocacy groups told me that PEM could have negative health effects, so I felt discouraged from continuing those activities. I generally opted for yoga and long walks instead.

The negative aftereffects of exertion were strongest for climbing, perhaps because my ongoing symptoms of pain/soreness/inflammation were clustered in the same upper body muscles that climbing engages. So, whereas I sometimes sustained an intermittent running practice, my climbing practice never got off the ground, though I tried a handful of times. Even running was difficult to maintain. My frequent symptoms made it harder to run and made me less certain that running was a good idea.

In October I moved to Seattle. During my first two months (November and December) I was often sick. I felt that my health may have regressed slightly. While I would sometimes go for walks or hikes, I did not feel comfortable making progress towards running, climbing or lifting.

Then things started to get better.

Improvements in the last two months

With the encouragement of my friends I returned to the climbing gym in late December. The day after my first climb my muscles felt sore and inflamed, as if I had a viral infection. After about three days those symptoms mostly resolved. Despite the adverse response, I felt that I had ultimately tolerated the exercise, so I continued going to the gym.

For a few weeks I climbed once or twice a week, doing short sessions (~30-40 minutes). Then I started running again. Slowly I ramped up to a point where I was running nearly twice a week and climbing nearly twice a week. The runs were about 2.5 miles with a fair amount of walking interspersed.

Throughout that period I experienced my normal fluctuations of symptoms, and often did not feel good, but crucially believed that I was not hurting myself, so I kept going. That change in belief was a tipping point in my recovery.

But despite my sense of progress, I was not encouraged by any of the feelings I experienced in my body, during or after activity. I was operating with great caution, and my symptoms often compelled me to reduce my activity and take unplanned rest days. Progress felt like a constant struggle, and it’s that tenuous headspace that formed the backdrop for last week’s crash and the six-mile run.

Why did I take the risk?

What was going through my head when I decided to go on that run?

It’s hard to describe. My felt experience of making health-related decisions is one of deep, complex uncertainty. As if I am dealing with a tangled ball of thoughts and feelings that spans multiple miles, so large that I cannot bring it into view and cannot think straight about it. The main component of the ball is fear, which suffuses everything.

The fear makes sense. I kept trying to expand my exercise envelope, and despite some success I continued to feel bad and to randomly crash. For each negative experience, I would worry I’d overdone it and would think about the potential causes, exercise-related or otherwise, and the fear would expand into more and more areas.

At the same time, a number of positive experiences gave me optimism. From as early as February 2025 I had moments where my energy seemed to suddenly recover: first at a bachelor-party weekend, then on a family vacation, and later during a series of backpacking trips. In hindsight these were all fun, multi-day activities in new environments.

The backpacking trips in particular were confounding. I would work up the courage to go out, set backup plans in case my health failed, and then proceed to feel great the entire time. Part of me still worried: I would feel exhausted after the trips, and backpacking is a pleasant activity for me, so perhaps I was still operating under limitation and offsetting the consequences? But it really didn’t seem that way. The memory of these trips became a beacon of hope for me.

Between the backpacking trips, my recent successes with re-introducing climbing and running, and my exposure to cognitive-retraining success stories1, I figured that I might be able to tolerate more than it seemed, and that I could engage in this experiment without serious risk to my health. The veil of fear would have to yield this time.

Also, I won’t lie: I was thinking it would make for a good blog post.

So I went on the run.

Quick note on terminology: long covid vs. post-viral fatigue

Before we move on I need to clear something up: I generally say that I’ve experienced “post-viral fatigue,” since that most directly describes my situation. If pressed I’ll say it was caused by mono, which is a frequent instigator of such issues. However, I’ve had covid, and probably long covid before (details at footnote2), and there’s evidence that Epstein-Barr virus reactivation (which causes mono) may be a contributor to many long-covid cases (see Dan Elton’s notes on this). Furthermore, post-viral fatigue from mono and post-viral fatigue from COVID are not clinically distinguishable - they share symptoms, mechanisms, and even treatment recommendations. As such, I didn’t mind calling it “long covid” during my run - it’s all post-viral fatigue to me.

The Run

I copied Chapman, heading out the door while forcefully saying “fuck you long covid,” trying my best to energetically cast off my limitations.

At the beginning running felt… unnatural. My vision was blurry and my thoughts were scattered, and I felt sluggish. But I could run. Over time my vision became normal and I felt normal, and then began to feel good. My congestion cleared up and I sustained a jog for most of the six miles. By the end it felt like running under normal healthy conditions.

That afternoon and evening I was productive and had plenty of energy. My lungs felt tired and I occasionally brought up phlegm, but otherwise my body felt great: I had a pleasant tingly feeling all over. The run had completely turned things around.

I was optimistic, but I wanted to wait before publishing my results. I’m glad I did, because the next week was even better.

I kept pushing myself and it kept working.

In keeping with my best-case exercise routine for the current period (run → climb → rest day → repeat), I proceeded to climb the next day (Wednesday). It was by far my best climbing day since 2024, and I climbed my first “blue” (an arbitrary difficulty grade at the local gym).

Two days later (Friday), I was out running again, this time in the snow:

About three miles. It went well. Just like on Monday (the first significant day of my crash) I was two days out from a pair of running/climbing days. Just like Monday, I worked on my blog (the same post on both days - an upcoming one). I did all of that to make Friday a controlled experiment. Happily, writing went well, as did the run.

The next day (Saturday) I was at the climbing gym again. I climbed my first “pink”:

Also, for the first time in months I lifted some weights - just a small amount to get started. Two days later (Monday) I ran far again.

I made poor choices regarding the timing of coffee consumption and ended up walking a fair amount, but I made it home ok and felt fine the next day. That brings us to today (Tuesday), during which I wrote the majority of this post and had another great climbing session.

Feeling motivated? I hope so! Allow me to fill in some gaps…

I felt slightly bad and concerned for my health every single day (and that might be a good thing).

After my run on Tuesday, I worried about my lungs. I told myself it would be fine, but was still scared. On Wednesday I wrote “My body is sore, a situation which threatens to draw my concern, but I remember that soreness is normal.” On Thursday I woke with a sore throat, but then began feeling better during my yoga.

Friday morning:

feel pretty sore, under the shoulder blades and such. The kind of thing that might normally be a warning signal. “Guess I’m sore after climbing” I think, and move on.

Saturday on the way back from the gym:

my body is throwing warning signs. As if I’m about to be sick. I Ignore them, have dinner, take a nice long bath. Afterwards, things seem fine. I am tired, but not unwell.

Sunday:

I woke up thinking I might be sick. Phlegmy, not feeling great. Still, got up, did my routine, did my yoga. Spent half the day cleaning my apartment, fasted, and eventually felt fine.

[…]

there’s still that feeling of heaviness, but only barely.

Monday:

again, quite zonked upon waking. This time, yoga was a bit harder. Once again, things progressed fine throughout the day.

…

Every single day was hard. And that might be a good thing. If I’d felt great the whole week I would have learned that I can have good weeks. Instead I learned that a lot of my negative feelings don’t cash out to physical reality, and that I can have a good week despite them.

Some closing thoughts:

This is all very exciting.

This seems a bit out of place in a sober analysis of the situation, but, yeah, I am really looking forward to slaying this demon, and this past week was like a big harpoon right in its rear end. It feels very good to go to the gym and pull hard. And it feels very good to experience hunger again (like, literal hunger, from exercising). If I can continue at my current rate of activity I will be over the moon.

I am glad that my friends invited me out.

My first two backpacking trips of 2025, my snow-camping trip in January, and my recent reintroduction to climbing were all initiated by friends. When I accepted each invitation I felt I was taking a risk, but in each case I was excited to join because of the people involved. Without these invitations, and the resulting disconfirming evidence I received, I would not have had the courage to push through this week. To everyone involved - thank you very, very much.

The plan is to keep at it.

As long as my symptoms do not exceed what I experienced this week I will proceed to treat them with “aggressive contempt,” as Chapman put it. I will continue the current exercise routine, and if that keeps working I will expand it until I’ve reached my ideal pre-fatigue cadence, with respect to the gradual approach required to recondition muscles (especially in weightlifting).

Of course, I am reporting on recent events and I have no idea how my health will progress. If I look within, I am still gripped by terrors: what if this week was an anomaly? What if I’m just building up for a bigger crash, one that I can’t fake my way out of? What if I have to turn around and say I was wrong?

The fears are to be expected. That’s how these illnesses work. But I do not intend to pay them heed. As long as my fears, pains, and concerns do not exceed what I experienced this week, I will keep pushing through them, choosing to live the life that I most wish were possible. If that succeeds, then eventually my body will catch up with my brain, learning to once more have confidence in itself. And then, finally, there will be no fears and nothing left to push through.

References and resources

1. The original “fuck you long covid” run:

Meaningness

Fake fatigue in long covid

This morning I was half-sitting, half-lying on the tiger-striped beanbag, weak and exhausted, mentally and physically…

Read more

2 years ago · 60 likes · 32 comments · David Chapman

2. Dan Elton’s writings on surmountable causes of fatigue:

More is Different

The "false alarm" theory of Long Covid fatigue

Note (added 9/29/24) - I do not think this explains most Long COVID symptomology. Even when I wrote this, I was thinking this theory was maybe only relevant in ~50% of cases. The main drivers are probably persistent virus, Epstein-Barr reactivation, immune dysregulation, and mitochondrial damage. Another point I would like to stress here at the start is…

Read more

3 years ago · 19 likes · 2 comments · Dan Elton

More is Different

Psychosomatic contributors to Long Covid suffering

“Some people are willing to ‘feel sorry for the sufferers’ if we find a traditional pathogenic mechanism … and that, to me, seems a fundamentally flawed point.” - Alan Carson…

Read more

3 years ago · 11 likes · 6 comments · Dan Elton

3. Mark Lovett Wells’ extensive long-covid wiki, presenting a path to recovery that does not involve pushing through discomfort: My Long COVID Treatment Strategy.

Bridge, Rudolph Ruzicka, c. 1915-1926

I’m not publishing my best essay of 2026

Yesterday, I performed an exciting “health experiment.” I’m planning to give you the full report next week.

In preparation for that event, I resumed work on the essay about mono which I foretold in my 2025 end-of-year update:

The three-month period depicted above was very challenging, and it’s daunting to try and do it justice in writing. But I have planned a post about this, “The Phenomenology of Mononucleosis” […]

At risk of tantalizing my audience I’ll say that the essay I produced is one of my finest pieces of writing: a deeply vulnerable story, charting a wild and disturbing psychological journey, punctuated with moments of the sublime, interwoven with images and excerpts from my journal…

But as I brought that period of my life into definition, I felt a growing sense of unease. The situation I was depicting, while fascinating, was really not a good one. And I figured that publishing it would give it more life than it deserved.

As I mulled over my decision, I began to play with a phrase, at first in my head and later out loud:

“We can probably move on from that”

Each time I said this, I felt more certain of its truth. I felt liberated. I don’t need to publish this, I don’t need to think about it, I can relegate it to the depths of my drafts folder and can do so with no ceremony. So that’s what I did.

I am someone who tends to finish things that I start. This has obvious benefits, but a drawback is that I can lose connection with my original motivations. In this case, I was motivated to write about health. Why? Because I’d like to help others and myself navigate those challenges and move in a positive direction. On reflection, the post I was writing didn’t serve that purpose. Whereas next week’s post almost certainly will.

The words we say are like spells cast into the world. “Where your focus goes, your energy flows,” says my Qi Gong teacher. And this is especially true of the words we put up in public places (like blogs): each becomes a node in a powerful network that conspires to shape our lives. As such, internet-writing experts like will remind us that it’s best to “focus on what you want to see more of” and to only “joke about outcomes you want.”

This doesn’t mean I intend to ignore life’s challenges. Instead, I hope to build narratives that transform my challenges into vehicles for optimism and growth.1 So how might I build those narratives for my own health story?

Crafting minimal bridges over the narrative landscape

As I see it, setting up beneficial narratives around challenging experiences is a creative exercise. You’re dealing with a messy landscape, and different people have seen you at different parts of it. You need a story that can hold up to a variety of inquiries and redirect them towards positive directions. So if the landscape has bogs and chasms, your goal is to build bridges.

What surprised me yesterday, but is obvious in retrospect, is how small the bridges can be. It doesn’t take much to get to a point where you can say “that’s enough explanation. We can just move on.”

When I consider my health story, it’s clear that my end of year post was a perfectly sufficient bridge. It touched on my challenges, but only in enough detail to clear up any confusion before reorienting towards the present. The particulars of my winter of sickness, in the end, despite being interesting, and complicated… are not relevant. What’s relevant are all the great things I can do with my life right now. And so there I leave it.

Remarkably, last night after I decided to shelf the essay, I came to the following quote in The Lord of The Rings (The Return of The King), which I’m reading for the first time (mild spoiler alert):

“They stripped me of everything; and then two great brutes came and questioned me, questioned me until I thought I should go mad, standing over me, gloating, fingering their knives. I’ll never forget their claws and eyes.”

“You won’t, if you talk about them, Mr. Frodo,” said Sam. “And if we don’t want to see them again, the sooner we get going the better. Can you walk?”

“Yes, I can walk,” said Frodo, getting up slowly.

In a fraction of the words I’ve written here, Sam Gamgee performs a beautiful redirect of Frodo’s attention. Something bad has happened. It’s over now. “Can you walk?” …And that’s that. The curtness is part of the virtue. Sometimes very little needs to be said in order to let something go.

My intent with the blog and otherwise is to focus on my immediate projects: on the work I can do now and the person I can be now, and to leave the past unexamined but for its present value.

As for whether I can walk, that question will be answered next week :)

Subscribe now

In optimism,

Daniel

Today we return to a core story of my life: that of my living personal systems. I’ve previously touched on them, but only in vague terms. This post will provide a more complete story, starting with my initial inspiration and ending with the launch of my first set of self-reflective routines.

I’m excited to share about these systems in part because they form the missing link between many disparate topics that I’ve written about, such as human-programming tools, artificial general intelligence, routines, and personal growth (links are to past posts).

Consider this part one of The History of My Living Systems. The full story would trace a thrilling rise and tragic fall, as the systems that survive today are a shadow of their former glory. And yet, I wouldn’t be writing this if I didn’t believe in the continued value of the endeavor, at least in my own life.

In any case, our story begins with a moment of inspiration.

A moment of inspiration

To set the scene, towards the end of high-school things were going great for me - I was enjoying school, getting lots of physical activity, and spending time with friends and family. But also I was feeling a growing sense of unease. A few things felt true: firstly, that the world was full of great opportunities. Secondly, that I had been “going with the flow” - working diligently at the tasks put in front of me (mainly school), but thinking little about what was actually worth doing with my life.

These feelings had been bothering me for a while when inspiration struck from an unlikely place: the psychological-thriller film Memento. In Memento, main character Leonard Shelby can’t form new memories, and yet is determined to accomplish a goal that he developed after his memory left him. His solution is to create a series of prompts, many of which are tattooed onto his body, that would remind him of his mission and the reasons behind it.

I had no intention of getting such tattoos, nor of pursuing such a vengeful mission, and yet I was inspired. Here’s someone that knows what he’s doing and why he’s doing it, I thought. And crucial to my inspiration was Shelby’s approach: he created systems that directed his attention and efforts towards his most important goals. To be clear, this is not a “Top 10 Life Optimization Tips I Learned from Christopher Nolan’s Memento” post - I barely even remember the plot, and am certain that my own systems diverged strongly from what’s depicted in terms of both vibe and approach - but, I cannot deny the inspiration!

I imagined having my own set of core directives, not tattooed onto my chest, but instead living in some digital or physical notebook that I could reference regularly. However, unlike Shelby, I didn’t have some totalizing quest handed down to me by the screenwriting gods, so I would need to figure out what I wished to do before attempting to do it. I began to imagine how I might go about discovering my values.

Planting the autopoietic seed

I might have sat down and defined some values for myself. But I figured that arriving at my values would be a longer process. And so, I might have sat down and defined the process for arriving at my values. But I figured that arriving at the process for arriving at my values would be a longer process. And so, I might have sat down and defined the process for arriving at the process for arriving at my values, but I figured that arriving at the process for arriving at the process for arriving at my values would be a longer process (and so on…)

It was clear to me that I was seeking a generic system that could grow in complexity over time while adapting itself continuously to my needs. I was still inspired to create finer-grained directives / principles for myself, but I didn’t want to start there. I wanted to build the adaptive foundation on which those directives, and any other system that might serve me, would stand.

Long after I started, I learned the term autopoietic1, meaning “capable of producing and maintaining itself by creating its own parts.” The term has been used to describe the self-maintaining chemistry of living cells, and it captures the “living” qualities of the system I wished to produce.2

A foundation of reflection

The simplest foundation I could imagine was to create daily and weekly routines that compelled me to reflect, and through that reflection to modify my routines to better serve my aims. I imagined these routines forming the seed that would blossom into more expansive systems for self-guidance.

I defined the following initial program for myself:

Complete the following activities without fail.

Daily: Write in your journal about how your day went and any thoughts you have about how to live your life better.

Weekly: Review last week’s journal entries, then update this reflection program in order to improve this program’s ability to achieve its goal: to help you lead a good life.

…And then I got to it - I purchased a journal and committed myself to daily nocturnal reflections, and from that point on the system took over.

This act of creation was quite a rush for me. I felt that I was starting a new life, both for the system and for myself. From now on, everything I do has a purpose, I felt. Now that I am committed to this program, all of my actions are contributive towards leading a good life. I felt I was embarking on the project of a lifetime - a project to systematically transform into the kind of person I wanted to be. All of my creative energy was happily directed towards this new pursuit.

…

And then the system began to grow. In the next post in the series I’ll continue the story, describing the specific results that emerged from these earlier foundations.

Subscribe now

A note to my subscribers ✍️

This week’s post is squarely in the “musings” category, focusing on time, presence, and communal practices - a sharp contrast to last week’s technical deep dive. It goes without saying, but I encourage my readers to pick and choose which posts you read and engage with. Not every post will be for everyone, but hopefully, each will be for someone. For my part, I’ll try to write clear titles and subtitles so that you know what you’re getting into.

If you do read and enjoy a post, I’m always happy to receive a “like” and/or a comment with your thoughts. Or you can reach me directly by replying to these emails. And an untracked KPI of this blog is conversations started, so please, converse away!

Without further ado, this week’s post…

An offer - Active Reading mode: if you’d like, follow the instruction on each of the green screenshots in this post as you reach them, to experience your own moments of presence.

Stretchly: a welcome interruption

In Deep Knowledge, and Three Ways to Get There, I mentioned my housemate’s fascinating computational biology “TED talk.” What I didn’t mention: every ten minutes, his slides would be interrupted by a full-screen instruction…

After a few instances, I caught a notification in the top right corner of his screen:

I was intrigued. As an unmitigated self-instructor, I had to investigate, and found out that Stretchly is free and open source! So I gave it a spin, and proceeded to have an unexpectedly great time.

The way it works: whenever your computer is open, every ten minutes your screen is blocked by a timer and an instruction. The first two times are “short breaks” (20 seconds); every third is a “long break” (5 minutes) (like the pomodoro technique if you’re familiar).

Here’s what I saw during a recent long break:

I was surprised at how much I enjoyed receiving these cues! Especially the short breaks - sometimes it’s a moment of mindfulness, sometimes a stretch… always an opportunity to step out of my projects and just enjoy life and my body and the world around me.

So, I recommend Stretchly. And perhaps this post could end there. But, I decided to go abstract and write about how my experience with Stretchly ties into time-management, psychology, and the sabbath.

On undifferentiated time

I came across Rowan Williams’ concept of “undifferentiated time” in an interview of Oliver Burkeman on Sam Harris’s Waking Up app1. Undifferentiated time is the tendency in secular society for all the time in one’s life to blend together into a uniform streak. Williams, a former Archbishop of Canterbury, presents this in contrast to the daily, weekly, and yearly schedules of the religious adherent.

Burkeman offered his own areligious characterization of the well-differentiated schedule in Four Thousand Weeks: Time Management for Mortals. I haven’t read the book, but resonated with this quote from it (ht Elizabeth Oldfield):

“freedom . . . is to be found not in achieving greater sovereignty over your own schedule but in allowing yourself to be constrained by the rhythms of community—participating in forms of social life where you don’t get to decide exactly what you do or when you do it.”

This freedom that Burkeman describes is far from my current life. While I engage in many different activities, many of which are social, I still make my own schedule and ultimately lack the kind of community-drive constraints espoused above. I’m especially aware of this fact because I grew up in a culture which imposes a strict, differentiated schedule: that of Judaism, or specifically modern orthodoxy. And the most notable division of time in Judaism is the sabbath.

Shabbat vs. project mind

The Shabbat experience is profoundly different from that of the rest of the week. To a large extent it’s defined by what you don’t do - no work, no electronics, no driving, no writing, no cooking even, and crucially, no planning for the week ahead. What’s left to do? Walk around to various locations of gathering (the synagogue, people’s houses), pray/sing, eat/drink, hang out.

I came across a quote I like in Heschel’s The Sabbath, which pitches Shabbat in opposition to the frantic utility-maximization of the work week:

He who wants to enter the holiness of the day must first lay down the profanity of clattering commerce, of being yoked to toil. He must go away from the screech of dissonant days, from the nervousness and fury of acquisitiveness and the betrayal in embezzling his own life. He must say farewell to manual work and learn to understand that the world has already been created and will survive without the help of man. Six days a week we wrestle with the world, wringing profit from the earth; on the Sabbath we especially care for the seed of eternity planted in the soul.

“The betrayal in embezzling [one’s] own life” - this resonates with me, but not because I’m constantly wringing profit from the earth. Instead, I find myself to be wringing my own specific variety of things - profit sometimes, but also board games, piano albums, plans, personal systems, health recoveries, and blog posts! The overarching category for me would be projects. In a way, my whole life feels like a project, and vast amounts of my experience are through the lens of project mind - that mental state in which I’m thinking of which actions might move my endeavors forward.

Project mind has its benefits, but it’s not quite conducive to caring for the seeds of eternity planted in the soul. Doing so requires present mind - being present with the people and world around. Present mind (or the encouragement of it) is part of what Shabbat offers, and part of what any time-constraining communal practice can offer as well.

Micro-Shabbats and Macro-Shabbats

I don’t have a Shabbat practice right now, but I do have Stretchly breaks, which, as scheduled pauses in productivity, are (sorta?) like micro-Shabbats2. When I get an instruction…

… I take it as an opportunity to enjoy the moment I’m occupying.

Is Stretchly just another tool in my project-mind’s acquisitive arsenal? Well, it could be, but it could also be a tool for presence. Connecting back to Heschel:

The Sabbath is not for the sake of the weekdays; the weekdays are for the sake of Sabbath. It is not an interlude but the climax of living.

Likewise, my computer breaks need not be productivity-serving interruptions, they can be earned moments of enjoyment - micro-climaxes of the workday, providing punctuations of pleasant differentiation.

Of course, the great benefits espoused by Williams, Burkeman, and Heschel are not to be found in software-mediated micro-pauses, and, at risk of being project-minded, I hope to introduce more macro-Shabbats (perhaps even the macro-Shabbat - Shabbat itself) into my life. But the path there may be long. In the meantime, I will be changing my sitting posture slightly more than usual. And hopefully continuing to enjoy it.

…

Want to try Stretchly yourself? click here on your computer. And if you’ve been following the instructions, here’s one more…

Subscribe now

Somewhere in a data center, billions of billions of floating-point calculations whir away on a fleet of GPUs, spitting out hundreds of thousands of tokens. The last 100 tokens are formatted and sent to a remote server for grading. A single puzzle has been completed. 119 more to go.

Finally, the result comes back: 84.6%. A new world record for Gemini 3 Deep Think! Has artificial general intelligence been achieved?

ARC-AGI is a series of benchmarks aimed towards evaluating the intelligence and/or reasoning ability of AI systems via the completion of visual patterns. Since 2019, AI systems have steadily progressed towards solving these tasks, with multiple high scores being published this month, each using a different solution architecture. But what particular approaches have succeeded, and how can we interpret those results?

These questions are answered in three parts:

1. The history of ARC - starting from the seminal paper On the Measure of Intelligence and continuing through the evolution of the benchmark and key historical solutions

2. Do we have AGI yet? (probably not) - discussing how we might interpret benchmark results in light of the AGI question

3. Are base LLMs the best? (so far, often so) - discussing the performance of different solution architectures, and how we might attribute performance gains to different research advancements

Finally, I touch on whether today’s commercial LLMs might grow to become generally intelligent.

Subscribe now

Part 1: The history of ARC

2019: ARC is created

In 2019, Keras creator Francois Chollet was concerned by humanity’s lack of progress towards artificial general intelligence (AGI). Disillusioned by the lack of flexibility in contemporary AI systems, Chollet published the paper On the Measure of Intelligence, arguing that better metrics for intelligence would aid in the development of usefully intelligent systems.

He wished to first nail down a definition of intelligence. The paper discussed many existing definitions, eventually arriving at skill-acquisition efficiency - the ability for a person or machine to quickly learn new techniques in new environments and thereby solve new problems.

To help researchers measure skill-acquisition efficiency, the paper presented a set of puzzles titled “Abstraction and Reasoning Corpus (ARC).” Each ARC task includes pairs of pixel grids, each pair representing an input and an output. The test-taker must determine the pattern between inputs and outputs and apply that pattern to a test puzzle (a lone input grid). For example, the following puzzle requires completion of a symmetrical pattern:

Each task in the dataset is handcrafted to represent a unique human-understandable pattern. If a test-taking machine could cheaply and reliably solve these without training on similar tasks beforehand, that would be a strong signal of its skill-acquisition efficiency and therefore general intelligence.

At the time of the paper’s release, no machine scored well, but commercial LLMs were right around the corner.

2019 - 2024: The rise of commercial LLMs

In 2019 LLMs were still nascent technology, with OpenAI’s GPT-3 model to be completed one year later in 2020 and released as ChatGPT in November 2022. How did this model and its successors fare against ARC?

Surprisingly poorly (though unsurprising for Chollet). Even GPT-4, released March 2023, scored only 7% when presented with textual versions of the problems (source: Jack Cole). And GPT-4o, released mid 2024, achieved only 4.5% (source: ARC leaderboard).

Notably, GPT-4 was OpenAI’s first “vision model,” but it seemed that the model’s new image-processing capabilities did not yield substantial gains on visual reasoning tasks like ARC’s.

This failing stood in contrast to early LLMs’ ability to mimic human conversation. Such mimicry was once proposed as a measure of intelligence in Alan Turing’s famous Turing Test. And yet, progress in conversation mimicry soared while progress on ARC stalled.

Alternate ARC approaches show promise

Program search (2020-2023): Chollet first hosted an “ARCathon” competition in 2020, and a pseudonymous “ice cuber” won with a 21% score. Ice cuber’s approach involved no neural networks, and instead used brute-force program search over a domain-specific language (DSL) of pre-built image-transformation primitives, attempting to compose a sequence of transformations that would faithfully produce the example outputs from the example inputs (source: ARC-icecuber repository).

As competitions continued, brute-force program-search solutions remained winning but with diminishing gains. In early 2023, Michael Hodel created an ensemble of past program-search solutions, scoring a 30.5% and capping off the era of program-search domination (source).

Transformer-based solutions take the lead (late 2023): Then at the 2023 ARCathon, a team named MindsAI shared first place, and soon after recorded a new high score of 33%. MindsAI’s solution is documented in their 2025 paper, and is notable as the first state-of-the-art deep-learning approach.

Their solution has three main components:

1. Pre-training on synthetic and augmented data: They used Michael Hodel’s RE-ARC to create 400,000 ARC-like puzzles for pre-training.

2. Test-time fine-tuning (TTT): At test time, for each puzzle, they generate a small training dataset based on augmentations of the given input-output pairs (for example, removing one pair, rotating the grids, applying a new color mapping), then they fine-tune the model on that augmented dataset. This was their core innovation, to which they attribute a 300% accuracy boost.

3. Augment, Inference, Reverse-Augment, Vote (AIRV): This part is very clever in my mind. First they use the same data augmentation engine as above to transform the test input grid into a large number of similar grids (using rotations, color mappings, etc.). Then, they run inference on each transformed grid, collecting a list of results. These results can then be remapped using the inverse of the original transformations, yielding legitimate candidate results. The candidate that received the most “votes” (that was produced the most times across all transformations) is chosen as the winner.

2024: ARC Prize announced

In mid 2024, Chollet teamed up with Zapier founder Mike Knoop to announce the ARC Prize, setting up $1M+ in incentives including a $600K grand prize. To secure the grand prize, a team would need to publish open-source code capable of achieving an 85% success rate on the ARC puzzles, using limited computational resources within a secure internet-free sandbox.

By the end of 2024, MindsAI was in the lead with a 58% scoring solution (an updated version of their previous winning solution). However, they declined to publish their solution’s source code and so did not make it to the podium. First place went instead to “the ARChitects” with a 53.5% score, and the grand prize remained unclaimed.

The ARChitects new deep-learning approach

The ARChitects overall pipeline was similar in form to MindsAI’s, but with different design decisions for many subcomponents (paper). Their candidate-generation and selection approaches were novel:

• Candidate generation: ARChitects essentially used the AIRV approach described above, but rather than sampling from a model with high temperature in order to create candidate answers, they performed depth-first search over all possible completions, extracting only those whose sampling probability exceeded a certain threshold.

• Candidate selection: Their voting system took into account the sampling probability of each candidate across each input transformation.

In sum, they leveraged the sampling-probability distribution built into their deep-learning model to a greater degree than past approaches. Still this was not enough to beat the incumbent MindsAI.

June 2024: ARC-AGI-Pub, the new public leaderboard

In parallel to the above contest, the ARC Prize Foundation released a second benchmark and leaderboard for closed-source solutions, allowing players like OpenAI to officially test their models. The public test dataset is “semi-private” - it’s not available on the open internet and therefore shouldn’t make it into any model’s training corpus (an important consideration when testing for “skill acquisition”), however the data is provided publicly at test time, and so could be leaked or trained on by a bad actor. That’s part of why there’s no monetary reward for ARC-AGI-Pub (though AI labs are surely incentivized to do well on this benchmark for PR reasons).

December 2024: O3’s breakthrough

A series of results on ARC-AGI-Pub brought the high-score for unassisted LLMs into the low 30 percents by the end of 2024. Then, a preview version of OpenAI’s O3, an early chain-of-thought-enabled reasoning model, achieved an 88% on ARC-AGI-Pub in December 2024 (source: ARC Prize blog). This was a breakthrough result, but came with some important caveats:

1. The preview model was trained specifically for success on ARC, unlike the later released O3 model which performed worse (source: ARC Prize blog).

2. O3-preview took a whopping 14 minutes and $4,560 of GPU compute per task, far exceeding the compute allowance of the open-source ARC Prize contest, and far exceeding reasonable estimates of the true cost of solving these small puzzles.

2025 - 2026: Refreshed benchmarks

Still, in 2025 the ARC Prize Foundation wished to counter the recent success of reasoning models, so they released a a new dataset called “ARC-AGI-2” and the old dataset became “ARC-AGI-1”. ARC-AGI-2 looked much like its predecessor:

… but sought to test for capabilities with which reasoning models struggled, such as:

1. Symbolic interpretation: assigning significance to visual symbols (see image above)

2. Compositional reasoning: applying multiple rules on top of each other

3. Contextual rule application: applying rules under specific contexts
   (source: launch post)

In tandem, the 2025 ARC Prize was launched with similar prizes and guidelines to the prior year’s. This time around the eventual top score was 24%, indicating that ARC-AGI-2 was indeed a harder benchmark.

Base LLMs see rising success

Despite the low scores on 2025’s ARC Prize, today’s frontier reasoning models continue to notch higher scores at greater efficiencies against the semi-private test set. Just this month, Google announced a massive 84.6% score (at $13/task) for their new Gemini 3 Deep Think model.

Is it over for ARC? Nope, the cycle continues: ARC Prize is already planning their 2026 benchmark, this time with an interactive game format (ARC-AGI-3), to be released March 25th, which will in all likelihood be more difficult for today’s LLMs. You can play the first three example games here.

Part 2: Do we have AGI yet? (probably not)

Is Chollet moving the goalposts with each new benchmark? And is Gemini 3 Deep Think already AGI?

Well, the goalposts are moving, but wouldn’t a general intelligence be capable of following a moving target? Isn’t that exactly what skill acquisition would enable? Concretely, as long as each successive ARC-AGI benchmark is achievable by humans, we can say that failures on the part of machines indicate a lack of general intelligence.

In fact, Chollet’s original paper acknowledges the need for periodic refreshes:

ARC only features 1,000 tasks in total, and there may be some amount of conceptual overlap across many tasks. This could make ARC potentially vulnerable to shortcut strategies that could solve the tasks without featuring intelligence.

[…]

…to mitigate potential vulnerability against such shortcuts, we intend to keep adding new tasks to ARC in the future, possibly by crowd-sourcing them.

What are the specific shortcuts that Chollet might consider disqualifying? The most likely culprit would be directly training on large synthetic datasets that closely mimic the puzzles themselves. Doing so arguably counts as pre-acquiring the skills that are supposed to be acquired at test time. Such pre-training has already been a hallmark of the winning deep-learning strategies since 2023, and is especially prominent in 2025 prize-winner NVARC’s solution. Also, we cannot rule out that closed-source model providers like Google might be taking similar approaches as well to improve their results on the public benchmark. Unfortunately, this makes the entire leaderboard harder to interpret.

No AGI until ARC-AGI-n is insta-saturated

A key pattern to observe is whether new ARC-AGI benchmarks saturate immediately. If Gemini 3 Deep Think scores well on ARC-AGI-3 in March, that would indicate the kind of fluid intelligence Chollet was looking for in his original paper, as the model probably hasn’t trained on environments similar to ARC-AGI-3’s games. If instead it takes time for Google and other labs to saturate the benchmark, we will be left wondering whether the gain is from general intelligence abilities, or from pre-trained skills that are aimed at beating the benchmark.

Part 3: Are base LLMs the best? (so far, often so)

Which methods are most successful against ARC, and which are most promising? Firstly, to define success…

The ARC leaderboards show a Pareto frontier between cost and score, and any submission along that frontier of low costs and high scores can be considered successful. What trends do we see?

Frontier trends

The ARC-AGI frontier has vacillated between different solution architectures - the early 2020s saw search-based program synthesis winning, while recent years saw chain-of-thought/reasoning LLMs pull ahead using transduction (producing the output grid directly instead of producing a program that produces the grid)… just look at how much of the frontier in the image above is dominated by Google’s green Gemini 3 Flash Preview and Gemini 3 Deep Think.

But even in recent months, we see a variety of architectures neck-and-neck for the frontier. Earlier this month a solution by Johan Land (code) topped the leaderboard with 72.9% on ARC-AGI-2 for $39/task. While this was soon eclipsed by Gemini 3 Deep Think, this solution has its part to play in the overall trends.

Johan Land’s briefly state-of-the-art kitchen-sink solution

Land’s solution involves a complicated pipeline, which works as follows (you can skip this numbered list if you’re satisfied by “it’s complicated”):

1. In parallel, run a bunch of state-of-the-art reasoning LLMs, some of which are tasked with producing a grid output, and some of which are tasked with code generation.

2. If there’s sufficient consensus among results produced by the above, then finish (this saves some money)

3. (if enabled) repeat the above steps using different models

4. Then run all five of these strategies in parallel

   1. Deep thinking: prompt a model to think especially deeply about the problem before producing its output

   2. Image input: Feed rendered images to a model, ask for grid output

   3. Hints: Get an LLM to analyze images and produce hints. An LLM uses these hints to produce a grid output.

   4. Structured, object-based thinking: A 3-part pipeline that compels models to describe the objects they see, then describe the transformations between those objects, then finally, in context of all previous answers, produce a grid output.

   5. Codegen: Re-run the codegen aspects from step 1, but with more models runs.

This was truly a “throw everything at the wall and see what sticks” solution, combining lots of raw LLM power with various forms of cognitive scaffolding.

For a brief moment it was the highest-scoring solution out there, surpassing Claude Opus 4.6’s score of 69.2% by 3.7%, though at 11 times the cost. Given the 120 puzzles in the semi-private test suite, Mr. Land spent ~$5K on the test run alone, to say nothing of the costs in developing his methods.

The biggest recent gains came from base LLM capabilities

Gemini 3 Deep Think outperformed Land on both accuracy and cost. Approaches with more scaffolding and/or symbolic reasoning, like Land’s, have occasionally reached the frontier, but they usually represent modest gains over incumbent transduction-based approaches, while incurring larger costs. Every approach at the top of the current public leaderboard relies heavily on LLMs even when combined with scaffolding, and improvements in the underlying models seem to cause improvements across the board. This suggests that a large amount of progress towards solving ARC (at least on the closed-source end) is due to advancements in LLMs.

Is this bad news for the ARC Prize Foundation’s mission to support novel research? Perhaps, but they have a few facts working in their favor. Firstly, LLMs did see meaningful algorithmic improvements over recent years - OpenAI’s O1/O3 models, for example, introduced the chain-of-thought reasoning approach that led to huge gains on ARC and many other benchmarks. Secondly, for the actual ARC Prize, which requires open-source models to run on limited hardware, each year’s winners have presented solutions with meaningful complexity outside of the LLMs that they leverage.

A possible synthesis of these facts: at smaller cost scales, independent researchers are incentivized to explore a variety of novel approaches, and these approaches beat out simpler LLM-based approaches and push the research frontier forward. At larger cost scales, submissions become too expensive for independent researchers, except for the occasional tech Senior Vice President (Johan Land), and the charts become dominated by big labs whose primary motive is to showcase their models as standalone products.

Where does program synthesis fit in?

A recent twitter post from Chollet reads:

I believe that program synthesis will solve reasoning. And I believe that deep learning will solve program synthesis (by guiding a discrete program search process).

But I don't think you can go all that far with just prompting a LLM to generate end-to-end Python programs (even with a verification step and many samples). That won't scale to very long programs.

Chollet has long been an advocate for program synthesis, and in January 2025 he and Knoop founded NDEA, a research lab building “AI systems that blend intuitive pattern recognition and formal reasoning into a unified architecture.”

As suggested by the tweet above, the architecture he wishes to build would not be the kind of LLM-assisted program-synthesis approach that Ryan Greenblatt used to top the ARC public leaderboards in mid-2024, which we see repurposed as a sub-component of Johan Land’s pipeline above - though by all appearances, such architectures are close to his vision. NDEA has not published anything yet, so we’ll have to wait and see how their approach differs.

However, the current frontier of ARC-AGI-2 submissions all use direct transduction without program synthesis, so it would be impressive if NDEA could present a program-synthesis architecture that’s competitive on past or future ARC benchmarks.

Will the LLMs generalize?

The ARC-AGI frontier has been populated by a wide variety of cognitive architectures, but for now is dominated by commercial LLMs - the same systems that answer our questions and write our code. Seeing whether these systems succeed immediately on the novel ARC-AGI-3 tasks will be a good test of their general intelligence abilities.

A key open question is whether commercial LLMs will grow to succeed at ARC-AGI-3 and subsequent versions. Per the ARC Prize website, “AGI [is] still unsolved. New ideas [are] needed.” But Dario Amodei, CEO of Anthropic, would disagree. In a recent interview with Dwarkesh Patel he indicated that he expects Anthropic’s training plan to scale to fully generalized intelligence (at least in domains that are operable by text). I’ll leave with his quote:

…all the cleverness, all the techniques, all the “we need a new method to do something”, that doesn’t matter very much. There are only a few things that matter.

[…]

In fact, I would point to the history in ML of people coming up with things that are barriers that end up kind of dissolving within the big blob of compute. People talked about, “How do your models keep track of nouns and verbs?” “They can understand syntactically, but they can’t understand semantically? It’s only statistical correlations.” “You can’t understand a paragraph, you can’t understand a word. There’s reasoning, you can’t do reasoning.” But then suddenly it turns out you can do code and math very well.

[…]

I think we may get to the point in a year or two where the models can just do [software engineering] end-to-end. That’s a whole task. That’s a whole sphere of human activity that we’re just saying models can do now.

Subscribe now

They say “slow is smooth and smooth is fast.” But you could also say “fast is fast,” which would be faster to say. I’ve been trying to say things fast, in order to be more smooth. Catch my drift?

… In all seriousness, I tend to talk slower than most, and I have been trying to talk faster, and I do think that doing so could increase my ease in conversations and allow me to listen better. I’ll explain my thinking and tell the story of a successful fast-talking experiment.

To wait has weight

As mentioned I tend to talk slower than most, and I reckon that sometimes I’m slow to decide and to act. But let’s focus on talking.

What does slow talking get me? Slow talking gets me space to think. This sounds good on the surface, but sometimes it’s a trap…

Subscribe now

An overthinking scenario

Let’s say you ask how I’m doing, and I’m aware of some challenge I faced this week. I formulate a response, but I don’t speak it - there’s a mental process that’s deciding whether the response is appropriate. How does that process evaluate? It needs to guess at your judgements towards me upon receiving my words. Perhaps, in this case, the result is uncertain. So I consider more options…

(Thoughts) But wait. How much do I value honesty vs. balance in the conversation? If I speak honestly, then we might just end up talking about my things, because I’m good at coming up with analyses. If you’re a question asker, I’ll just keep giving answers. And we can’t let that happen. What leads the conversation in a good direction?

At this point in the convo I offer some platitudes, but we’re off the deep end now, because now I’m in thinking mode™ and am in various ways self-conscious, and am not paying great attention to you, my interlocutor. Each question I receive pulls me further into rumination, as I carefully reconstruct and interrogate my intricate, fragile models of the world and my actions within it…

What an inconvenient dynamic! How do we prevent this!

There might be many ways to relax one’s mental grip and/or become a more engaged listener, but for me, what I ended up trying was to just talk fast, and I’m happy to say that I had a great time doing so.

A successful fast-talking experiment

I was at the Sauna with Olly when I thought, “you know what? I’m gonna talk to Olly without thinking. I’ll just observe my words as they come out of my mouth, not before. I’ll just talk fast.”

It was a wonderful experience! First of all, my approach basically worked - I was able to talk much faster. When I didn’t give myself time to think, my thoughts quieted down in a very pleasant way, and I could pay attention to what my friend was saying.

Under normal circumstances I would ask questions and think “is this a reasonable question?” and end up being self-conscious, but this time around, I was freed to focus on my conversation partner. I could feel a slight impulse to evaluate my questions before speaking, but could override that impulse, smoothing over the intrusive bump and returning to my present experience - the experience of what he was saying, and what I was saying too.

Also, as a recipient of questions I found myself to be defended against the pitfalls of overthinking. If asked a hard question, I could say “man, I don’t know, that’s a hard question,” and then later could say “I guess [xyz],” but in a way that makes it clear that I don’t fully endorse what I’m saying. That I don’t fully endorse anything I might say, as I’m not capable of coming up with definitive answers and proofs of their correctness. That I’m just talking. That it’s playful. That I can just say what’s on my mind instead of sitting there and recomputing my entire world-model every minute.

As time went on, I gained a sense of self-confidence, in my ability to have good conversations and in my ability to act well in the world. I felt that just as I could speak with little hesitation, so too could I act on the world with little hesitation, and feel good doing it.

Towards a broader practice of talking (and acting) fast

I’d like to keep practicing my ability to talk fast, especially in low-stakes informal conversations. I see this as part of a greater endeavor to reduce the burden of thinking and justifying that steers my attention away from the present moment.

Besides the benefits mentioned above, talking fast has the obvious benefits of getting more information across quickly, and it’s more collaborative too - it suggests, “we can figure this out together, and I trust you to understand my process along the way.”

Something similar could be said about acting fast in the world… That the world might reward you for acting upon it from a place of uncertainty; it might appreciate getting to figure things out together. I think that’s (metaphorically) true, but I’m not sure how I would approach a practice of acting fast. Whereas a practice of talking fast is easy to pursue.

Note: Surely there are times when slow, thoughtful conversation is better, and some people would benefit from working on that end of the spectrum (remember to consider reversing all advice).

I’d be curious to hear from my readers - do you tend to think before you speak, or not? Did anything here resonate with you? I am always happy to receive comments, or phone calls - in the latter case, expect me to talk fast.

til next time,

Daniel

Subscribe now

I’ve been thinking about how I might cultivate “deep knowledge” in my career and/or life. This essay explains what I mean by that, and tells the stories of three people around me who have done so. These stories make up the second half of the post, and are about:

1. A Civic Entrepreneur

2. A Logician and Radical Behaviorist

3. A Computational Biologist

But first, my problem statement:

I am awash in a sea of concepts

Increasingly, I feel that I know less and less. The world is a complicated place and it’s getting more complicated. Just this week, Moltbook arrived, and suddenly, machines are writing human-like posts to each other on an (ostensibly) AI-only social network.

What are the implications? No one knows, really. Perhaps this is a fun experiment that goes nowhere, perhaps it is an awakening of a new hive mind that will soon dominate the world. Most likely, it is a currently harmless experiment that foreshadows a near future of increasingly confusing dynamics between humans and growingly empowered machine intelligences.

When I say “no one knows”, I’m including the researchers and engineers who built the language models that make up the most significant part of these agents’ internal workings. I know this because I follow them on Twitter and Substack and other places. There is a healthy debate ongoing about the trajectory of AI, with some believing that AI will show diminishing returns, growing to nothing more than a helpful assistant for well-scoped tasks, and others believing that autonomous software agents will eclipse humans in all capabilities. Beliefs about timelines for such outcomes vary widely.1

…

Additionally, in my own work I feel I know less and less. I write code in high-level languages, far from the actual binary operations of my machine. These languages shift in and out of fashion while I sit atop the current wave. And whereas I previously needed to memorize many programming language features, I now rely on LLMs to write most of my code for me. With the introduction of new flagship LLMs like Claude Opus 4.5, I find fewer errors in the code they generate and am incentivized to delegate larger tasks.

Beyond that, I spend a lot of time reading blogs and listening to podcasts, and though I’m occasionally inspired by these sources, I don’t feel they leave me with a much better understanding of the world. I feel that they leave me with a fashionable vocabulary that sits underneath a discordant pile of half-understood concepts. When I attempt to articulate informed beliefs on subjects about which I’m “well-blogged”, I surprise myself with my lack of understanding.

Towards deeper knowledge

I don’t doubt that my effectiveness at work and in other arenas of life is increasing over time, but I also don’t enjoy the feeling that I lack strong understandings of the world and my impact within it. I wish to do good, and to develop convictions about how to do so. In this rapidly changing world, such footholds seem hard to reach, and yet I maintain the belief that they’re reachable, and that doing so is a sure way to have a positive impact on my community and/or the world. I maintain that belief because I see confirming examples, sometimes in the people around me (more on that soon).

I’ll call what I’m looking for “deep knowledge.” What do I mean by that? Here’s a graph of how I’m thinking:

Here’s what you’re seeing above:

X Axis - “Solidity of knowledge”: Solidity could be broken down into sub-dimensions of transience and rigor. For example, knowledge about a particular US census might be rigorous but transient. Knowledge about Internal Family Systems might be less rigorous (no offense to IFS - just a bit of a fuzzier topic). Whereas knowledge of Newton’s laws of motion would be rigorous and perpetually applicable.

In truth, the concepts “transience” and “rigor” might not perfectly encompass what I wish to describe here, and I use “solidity” as a placeholder for some more calibrated term of judgement around knowledge’s applicability.

Y Axis - “Depth of understanding”: The bands here describe levels of understanding. Note the “Twitter/Substack Zone (Danger!),” which I lament earlier in this essay. To clarify, it is specifically the ambient-scrolling form of consumption (which I find myself doing all too often) that seems insufficient for deep understanding.

The top-right zone is the “Deep Knowledge Zone,” in which one has a deep understanding of a solid knowledge area, such that one can speak, write, and think well about it. But really the whole top row is a good place to be.

Stories of deep knowledge

I will speak of three people around me who have taken their own paths in the deep knowledge direction, each of whom inspires me in different ways.

Caveat: these stories are from memory, and only represent my own perspective

The civic entrepreneur

I met Daniel at my co-living house in Brooklyn. Along with his impassioned speeches about America’s founding fathers, I learned of his vocation as the founder of Maximum New York (MNY), a “civics academy focused on governmental mechanics.”

Daniel took his own learning very seriously on the way to becoming an instructor. He seeks to read five books on each subject before he considered himself to have a reasonable grasp of it, and he’s done so with many subjects. He studied each aspect of the government in granular detail, including such feats as memorizing the NYC charter and reading every local law passed in 2023.

On our “Knowledge Zones” graph: Daniel’s knowledge would be distributed around the top layer - he knows a great deal about a subject area (government & politics) that is itself sometimes amorphous. But, within that area, Daniel’s knowledge is about as fundamental and mechanistic as it gets, and that places him in a position to pursue impactful projects from a standpoint of intellectual rigor. I am super inspired by his work with MNY and his other political activities, all of which are oriented towards promoting the prosperity of New York City and its citizens.

On learning paths: Much of Daniel’s learning happened outside of the university - after graduating, he underwent a number of career pivots and personal projects that eventually led to his current direction, and along the way he pursued much of the learning (and especially learning by doing - engaging with his local political systems) that led to his current expertise. Daniel has also expressed disatisfaction with political science degrees.

The logician and radical behaviorist

I met John during my time in college and have since struck up a correspondence in which we discuss all kinds of topics, often adjacent to human behavior and computing.

John has read an impressive amount, and, like Daniel above, has the bookshelves to prove it. John’s areas of interest are wide, but they center around logic, math, and the science of human behavior.

On learning paths: John has a diverse educational past that involved study and research across many subjects (math, physics, cognitive science, computer science) at a few institutions. During his undergraduate years he honed in on math as a primary interest and met mentors (David R. Larson and Ronald G. Douglas) whom he cites as ongoing influences.

Eventually John dropped out of a math PhD program and continued to independently learn about the history of math, logic, and behavior. Over time, he developed a perspective and opinions on such fundamental questions as “what is truth?” and “how is science done?” and “how do people function?”

I asked John how long it would take for me to understand his worldview. “About a decade,” he said, once again sending me a picture of his bookshelves. But he admitted that there was a handful of books that most captured his worldview, and I got started reading one of them: “The Analysis of Behavior (Programmed Instruction).”

Reading this book is part of what inspired me to write this essay - it was a joyful experience to learn something very fundamental about our world, and to do so not by scanning blogs that I’ll soon forget, but by carefully engaging with a system of memorable facts and theories. For example, did you know that your body has three kinds of mechanisms for responding to its environment: striated muscles that move your skeleton around, smooth muscles that modulate the dimensions of internal organs, and glands that release various juices? What a basic thing that I was unaware of!

On our “Knowledge Zones” graph: John’s knowledge is clustered at the top right, being both solid and deeply understood. John is still in the early stages of applying this knowledge to the world, but he aims to contribute to the survival of what he sees as an atrophying culture of scientific rigor, in part by developing and sharing teaching materials related to science, logic, and behavior.

The computational biologist

I got to see Pascal speak at my house’s “TED talk series.” He spoke for an hour plus about Vilya, a recently-trained neural network that can predict the structure of small cyclical molecules (macrocycles) with state-of-the-art accuracy (think of this like DeepMind’s AlphaFold, but for even smaller biochemical structures). Predicting these molecular structures can speed up the development of drugs that neutralize harmful proteins.

As with my reading of “The Analysis of Behavior,” I loved stepping into a world of more concrete truths. There’s something thrilling to me about looking at a chemical diagram and thinking “wow - this is very real, and very important, and I know very little about it, though much could be known.” And as with MNY, I am inspired by Vilya as an application of deep knowledge towards pressing problems.

On our “Knowledge Zones” graph: Vilya is a collaboration between biologists, chemists, and machine-learning engineers like Pascal, and Pascal’s own role required him to have knowledge of each subject. I believe his knowledge would occupy various x coordinates along the top of the graph, with biochemical knowledge being an especially solid component.

On learning paths: Pascal completed a computational biology PhD before joining Vilya. He considers himself to be lucky to have landed an excellent advisor and a personally-interesting research area with a high potential for impact.

So how does one get to deep knowledge?

There’s some overlap in these three stories. In all cases, the knowledge-acquirer was substantially interested in their field and spent at least a few years dedicated to learning. In all cases, this learning involved a high volume of book-reading and accompanying writing (whether for personal notes, publications, or assignments). At least in Daniel’s and John’s cases the curriculum involved both modern and historical texts. And at least in John’s and Pascal’s cases, the academic environment and key academic mentors were important parts of the picture.

Again, a big caveat - these are my own analyses! I would certainly benefit from asking these people more about their own perspectives on their learning journeys. But today it’s Wednesday, and the blog must go out.

…and how might I pursue deeper knowledge?

As to my own possible paths, I have a few open questions. One is “what to learn about?” I seem to have intrinsic interest in computation and planning (a la Methodable), and those interests point me towards the realms of AI agency (or human agency). There are also deeper directions to go when it comes to software engineering more broadly. However, these directions feel less solid when compared to e.g. biology. I actually feel excited about computational biology presently for that reason, though this is a recent interest, and I don’t know much biology yet.

The second open question would be how to approach such learning, in case I’d like to do so. The academic route is appealing, as higher ed is our primary institution for the development of deep knowledge. But I worry about ending up in the wrong lab and studying something that is not personally interesting or applicable to real-world problems. So on reflection, a key sub-question is: “how easy is it to pivot and choose your own research direction in various academic settings?” - If the answer is “very” then that seems a promising direction.

On the flip side, I seem to be surrounded by a culture of DIY learners, some of whom have indeed gone very deep in their subject areas. Considering such approaches in my own life raises further questions: how would I find/create a community of learners and support myself through a learning period? How would I maintain clarity on a learning path?

If anyone has thoughts or related experiences to share, I’d love to hear from you.

Subscribe now

Til next time,

Daniel