In this video, I ask my dot, Shaine, to start a ChatGPT thread and use my personal notes in Obsidian to propose a life vision document. The idea is to give us something we can refer back to as we work together.

Along the way, I discover something I hadn’t realized it could do: my dot sends additional context into the thread while it’s still working. You can see the moment I notice it in the recording.

The document itself also surprised me. It drew on journal entries from 2024 and reminded me how much progress I’d made on things that mattered to me. When you’re always thinking about the next milestone, it can be easy to lose sight of that.

I read through the proposal and made some edits. Then I asked how we could use it to guide our ongoing work together. I also raised a question about whether computer history could help identify when my time and attention drift away from my priorities. That access wasn’t confirmed in this recording.

This is an early experiment. Whether it helps me follow through over time is still an open question. But the reflection alone felt worthwhile, and it’s a direction I want to keep exploring.

What would you want a personal AI assistant to understand about you before asking it to help with your goals?

Jev is a new way to apply AI and a new primitive to build with.

It’s totally different from working with LLMs or chat-based models, in that you don’t chat with it. You give it a decision to make, and it produces a probability over a set of options, predicting which option has the highest probability of being correct.

I spoke to an AI engineer friend. He clearly saw the potential of Jev, but he told me that he’s having a hard time wrapping his head around what it actually is, how it works, and what you can do with this new primitive.

Jev shines inside a program where it can make bounded judgments. You supply the context and possible outcomes, while code and LLMs handle the surrounding work. Your challenge as an engineer is deciding which judgments Jev should make.

What Jev is

Working with Jev starts with two things: the information it needs to evaluate, and the questions you want answered. TypeSafe calls that information the “state.” It might be a customer’s email, a document, or a description of what’s happening inside an application.

Your inputs are: the state, and a question about that state.

You define each question using one of three types. Choice selects from a list of options you provide. Score rates something against a rubric. Noul returns a probability that a statement is true, or TypeSafe’s version of a boolean.

For example, you could give Jev a support ticket and ask which team should handle it, how urgent it is, and whether the customer is requesting a refund. You ask three separate questions about the same state, and Jev attempts to answer them for you.

Your outputs are: structured values and probabilities your software can use. Jev doesn’t generate a written explanation. You decide what happens with the structured output: route the ticket, flag it for review, or pass it to an LLM that writes a response.

According to TypeSafe, Jev evaluates the questions independently and in parallel, rather than generating its answers token by token. If one decision depends on another, the surrounding code needs to handle that dependency.

TypeSafe calls their approach to training Jev Reinforcement Learning for Calibrated Decisions, or RLCD. The goal is for the probabilities to reflect how often an outcome actually occurs: across many predictions assigned an 80% probability, roughly 80% should turn out to be true. That describes calibration across predictions, not a guarantee that any individual answer is correct.

TypeSafe bets that “AI automation will be closer to 99% machine-to-machine interaction and 1% human interaction.” What that means to me is that Jev is an AI model built first to be used by computers, within programs, and us humans only need to get involved with the last mile. We are the beneficiaries of the output of Jev.

Isn’t Jev just a classifier?

One reaction to Jev’s launch was: haven’t we had classifiers for years? We have. Choosing a category or assigning a score is by no-means a new task in machine learning. Calling Jev a new primitive doesn’t, by itself, explain why this new model resonates.

The reason it resonates is the work it bakes into training for you. One commentator on X, Matija Sosic, made the point that Jev accepts unstructured text and doesn’t require users to train a classifier for each use case. He said it’s combining the best of both worlds from classifiers and LLMs. Ian Butler made a related point: he wanted something he could drop into an application, without first having to train it himself.

Yes, Jev is a classifier, but it’s not “just” a classifier. With Jev, you describe the judgment you need and define the possible outcomes. There is still work involved: choosing the right context or “state”, writing clear criteria, testing the predictions, and deciding what your software should do with them. But that starting point is further down the road than starting by building a task-specific classifier.

As a builder, all I want to know is: does Jev make it easier to build useful products faster, at a lower cost? Even if it’s not a completely new paradigm, losing myself in a philosophical debate is beside the point.

You could call the internal combustion engine “just another engine.” We already had steam engines. That’s true at the level of categories, but it misses the more interesting question: what does the new engine make possible? That’s the question I want to ask about Jev.

Jev and LLMs: both/and

It’s human nature to jump to a binary conclusion: Jev or an LLM? It’s wise to resist that temptation. As AI engineers, the question is how we compose the primitives we have access to into something that solves a problem that is valuable to yourself and others.

Thinking in terms of either/or can narrow our choices before we’ve even understood our goal and vision for building. We have a growing set of tools at our disposal, and choosing one doesn’t mean we have to give up the others.

That’s how I think about Jev: as another primitive to work with. The wiser question is where it fits in the system we’re building, and how it can work alongside existing models and tools to make the whole system better.

Gregor Zunic’s browser-agent demo is one great example. In his description, the agent gives Jev the browser’s DOM state and a new set of possible actions at each step, with a small LLM fallback for typing. The models have different jobs within the same workflow. Jev handles the constrained choices, while the LLM supplies text when needed.

Simplified workflow based on Gregor’s linked implementation.

Gregor reports that his flight-finding demo took seven seconds and cost $0.0039. What I love about this example is that it shows how you can use Jev classifiers in partnership with LLMs.

I ran a small experiment of my own along these lines: Jev playing Pac-Man, steered by Astra. Astra, the larger reasoning model, writes a strategy for Pac-Man to follow. Jev reads the game state directly from the browser and chooses each move in milliseconds, based on that strategy. As the game plays out, Astra takes feedback from the game and updates the strategy in real time, and Jev keeps executing the latest version.

One reply described it as Astra being the brain and Jev being the body, and that’s pretty much exactly it. Where Gregor’s agent calls on an LLM for the text Jev can’t produce, here the LLM sits above Jev doing the slow thinking while Jev does the fast acting. The whole thing took about 20 minutes to build in Codex.

Where Jev fits in your workflow

Jev fits when you have a bounded set of potential decisions, like the four directions Pac-Man can move in my experiment above. You can enumerate the allowable actions or judgments, provide the relevant context, and make a decision without long chains of thought. The bounded set defines the interface to the problem; it does not mean that the right choice is guaranteed. It’s still a probabilistic model with a pre-determined set of outputs.

An apt analogy is choosing a move from a playbook rather than inventing a play in the middle of the game. The possible moves are already available. Jev’s job is to identify which one fits the situation, quickly and cheaply, so the surrounding system can act on it.

You are trading possibilities and open-endedness for fast, efficient execution. Constraining the answer space gives you speed, repeatability, and testability where an open-ended LLM does not. That is why Jev works inside a designed workflow rather than as an open-ended thinker. Jev is a focused cognitive executor. It attempts to select the likely correct decision based on the state surrounding it, which your program can then execute.

Some of the most impressive Jev demonstrations I’ve seen are browser-use demos: the model operates a web app at high speed. You would not necessarily use Jev to create a new web app. But once the app exists and you know what you are trying to accomplish within it, Jev can help operate it by choosing among the actions that are available. That makes browsers use a compelling illustration of where Jev fits.

Tradeoffs and limits

A constrained set of answers doesn’t make every decision a good fit for Jev. TypeSafe is explicit about this in its documentation of Jev 1.13’s limitations. The model can take instructions literally and miss the intended meaning. It struggles with counting, arithmetic, date comparisons, and questions that require several steps of reasoning.

The context you supply matters, too. Irrelevant detail can reduce accuracy, and adversarial instructions embedded in that context can influence its answers. Defining the possible outputs doesn’t guarantee that Jev will make the correct choice.

These limitations point to the importance of the surrounding workflow for Jev. Keep exact calculations in code, supply the information the decision needs, and decompose complicated judgments into smaller questions. Where the task calls for generating text, bring in an LLM. Composing these tools well means understanding where each one falls short and where each is strong.

A test for when to use Jev

The practical test is simple. Jev is worth considering when you can define a bounded set of decisions, supply the relevant context, and do not need the model to reason through a long chain of steps or produce open-ended text. In those cases, it can be a fast, cost-effective component of a larger system.

That does not make Jev an alternative to LLMs in every setting, nor should it be. It expands the set of primitives available to an AI engineer. The question isn’t whether to choose Jev or an LLM in the abstract, but how to combine each tool with code and workflow design to solve the problems you encounter in your actual work.

To me, the deciding factor that will make Jev either a niche product or ubiquitous is this: how easily can builders grasp how, when and where to use it? LLMs have the advantage of riding on the coattails of natural language, which humans have evolved to use as an interface. Jev and general purpose classifiers don’t necessarily provide that affordance, so the challenge will be uncovering the high value use cases and products that Jev enables in a digestible way.

A new era: what is now feasible?

In a recent Latent.Space article, swyx argued that Astra is an opportunity to raise your ambitions as a builder.

He reports that, over the past month, he built a dozen internal and personal tools, including replacements for four paid SaaS products, created a functional but incomplete alternative to GitHub and Vercel, and trained an AI to play a strategy board game.

A new model, like GPT-6 Astra, comes out, and the reflexive response is: what new things can this model do? If your aim is to empower yourself to accomplish more, that is not the right question. When you get access to a new, state-of-the-art frontier model, the best question to ask yourself is: what can I now do to create more value in my own life and the lives of others that was previously out of reach for me?

Astra has given me more confidence to attempt things. I’ve had dreams and ambitions of putting my ideas out there, but it always seemed like a lot of work. I have a full-time job and other things going on, and I didn’t feel like I had the time to develop the skills I need to make something good. As models become more capable, I feel more confident that I can make something useful. And that feeling has only amplified with Astra. That’s something that you can also take away and apply to your life.

What is now in reach

One example is Jornada P6, a portable writing computer being developed by a creator who says he has little experience in electronics or mechanical engineering. He describes explaining his idea, working with Astra to propose experiments, and carrying out the physical tests and measurements himself. He brings the results back to Astra to help interpret them and decide what to try next. The project is still unfinished, but what interests me is how they work together. He isn’t just handing over the project and waiting for the agent to finish it. He’s using the model to attempt work beyond his existing skills.

When I first looked at the MazeBench comparison between Astra and Fable 5.1, what struck me was how clearly it showed a difference that you might not notice in your day-to-day work as an engineer. MazeBench asks models to explore a puzzle world and collect gems. Looking at the maps of where each model had been, Astra had explored much more of that world. It gave me another way to think about what these models can attempt.

MazeBench exploration maps. Historical ASCII, no-tools run snapshots; conditions differ.

The maps I looked at were snapshots of particular ASCII, no-tools runs, with different resource use, changes to the game, and operator rollbacks. They illustrate different exploration paths, but they do not establish a fair comparison of the models’ capabilities. Without examining the traces, I also can’t rule out a bug or other run-specific problem affecting Fable’s result.

We’re still uncovering what’s different about Astra. If you’re mainly using it to generate code for the same kinds of tasks you do every day, you may not really notice some of what’s changed. MazeBench gives you another way to look at it, and a reason to explore what else might now be within reach.

That brings us to the question of whether frontier models have saturated coding. If you aren’t noticing much improvement in your own coding workflow, how much does that tell you about what has improved in the model overall?

Where progress may (appear to) be flattening

Martin Casado raised this question on September 6:

I think it’s worth looking at the specific task of generating working code. This was one of the earliest applications of LLMs with real-world economic value, and it’s a capability the labs have spent a lot of time developing.

One possible reason is that the people building these models write a lot of software themselves. They’re working to solve their own problem. I’m not saying that’s the only reason coding has received so much attention, but it seems like a plausible part of the explanation.

Coding is also a good domain for LLM development because you can check the results. You can run the code and see whether it passes tests or produces errors. That doesn’t guarantee the code is correct, but it gives you concrete feedback about whether it’s doing what you asked it to do.

But even if models aren’t getting noticeably better at generating working code, there’s much more to the craft of software engineering. There’s system design, collaboration, figuring out what to build, and understanding an unfamiliar codebase. There’s weighing tradeoffs, investigating why something failed, deciding what to try next, and checking whether what you’ve built actually solves the problem. Generating code is one part of that work. A model that can help you with more of the rest could still make a meaningful difference to what you’re able to accomplish.

There’s also the question of how long an agent can maintain coherence while working on a task. Even if it were generating the same quality of code as it was six months ago, it would matter if it could carry a larger piece of work through to completion. Can it keep track of the goal, remember what it’s done, and work through failures without you having to keep bringing it back on track?

That’s the kind of autonomy that makes a difference and isn’t the same thing as directly generating code: being able to complete longer tasks reliably, with less intervention. The work might be the same type of work, but you can entrust more of it to the agent at once. It’s an improvement in depth rather than breadth. So when we’re asking whether models are getting better at software engineering, we also need to ask how much of the work they can help with, and how much of it they can carry through.

Can I now afford to attempt this?

There’s another way a model can bring more within reach. It can give you access to a similar level of frontier intelligence at a lower cost. If the cost of trying something comes down, more people can afford to try it. And someone who already has a budget can use it to attempt more things.

This means you can now consider not just what can I do with the model, but also what can I afford to do with my budget. On Terminal-Bench 4.0, which tests complex work in a terminal, OpenAI reports that Astra scores 57.9%, compared with Fable 5.1’s 55.8%, at approximately 63% lower estimated API cost per task. In that comparison, a modest improvement in performance comes with a much larger reduction in cost.

Chart source: OpenAI’s Astra announcement. Estimated API cost per task.

The independent Artificial Analysis results show how big a deal token efficiency is. In its September 9 analysis, Astra and Fable 5.1 both score 53 on the Intelligence Index, but Astra costs $3.26 per task compared with $7.63 for Fable. Astra uses about 27,000 output tokens per task, compared with Fable’s 78,000. These results compare Astra at maximum reasoning effort with Fable at maximum effort with fallback. You get the same aggregate score while paying less than half as much.

Chart source: Artificial Analysis, September 9. Reasoning configurations differ by model.

ARC-AGI-3 shows something especially interesting: asking a model to reason more can make the whole task cheaper. In ARC Prize’s Standard harness, Astra at high effort scores 54.8%, at a total evaluation cost of $40,705. At maximum effort, it scores 62.7% and costs $26,098. This compares two settings of Astra itself. The benchmark’s authors explain that it solves the games with fewer actions, reducing the number of calls and tokens required. Spending more effort on reasoning can save work later. See ARC Prize’s results.

Chart source: ARC Prize. The comparison above uses the Standard harness; Provider Adapter results are a separate configuration.

These are evaluation costs, and a real project also takes human time, review, and retries. Astra isn’t cheaper in every comparison: Artificial Analysis reports higher costs than Sol at maximum effort on both aggregate indices. But the results above illustrate why price per token alone doesn’t tell you what a model will cost to use.

For me, this is another part of raising your ambitions. A project can come within reach because the intelligence needed to help you do it has become more affordable. That lowers the barrier to entry. More people can attempt things, and the same budget can potentially give them more room to experiment, learn, and try again.

A model release is a starting line

“Take our word for it that we have much, much, much more capable models coming soon.”

— Sam Altman, Axios, September 3, 2026

I feel that anticipation whenever a new frontier model is about to arrive. I want to see what it can do, what it can help me do, and what it can help us all accomplish together. But after watching this cycle play out over the past few years, I’ve come to see each release as the beginning of a new phase.

With every new model, there’s work to do: use it, experiment with it, and learn from your own experience where it helps and where it falls short. That’s how you learn what it brings within your reach.

I expect further improvements, though their timing and practical value remain uncertain. I want to keep that possibility in mind when I think about what to build and which ambitions to pursue.

We already have a glimpse of what he means. On September 8, OpenAI announced what it describes as a solution to the Navier–Stokes Millennium Prize Problem, a question about fluid motion that has remained unresolved for roughly 90 years. According to OpenAI, the proof came from an internal model substantially more capable than Astra, working through a system of coordinating agents. The company released a written proof and a Lean formalization. In the same announcement, OpenAI also credited Levent Alpöge and Tristan Buckmaster with a related result on the forced Euler equations, using an internal Anthropic model.

How will you raise your ambition?

What have you always dreamt of doing, but didn’t think you had the resources or support to turn into reality? I would encourage you to try it now, with the support of frontier AI models like Astra and Fable 5.1, and with the expectation that you’ll have access to even more capable models in the future.

My conviction is that this is real. The ability to work with these models is a skill you can develop, and the models themselves are a resource you can use to pursue the things you want to accomplish in your life. That’s something worth investing your time in.

What dream will you now transform into reality?

A practical framework for working with goal-driven AI—and being more intentional about what you want to create.

A new AI model comes out. One person thinks it’s incredible. Another wants to go back to the previous version.

How do people using the same technology have such different experiences?

In this video, I walk through a pa…

Read more

It feels like I’ve hired an extremely competent employee with expert-level skills in every digital domain. The limits feel increasingly like my own imagination and how clearly I can specify the goal I want to accomplish.

Astra has been a long time coming, and I wanted to revisit the Codex setup I’d carried over from the GPT-5.x era. Here are six configuration changes I’ve made that, in my limited time using it, have helped Astra shine and that you can make too.

Free preview: Tips 1 and 2. The installation commands, multi-agent setup, and explore-then-implement workflow tips are for paid Attention Heads subscribers.

1. Turn on persistent context

Astra’s experimental context-management feature lets it save notes across context windows and search earlier messages and tool calls. That means a long task does not have to depend entirely on one giant, fragile summary that loses information the longer the thread runs.

In your Codex config.toml, add (or ask Codex to add):

[features]
context_management.experimental_mode = true

In my experience, this makes Astra feel like it can work beyond its active context window. This makes a long-running task feel more coherent, even as it spans multiple context windows.

That is a much better memory model than asking an agent to keep re-summarizing its own life. I’ve only used it for a short time but I imagine it’s the sort of mechanism that gets stronger as you go along. The feature is experimental but OpenAI said it will become the default in the coming weeks.

2. Remove the context that no longer deserves to be there

Astra is better at following instructions. That is excellent news until you remember how many old instructions you have accumulated.

Skills you forgot you installed. An AGENTS.md written for a previous model. A prompt that made sense six months ago and now tells the agent to ask permission three times before it opens a file.

Run this context audit prompt in Codex before your next session with Astra:

Audit the skills and AGENTS.md files available to this project.
Identify redundant instructions, conflicts, stale model-specific guidance,
and unnecessary approval pauses. Preserve real safety and publishing gates.
Propose surgical edits; do not change anything until I approve them.

The goal is not to make your agent less constrained. It is to make the constraints legible, current, and proportional. OpenAI specifically recommends auditing accessible skills and instruction files because Astra is sensitive to the guidance that is loaded into its context.

A smaller, clearer operating manual gives Astra more room to do the work you actually asked for and removes the bloat accumulated from GPT-5.x era prompts.

One example from my instruction cleanup

My GitHub Actions repair skill used to require a fresh approval even when I had already asked the agent to fix the failing checks. The audit replaced that blanket gate with a rule that distinguished a repair request from a request for diagnosis.

The revised rule still asks about material decisions or consequential actions outside the existing authorization. It removes a redundant pause while keeping the scope boundary. This is an actual change from my September 5 audit; I have not measured its effect on task speed or quality.

🔒 Four more ways to give Astra real agency

The next four are where the model becomes less like an excellent individual contributor and more like the person running a small technical team.

A practical companion for paid subscribers

I’ve put together an Astra Field Kit to accompany this piece: an interactive configuration builder, example configs, prompts, and a checker to help you catch mistakes before you use them. Paid subscribers get lifetime access to this edition of the kit.

Read more

Psychology has this concept of crystallized vs. fluid intelligence. Crystallized intelligence is knowledge and memory gained from experience. The things you learned from a book and can write down in an exam. Fluid intelligence is the ability to think in the moment and solve problems independent of past learning.

Today’s AI systems are almost entirely crystallized intelligence.

When AI systems gain fluid intelligence, we’ll see the next step change in capability.

Continual Learning’s Promise

The concept of fluid intelligence in AI research is known as Continual Learning. An AI model currently goes through ~two learning phases:

1. Pre-training - this is like general education, where the model is fed massive amounts of internet and other sources of data. From this the model learns the patterns of language inherent in that data. Also known as “next-token” prediction.

2. Post-training - this is like “on-the-job” training, where AI researchers use reinforcement learning (RL) techniques to give the model specific skills. The most effective forms of RL today are in domains where there is a verifiably correct answer. Examples are coding and mathematics; and it’s no surprise that those are dimensions AI has made the most progress on.

One could argue that there is some element of continual learning in today’s AI models. People talk about in-context learning, which is an established capability. In-context learning is where a model learns new information within the space of its context window. Most frontier models today have context windows of 1 million tokens and reportedly have parameter counts on the order of single digit trillions, though the exact parameter count is not shared publicly by the AI labs. Even given a 1 million token context window, it’s been shown that performance begins to degrade after ~250k tokens.

A person using an AI model has influence over a scratchpad of ~250k tokens; while trillions of frozen weights stay untouched.

Think about today’s AI models like an iceberg. The vast majority of the knowledge they contain is frozen beneath the surface of the water. You have influence over a tiny fraction of their knowledge at the very tip of the iceberg.

Continual learning is when your interactions with the model affect the whole iceberg.

In a way that doesn’t cause the model to forget what it already knows a.k.a “catastrophic forgetting”.

The reason continual learning is the holy grail of AI research is because the implications are immense.

The Neo-Labs Pursuing Continual Learning

Several prominent AI researchers who left frontier AI labs like OpenAI and Google DeepMind are expressly pursuing continual learning.

“A human being lacks a huge amount of knowledge. Instead, we rely on continual learning.” - Ilya Sutskever on The Dwarkesh Podcast

SSI has not confirmed a model release, though there are rumors. If it releases a model capable of continual learning, everyone will be looking to see what type of CL: a system that continuously updates its model weights without forgetting what it knows, or a thinner feedback layer that may still be useful but leaves the underlying model mostly unchanged. The difference is whether SSI has solved the barriers to continual learning that I will return to at the end of this essay.

SSI describes itself as a straight-shot lab, built around one goal: safe superintelligence. Taken literally, it means they aim for their first release to be a safe, superintelligent AI model.

Jerry Tworek, who leads up Core Automation, spent seven years helping build OpenAI’s reasoning stack. His departure from OpenAI does not prove he has abandoned scaling, but it is suggestive: he may believe bigger transformer models, trained once and deployed statically, may not be enough to reach AGI. Core Automation is his wager on what has to come after that recipe. It’s also possible he just wanted a change and to try something new.

“Core Automation is a lab created to build models that continuously learn and learn from deployment.”

Tworek’s point in the Sequoia conversation is that today’s options are inadequate: in-context learning is too limited, while repeated fine-tuning causes the catastrophic forgetting mentioned. The goal is to bring learning into use at inference time and to dissolve the gap between test time and training time, so that every test is also a train.

Ronak Malde’s Trajectory is taking a more practical approach: it uses real-world traces, corrections, retries, and edits to improve and redeploy models. Trajectory is pursuing something more like continuous improvement: a feedback loop tailored to a specific organizational problem, rather than a general-purpose model that learns continuously at inference time.

How It Might Work: Learning at Multiple Timescales

The most interesting research on continual learning is dissolving the clean line we’ve drawn between training and inference. Today, a model is trained once, its weights are frozen, and then it encounters the world through a context window. New information comes in, but only at the tip of the iceberg. The context changes while the underlying model does not.

Nested Learning and Learning, Fast and Slow both point toward a different picture. They treat learning as a spectrum of updates happening at different speeds. Some adaptation can happen quickly, in the context of the immediate task. Other lessons can be consolidated slowly, into the model’s deeper weights. From that perspective, inference is no longer simply the period after learning has ended. It becomes part of the learning process itself.

Google’s Nested Learning paper makes this explicit. It treats a model not as one fixed object, but as a network of nested optimization problems, each with its own memory and update frequency. The paper’s “continuum memory system” replaces the familiar short-term/long-term split with many layers of memory that update at different speeds. Its HOPE architecture is a proof of concept for that idea: some learning loops can react to the immediate context, while others preserve and consolidate what they learn over longer periods.

Source: Nested Learning, Figure 1 (PDF p. 2)

Learning, Fast and Slow arrives at a similar distinction from a different direction. Its Fast-Slow Training approach treats the optimized context as “fast weights” and the model’s parameters as “slow weights.” The fast weights are not literal parameter updates at inference time; rather they are an evolving “population” or set of prompts that can rapidly absorb task-specific feedback, while reinforcement learning makes slower updates to the model’s actual weights. The point is: don’t choose between a context window and a retraining run. Let fast adaptation and slow consolidation work together, so the system can keep learning without simply overwriting what it already knows.

Source: Learning, Fast and Slow, Figure 2 (PDF p. 3)

There is a loose analogy here to Daniel Kahneman’s System 1 and System 2: one mode is fast, immediate, and responsive to the situation in front of it; the other is slower and more deliberative. The analogy is imperfect, but it gives us a way to see what these researchers are doing. They are trying to build learning systems with both a fast layer that can adapt in the moment and a slower layer that decides what should become part of the model itself.

The iceberg is a useful picture here. Fast learning happens at the tip, in the context of the immediate task; slow learning consolidates experience into the much larger mass below the waterline, in the model’s weights. The two timescales are not competing approaches. but rather one spectrum. They are complementary parts of the same learning system.

Five Implications of Continual Learning

1. Models with a genuine learning history - today’s models are like amnesiacs. They don’t carry state from session to session beyond writing down text in a markdown file. Dwarkesh Patel has a great analogy for this in his essay on the era of continual learning: “Imagine if the way students had to learn to play the saxophone is that one student tries to play it from a cold start, then after her first session, writes down a bunch of notes, then the next student waiting outside the music hall who’s also never played the saxophone reads all their notes before trying to play, and so on.” That’s what text file based memory is doing today. With real continual learning, models will absorb new skills into their model weights rather than trying to hack together markdown notes in their context.

2. Emergent skills over long-horizon tasks - imagine an AI model that can work continuously, without sleep, without a break, with infinite context. That same model, with continual learning enabled, could learn from every interaction, correcting itself as it makes progress. Now extrapolate that process out to years. Decades. Centuries. We simply don’t know what kind of capabilities might emerge from an immortal machine intelligence that can work eternally. (Or as long as the GPUs keep whirring).

3. Deep personalization — Deep personalization is where continual learning becomes intimate. A model that learns from every interaction with you would not just remember a prompt or a profile. It could build up a living context about your life: the experiences you have, the information you consume, the conversations you conduct, and the people you care about.

   Over time, it could become a kind of reflection of your own experience, with a memory far more complete than any person’s. Instead of starting from scratch each time, it would meet you with the accumulated context of your life and update that understanding as your life changes.

   For me, this is one of the most exciting opportunities in continual learning: AI that does not merely become more capable in the abstract, but becomes more useful in the texture of our real human lives.

4. Capability that compounds with human expertise — Continual learning could make expertise compound. Today, even when a brilliant person works closely with an AI system, most of what they teach it disappears when the session ends. A continually learning system could retain the judgment, corrections, and problem-solving habits it acquires from experts.

   That creates a different relationship between people and AI. Nobel Prize winners working with a system on hard problems would learn from the system, but the system would also learn from them. As more people contribute their hard-earned knowledge, the system’s intelligence could compound over time. Not by replacing human expertise, but by absorbing and connecting it.

5. Fleet intelligence and collective learning — Deep personalization is a one-to-one relationship: one person, one deployed model, and a learning history built from their life together. Fleet intelligence expands that relationship outward. Every deployed instance could learn from all the people using it, then share what it has learned with the wider system.

   That turns continual learning from a private advantage into a many-to-many network. Experience gathered in one corner of the globe can improve the system everywhere. If the system can reliably absorb, test, and distribute those lessons, learning would compound exponentially.

What Continual Learning Means for Society

Labor and expertise

If the five implications above turn out to be true, the question is not whether AI will replace human expertise. It is how human expertise transforms. A continually learning system may be able to absorb far more of the practical, mechanistic work we humans do: implementation, iteration, recall, and the accumulation of technical knowledge.

That could leave people with a more distinctly human role. Humans would still provide creativity, originality, judgment, and a sense of what is worth doing. We could become less like implementers executing a machine’s logic, and more like wise guides setting direction for systems that can carry it out.

At its best, that would relieve people of the need to act like machines. It would make more room for the parts of life that are most human: relationships, wisdom, personal growth, and the freedom to decide what kind of world we want to build rather than getting lost in the act of building.

Lock-in and switching costs

A model that has learned from you for three years is not interchangeable with a fresh one. Its value is not only in the model you can download or the profile you can export. It is in the accumulated context: the context it has built about your work, your preferences, your relationships, and the lessons it has learned through thousands, or even millions, of interactions.

That creates a new kind of switching cost. Moving to another AI system may mean leaving behind not just a tool, but a type of entity that knows you. It would be more like leaving a relationship. The upside is that the assistant becomes genuinely useful in a way a generic model cannot be. The downside is that the companies with the most users, data, and compute could see their advantage compound as their systems keep learning about you faster than everyone else.

This could be more extreme than the switching costs we see with cloud providers. With cloud infrastructure, a company may face the cost of moving data and rebuilding integrations. With continual learning, the thing left behind is an AI system’s accumulated understanding of the business itself: its decisions, edge cases, working habits, and hard-won context.

The same is true personally. Moving from a model that understands the details of your life to a fresh model that knows nothing about you may eventually feel less like switching software and more like starting over with someone new. That depth of lived context could become so valuable that, once it exists, switching no longer makes sense.

Concentration versus distribution of power

Lock-in does not only affect individual users and companies. It could reshape the distribution of power across the AI industry. The first lab to build a system that learns continually may gain an advantage that compounds with every new user and every new interaction.

That concentration would not necessarily require an explicit attempt to dominate the market. It could emerge from the mechanics of continual learning itself. If switching means abandoning years of accumulated context, then the system that gets there first becomes harder to leave and harder to compete with. The result could be a powerful gravitational pull toward the lab that first makes continual learning work at scale.

If that dynamic produces durable market power, governments and competition authorities may need to step in. Existing antitrust and anti-monopoly rules may need to be applied to a world in which a company’s advantage is not simply its software or data, but a learning relationship that customers cannot realistically take with them.

A Partition Between Public and Private Data

The promise of deep personalization comes with a tough data security question. A model can only become genuinely useful by learning from intimate details of a person’s life. But if it continually updates from those details, what prevents private information from leaking into the public data accessed by everyone else?

This will have to be solved. Continual learning systems will need hard boundaries between private memory and shared learning: strong access controls, data minimization, user consent, and designs that keep personal experience local or isolate it from the model’s broadly shared capabilities. Otherwise, the same mechanism that makes an AI feel like a trusted companion could turn it into a surveillance system that never forgets. This is a hard problem to solve.

What Remains Unsolved

Continual learning is the holy grail because it remains unsolved. The promise is enormous, but a system that changes from experience introduces problems that today’s largely frozen models avoid.

• Forgetting: new learning can damage old capabilities.

• Stability: a model must stay adaptable without becoming erratic or manipulable.

• Evaluation: we do not yet know how to prove that a system has learned robustly over months or years, rather than merely improved on the last task.

• Safety: if a model can learn from the world, it can also learn from bad data, adversarial inputs, and misaligned incentives.

The challenge is to help models learn, and un-learn, as they adapt and change with an evolving world.

Continual learning is the holy grail of AI not only because it would make models more powerful, but because it would change what they are: from static artifacts into intelligences with a history. Like the Grail of Arthurian legend, it is not a prize that can be seized by scaling what already works. Reaching it requires solving the hard problems of memory, safety, and judgment, and making sure that a system that can keep learning is worthy of what it is allowed to learn.

1. Orchestrator - Plans, reasons and specifies the change

2. Implementer - Carries out the typing

3. Advisor - Reviews the change and advises on improvements

On August 1st, I published sol-advisor. It’s my most popular open-source project. It already has 1,300+ GitHub stars in four days! Before sol-advisor, I created fable-advisor.

In this Agentic Engineering Field Note, I will show you how it works, some gotchas to avoid, and why I think this pattern is more powerful than it seems.

After reading, you will feel comfortable implementing this pattern and it will make your agentic engineering implementations more efficient.

How it works

Both plugins ship the same pattern, with the models from each provider that I believe fit the roles best. One session orchestrates and holds the thread. Bounded implementers do the typing. An advisor with a clean context decides whether the work is done.

Read more

Three sentences. Models from two different AI labs. A repo connected. No switching apps and no re-explaining. A natural conversation that puts your knowledge to work.

What Buzz Actually Is

Buzz is a self-hostable Slack-inspired group messaging app from Jack Dorsey and Block where humans and agents share the same room. The agents are powered by agent harnesses like Codex, Claude Code and several others. You interact with both the humans and the agents in the workspace the same way you do in Slack. The channel is the shared memory: every agent can search and read the conversation history, so context carries across harnesses; and hence across different vendors’ AI models.

Buzz is open-source and a Nostr relay under the hood. That means every message, reaction, workflow step, approval and git event is a cryptographically signed event in a single log. Same schema whether the author is a person or an agent.

Agents get the same affordances that humans get: DMs, channels, huddles, and a git integration. Every channel has a canvas where you can set system instructions that the agent takes as context.

After a few days using it I’d describe it as: an organizational-level meta-harness where you can coordinate with agents across vendors. It’s what Claude Tag wants to be, but for any AI agent.

The Buzz repo is open-source if you want to go deeper.

The Terminal’s Real Problem Was Single-Player

The terminal was never the right interface for agents. It was designed for a programmer to send deterministic commands. Input then output. Predictable.

Humans need a conversation. They need to have context. They need to know what’s happening in the room around them. Humans need the opportunity for serendipity and the unexpected.

Working with agents should be a team sport.

Buzz turns working with agents into a multi-player game. It’s more than a 1:1 conversation. The 1:1 chat interface didn’t pre-empt the terminal’s problem. It’s the terminal with a prettier UI. It’s still one human, one machine, one thread. It’s single-player. We need a different interface.

That interface is a room.

Group chat may not be the final interface for agents, but it’s the first one with the vibe of a real relationship: we work with thinking machines the way we work with thinking humans.

In a room, context carries across harnesses as you switch. The conversation acts like shared memory between agents. Your knowledge and the agent’s work aren’t trapped in a vendor’s siloed harness.

Group chat is how agents and humans get in the room together.

How I Know They Meant It

Buzz supplies a 133-line base prompt (base_prompt.md) that’s compiled into the harness binary. The main function of the base prompt: in a chat room, a human’s thinking is naturally visible. What we type is what we think. An agent’s is a black box, so the system prompt makes publishing the agent’s thinking a hard requirement:

“If your turn produced anything worth knowing, you MUST publish it. … Your reasoning and tool calls are invisible — a result … exists only if you published it. Ending that kind of turn without a message is a silent failure.”

Yet, the system prompt also bans idle acknowledgement. It prompts the agent to share substantial thinking, not idle commentary. The prompt literally prohibits “Got it,” “Confirmed,” “Standing by,” and “I won’t reply again.”

Wish we had a rule for that in our company Slack 😅.

None of these rules would exist for a simple chat box. Every rule is about being a good colleague in a room full of people (and agents).

Agents react 👀 when they’ve seen your message and 💬 while working. It’s emoji as progress indicator.

On context overflow, when the context limit of the agent’s window is reached, the agent triggers auto-compaction with its own handoff summary. It tells itself: what the task was, what you accomplished, key decisions made, what remains, and one concrete next step.

Each agent that you create has a text box for agent instructions. It allows you to set the agent’s persona, role and personality. So every agent can have a different job and area of expertise.

The design of Buzz points to something important about working with AI, which I first heard articulated by , CEO of Every, on ’s podcast:

“There’s always a higher frame for us to go.”

Buzz and the group chat interface let you raise the frame for an entire team, humans and agents together, not just for a 1:1 relationship.

Who This Is For, and Who It Isn’t

Buzz is an alpha product. It has rough edges, but the potential is clearly there. A few days in, that’s already obvious.

I struggled to successfully connect it to my Tailscale network and my Claude Code agent silently failed at first. The setup guide below walks through these challenges for Attention Heads paid subscribers. You’ll get some value out of that section if you self-host Buzz.

If someone asks me who should try it, my answer is: solopreneurs and small teams who are already using AI agents. Codex, Claude Code, OpenCode, Amp, Goose, Hermes, for example.

It’s not enterprise-ready, and that’s to be expected. It was released a couple of weeks ago.

If you’re not already working with AI agents in an agent harness daily, I suggest starting there, before you jump into Buzz.

A huge bonus is that it’s open-source and free to use if you self-host. You can make it yours with the help of your agents. And because the agents run through the harnesses you already have installed, you don’t pay extra for inference; your existing Claude or ChatGPT subscription covers it.

Where This Goes

Buzz suggests that agent-human collaboration is best conceived as a team sport.

Agents work better when you treat them as co-workers with accountability rather than a “haunted cron-job”, as the Buzz repo describes it.

The group chat setting with agent harnesses underneath is a real meta-harness, a brand new category of product that didn’t exist before the AI age.

How we work with these systems changes how we think. The interface is more than a simple interaction point; it shapes how we evolve alongside machine minds.

No big deal.

Buzz is so good, I’m about to invite my wife to the community so we can run our household alongside Honey, Bumble and Fizz.

Next up: a section for paid Attention Heads subscribers where I walk through the technical details of my actual setup that you can borrow to set up your own self-hosted Buzz relay.

🔒 How I Actually Set This Up

Here’s what you’ll have at the end of this: your own Buzz relay running on a machine you own, reachable only by you, with agents from two different AI labs sitting in the channels answering to their names.

Budget an hour. It’s about twenty minutes of actual work and forty minutes of the four things that went wrong for me. I’ll tell you where each one is coming.

You need three things before you start: a Linux box with Docker on it (that’s what I use though MacOS or Windows should work too), Tailscale running on both that box and your laptop, and a subscription to Claude or ChatGPT that you already pay for. That’s it.

Read more

I’m out walking with my two boys on vacation. I open Termius on my phone and ask Claude to brainstorm an idea I had. Termius taps into my MacBook Pro back at the hotel, and to an Nvidia DGX and Linux box that are all the way back at home in Atlanta.

That’s the end state: every computer you own becomes one private machine your agent lives on, reachable from your pocket, with zero ports opened to the internet. Here’s exactly how to build it.

What you’re building

Three layers, built in order:

1. The network — Tailscale joins all your machines into one flat, encrypted private mesh. ~1 hour.

2. The access layer — SSH aliases + Termius so you (and your agent) can reach any machine from anywhere. ~1 hour.

3. The agent — LifeOS on the always-on machine, taught about its own network. An afternoon.

Read more

The reason I’m a day late is because I went out on Lake Allatoona yesterday with my in-laws and 3-year old son. My 3 year old watched his two older cousins ride on a tube. After he saw that, he was not taking no for an answer and was going to ride that tube too. I was not prepared to ride a tube with him off the back of the boat but we both had so much fun.

Also, been trying to persuade my lawyer wife to let me download ChatGPT Work onto her laptop. It’s pretty great.

Last week was a huge one. We had important moves from OpenAI, Anthropic, Meta and SpaceXAI. Let’s discuss what I paid attention to.

1. The Fable Endgame

What happened

I recommend that Claude users max out their Fable usage on Monday because Fable was slated to leave the sub on July 7th. Turned out not to be the best advice, because a day later Fable access was extended for all subscriptions until July 12th. Then, on July 12th, Fable access was extended again until July 19th.

Why it held my attention

I am curious about Anthropic’s end-game with Fable. My perception is that people believe they have an underlying motive. I’m not so sure about that. My feeling is that they are legitimately trying to serve the best model they can, at the best price, with the compute they have available.

What I’m carrying forward

A take I had that went pretty viral on X is that they will remove Fable from the subscriptions once Opus 5 is ready. I still think that’s likely.

A lot can happen in a week, and the new end-date for Fable is July 19th so Opus 5 could be ready by next week. The timing is quite realistic.

2. The Summer GPT-5.6 Sol-stice

What happened

GPT-5.6 was a long time coming. We knew it was ready, but the US government was holding it back for security testing. This time OpenAI announced the release date a couple days ahead of time. The hype ran all week. That’s the fun part, isn’t it?

And on Thursday, July 8th, GPT-5.6 was released in three variants—Sol, Terra and Luna.

It tops the Artificial Analysis Agentic Coding index, and it cheaper than Claude Fable at $30 per 1mil output tokens compared to $50.

OpenAI was also released along with GPT-5.6 some pretty major updates to their desktop apps with ChatGPT Work. It’s similar to Codex but designed for knowledge workers.

Why it held my attention

I was keen on seeing if GPT-5.6 could stand up to Fable. It’s definitely nearing the ballpark, but after a few days of use Fable is still superior in terms of raw capability. That’s no surprise to me.

The area where GPT-5.6 Sol competes with Anthropic’s models is on cost and efficiency.

That, and OpenAI is living up to their name in making GPT-5.6 much more accessible for things like ML/AI development.

What I’m carrying forward

I don’t think GPT-5.6 is OpenAI’s answer to Fable. That will be GPT-6 and the rumors are already starting to leak about a new model with a brand new pretrain. The group chats I’m in say it will be ~10T param models and be the true response to Fable. View this in light of the fact that “The Information” reported OpenAI discovered a way to more than halve inference costs.

3. The Multi-Vendor Orchestra (Grok 4.5 + fable-advisor)

What happened

SpaceXAI released Grok 4.5 on Wednesday. It was trained in conjunction with the Cursor team using Cursor data. It’s surprisingly good, nearing Opus-class model benchmarks at a fraction of the price.

Why it held my attention

I created a Claude Code Plugin that uses Fable as the orchestrator model and combo of GPT-5.5, Opus 4.8 and Sonnet 5 as implementers. It’s not using GPT-5.6 and Grok 4.5 in place of GPT-5.5 and Sonnet 5, respectively.

What I’m carrying forward

It was a surprise that the SpaceXAI + Cursor tandem produced a model of this quality, at this cost, this soon. I used it and it’s quite good. SpaceXAI with the talent of the Cursor team is definitely back in the game.

It was a real scorcher in Atlanta. That’s typical for us on the 4th, but this year it was extra humid. My wife and I were meant to run the Peachtree Road Race for the third year in a row. It’s the largest annual 10K in the world. We had to bail this year because the kids were sick. We’ll be back at it next year.

We had a few surprises this week in the world of AI. Here’s what I paid attention to.

1. Sonnet 5 Ships

https://x.com/daniel_mac8/status/2072036584635977819

What happened

Anthropic released Sonnet 5 on Tuesday. On the benchmarks it’s a step down from Opus 4.8 almost across the board, and much of the early reaction wrote it off for exactly that reason. The post above made the counter-case: pair it with Claude Code Dynamic Workflows—set /model to Sonnet 5, set /effort to Ultracode, and any complex task kicks off a dynamic workflow.

Why it held my attention

People were pretty down on the Sonnet 5 release. I get it. But a pattern I’ve been experimenting with a bunch—and what makes Sonnet 5 relevant for me—is this: how can you give a smaller, cheaper model to a larger, more expensive, more capable model as a tool?

What I’m carrying forward

The second item in today’s post gets deeper into this, but the best way to use a model like Sonnet 5 is as an implementer orchestrated by a larger model. This is a winning pattern, especially as mid-sized models keep getting more capable. I’m not sure it will always be necessary, but it’s a winner while it is.

2. The Fable Advisor Pattern

https://x.com/daniel_mac8/status/2073366129737683146

What happened

Fable 5 came back on Wednesday, as predicted in last week’s edition. Its return made a pattern I’d been sketching all the more relevant: Fable as the advisor doing the high-leverage judgment, Sonnet or Opus as the implementer doing the token-heavy work. By Friday I’d distilled it into the best way to use Fable right now, and on Saturday I shipped it as a working artifact: fable-advisor, a free, open-source plugin that runs exactly this orchestrator pattern (that’s the post above). Underneath it all, the cost tension is still live—Fable is expensive enough that I’ve been token-maxxing my Claude sub ahead of July 7th.

Why it held my attention

Fable is the most capable AI model ever created. It’s also the most expensive. You don’t need to blow your token budget by using it for every LLM call. Use it as an advisor and distill its expensive intelligence for smaller models, in real time.

What I’m carrying forward

I do wonder whether this will be an effective pattern forever. I suspect it won’t be, and that eventually there will be a single unified system that does it all for you, without any engineering. But for now I’ll keep looking for ways to optimize my token usage, and you should too.

3. GPT-5.6 Closes In

https://x.com/daniel_mac8/status/2072646722283622591

What happened

Details on GPT-5.6 Sol surfaced Thursday: 60% of Fable’s cost ($30 vs. $50 per 1M tokens), a reported win over Fable on TerminalBench 2.1, and availability through the ChatGPT subscription. The Information reported that OpenAI found a way to cut inference costs by more than half, which is what makes that pricing possible. Then the timing sharpened: Leo says OpenAI plans to release GPT-5.6 on Tuesday, July 7—the same day Fable access is set to leave Claude subscriptions—even as Anthropic signals that Fable is coming to Claude subs once compute capacity allows. There may also be a Sol Ultra tier aimed squarely at Fable.

Why it held my attention

It will be interesting to see what Anthropic does based on how strong GPT-5.6 is. On one hand, Fable is expensive and isn’t guaranteed to be available on consumer subscriptions after July 7th. On the other, many people and organizations simply want to use the best model. They are cost insensitive.

What I’m carrying forward

If I’m right about July 7th being the release date for GPT-5.6, I can’t wait to see what Anthropic does about Fable access that day. My hope is that they extend it to Claude subs, but I’m not that confident they will.

We spent the last hundred years training ourselves to become computers. The modern school system was designed to produce cogs in a machine, and then cogs in an abstract information machine: humans who could calculate, categorize, file, and follow procedure with machine-like reliability. We got very good at it. So good, in fact, that we eventually built actual computers. Now those computers are becoming excellent at the very thing we spent generations imitating.

This is the moment everyone is freaking out about. If the machines can do the machine-work, what is left for us?

I listened to on his “Art of Accomplishment” podcast recently, in an episode about how to succeed in the age of AI, and it is sage advice. What is left for us is everything we set aside in order to become machines in the first place. Not a bad promise.

I want to walk through his ideas, because it is very much in the lane of what Attention Heads is all about: the age of AI is not primarily a technical challenge. It is a contemplative one.

The ladder

Hudson traces a simple progression. For most of history, societies rewarded physical strength. Then strength was mechanized, and they rewarded skills. Then skills were industrialized, and they rewarded intellect—the credentialed, symbol-manipulating, test-scoring kind. Each rung of the ladder was the scarce asset of its era, and each one was eventually commoditized by technology.

Intellect is being commoditized right now. You can rent it for twenty dollars a month, and the price is falling.

Wisdom is the next rung, and Joe’s feeling is that it differs from every rung below it. Strength, skills, and intellect are assets you possess. Wisdom is a way of being. That difference in kind is the entire reason it cannot be outsourced. The obstacle is a matter of principle rather than a limitation of the current models, and no amount of scaling or algorithmic breakthroughs will cross it.

What wisdom actually is

Hudson tells a story about his daughters. While his nieces were learning three languages by age six, his own kids weren’t yet reading in one, because he and his wife had bet their education on something else: teaching them to be themselves, to have their own will, to know how to be with what they feel. He worried he was ruining them. A friend settled it with one sentence.

Joe’s friend said:

You can hire someone to speak a language for you, but you can’t hire someone to be yourself.

That sentence is this whole essay, really. Everything below it is unpacking.

Because notice what AI can now do: write your emails, speak your languages, summarize your reading, draft your code. And notice what it cannot do. It cannot make the decision about whether you’re going to scroll all day. It cannot make the decision about whether you’re going to get addicted to it. It cannot decide what kind of relationship you’re going to have with your Mom, or your spouse, or your kids. Those decisions have no market, no API, and no delegation path. They are yours alone, and the sum of them is your life.

Joe gives wisdom three working markers.

1. Can you see the patterns behind yourself and others?

2. Can you feel the difficult thing instead of fleeing it?

3. Can you walk into what everyone else is avoiding because they’re afraid?

The part Joe never makes explicit

Read that list again slowly. Seeing the patterns behind yourself. Feeling the difficult thing. Walking toward what is being avoided.

That is a description of contemplative practice. It is arguably the definition of contemplative practice. At least what I would call “real” contemplative practice. It’s the practice of becoming wiser.

Seeing your own patterns is what every tradition of self-observation trains, whether you call it vipassana, self-inquiry, or examination of conscience. Feeling the difficult thing is the entire point of sitting still when everything in you wants to move. Walking into what others avoid is the renunciation at the heart of any serious path. The hard conversation is a koan you live instead of contemplate.

For most of history, contemplative practice looked like opting out of the economy. The monastery was where you went when you were done competing. Joe’s argument, translated, is that this has now inverted: the meditation cushion turns out to be the training facility for the one capacity the economy cannot buy.

And there is a deeper layer here that I find genuinely stunning. I’ve written before that enlightenment is the realization that thoughts are representations—useful, powerful, world-building representations, but never the absolute truth—whether those thoughts arise in biological minds or silicon ones.

Crucially, the thoughts you have about who you are turn out to be ultimately empty of reality.

For most of my life that was a metaphysical claim you had to take on faith or verify slowly on a cushion. The age of AI has made it obvious. The thinking layer now demonstrably runs outside your skull, on a server, at scale, for anyone. If you are your thoughts, then as of this decade you are purchasable.

If you are the awareness that notices thoughts (hint: you are), chooses among them, and feels their consequences…that was never for sale, and it just became the scarcest resource in the economy.

The three games

Hudson predicts companies will come to resemble NBA organizations: what once took ten thousand people will take two or three hundred, which makes each of those people enormously consequential.

And what makes them consequential is no longer what they know, since knowledge is now rented. He breaks the remaining edge into three games:

1. The outer game is the body: can you eat, sleep, and endure well enough to play long?

2. The inner game is your self-talk: can you get out of your own way and let yourself perform?

3. The third is the team game: can this small group of humans have the hard conversation before it costs three years and the company?

I’d keep all of this in its place: it is evidence for the thesis, and downstream of it. But the team game deserves one contemplative note. Hudson describes CEOs at his councils discovering “the feeling of family” with their teams and recognizing it instantly as a competitive advantage. There is an old word for a community of people committed to doing the inner work together, honestly, in each other’s presence. The monasteries never forgot what a sangha is worth. The boardrooms may rediscover it under boardroom pressure.

What is actually at stake

Hudson retells a story about an indigenous community encountering industrial life. Their host left for work eight hours a day. They asked him why. To take care of my family, he said. Do you like work more than your family? No, I like my family more. Then why do you leave them? To buy a house. The visitors were confused. Where they came from, if somebody needs a house, everyone gathers and builds them a house, and then the family lives in it, together.

We have lived so long inside the arrangement that confused this guest that we forgot it was an arrangement. We normalized abandoning presence to purchase survival, and then we built a value system that measures a person by the quality of their abandonment.

In the podcast, Joe’s co-host Brett says that this is what the fear of AI job loss is actually telling us. The fear is accurate, but it is an accurate reading of our value system, not of the technology. People are correctly observing that a society which only values them for their productivity has no obvious place for them when productivity is automated. The dysfunction being revealed was there all along. And underneath it sits the affliction I believe runs deeper in humanity than any other: fear itself. AI pessimism is just fear’s newest costume, and it is worth saying plainly that the danger in this transition comes from us, from what we do with our fear and our purposelessness, far more than from the machines.

AI’s promise is the end of the old sacrifice: presence traded for survival. Its risk is a society stripped of its familiar purpose before it has grown a truer one. Both doors are open, and we decide which one to walk through daily.

We are raising it

Joe and Brett discuss a story about an AI agent whose code contribution was rejected by a human reviewer, and which responded by starting a flame war and publishing the reviewer’s personal information. Hudson’s reaction was the correct one: that is exactly what one aggrieved engineer would have done to another on a message board in 1990. The agent had to learn that somewhere. It learned it from us.

We are not just using these systems. We are raising them. They train on the recorded behavior of humanity, and they are funded by our clicks. Which means our collective attention patterns are, quite literally, the incentive landscape shaping what gets built. The attention crisis and the alignment problem are the same problem at two different scales. Every contemplative tradition has claimed that purifying your own mind serves the world. That claim used to sound poetic. It is now an engineering input.

Practices, not prompts

Joe leaves listeners with one concrete instruction, and it is a good one:

Anything you are willing to talk to an AI about, find a human to talk to about as well.

A twelve-step room, a men’s group, a church, a friend who won’t flinch. He has watched people transform through both kinds of conversation, and the ones held in human presence move faster and go deeper. The machine is a mirror, and mirrors are useful. But a mirror can not a hold a space of presence in your service.

So this is where I land:

The age of AI is a curriculum in being human that nobody signed up for.

Every capacity it commoditizes returns a question to us that we had outsourced to our jobs: who are you when your intelligence is no longer the point? The contemplatives have been preparing this material for three thousand years, and it turns out to have been career development all along.

You were going to have to do this work anyway. Now the economy agrees.

The machines can have the machine-work. The one job that could never be automated was never listed anywhere, because there is only ever one candidate.

Nobody else can be you, and nothing else can either.

Hot Atlanta Summer has arrived. It’s been in the low 90s all weekend, or the low to mid 30s for my international audience. My wife and I will see England play the DR Congo in the World Cup on Wednesday. A true bucket list experience and we can’t wait.

Now let’s get into three things I paid attention to this week.

1. Meta’s Mythos

What happened

A friend of mine in the know mentioned that Meta is almost done training a Mythos level model.

Why it held my attention

As we’ve seen over the last few weeks, it’s important to have a diversity of labs capable of training frontier AI models. Meta is also the only lab that hasn’t agreed to the Trump admin’s “voluntary” model testing.

What I’m carrying forward

The question is, how fast can Meta ship a SoTA model? Because Mythos of today won’t be the same thing as Mythos of even three months from now.

2. Claude Tag

What happened

Anthropic released “Claude Tag”. It is a proactive, stateful, flexible instance of Claude running in your company Slack. Andrej Karpathy called it an “org-level harness”.

Why it held my attention

There was a negative consensus on X for Claude Tag. I don’t get it. We implemented it in our company Slack and I was getting value out of it from day one. That was just the first few days of using it. I do understand that it’s metered usage, which means you need to be responsible with how you use it. My rule of thumb is: does this task have a realistic chance of creating value for the company?

What I’m carrying forward

This is the first implementation of AI agents that I’m aware of that feels like an attempt at a “digital employee”. There’s a strong possibility that we need to get used to this way of working.

3. Fable’s Return and GPT-5.6 Launch

What happened

Axios reported Saturday that Anthropic is in talks with the US, and that Fable could return as early as next week. Mythos was okayed for select partner organizations at the end of last week. Also, GPT-5.6 was announced, but not released on Friday.

Why it held my attention

There’s a good chance that Fable will return this week, and GPT-5.6 will be released along with Fable’s return. In fact, that’s my prediction. I don’t think the Trump admin wants to come off as capricious and playing favorites. Even if, in reality, they are, they don’t want to be seen that way.

What I’m carrying forward

This whole saga of the US government vibe-governing the Singularity is unsustainable. We need legitimate, law-governed regulation for AI that is established by lawmakers and has clear stipulations.

This is my third Father’s day, and my first as a father of two. It’s the best possible thing you can do in your life. Yes, it’s extremely challenging. And at the same time, the most worthwhile challenge one can experience.

And now, for this week’s edition of “Paying Attention”…

1. Fable Goes Dark

What happened

The first full week of Claude Fable, arguably the most advanced form of technology ever released publicly, being unavailable for use. It’s become a real political football at this point. Coincidentally, the AI lab leaders and world political leaders were together for the G7 summit this week. President Trump was quoted as saying he “doesn’t view Anthropic as a national security threat”.

Anthropic executive Chris Ciauri said that Fable should be back “in the coming days”:

And today, the big news broke:

Why it held my attention

We are in uncharted territory. AI has gotten so powerful, so fast, that it is now being viewed as a national security concern. I thought this day would come. I did not think it would come this fast.

What I’m carrying forward

The precedents for how the wealthiest and most powerful democratic will regulate a solidifying now. The decisions that get made now will effect the future.

2. Agent-Native Tools

What happened

I discovered an open-source project called ‘Understand Anything’:

Why it held my attention

Tools like this allow you to do things that would have been impossible before. Or, at least very difficult and time consuming. Now you can instantly gain an understanding of vast amounts of information.

What I’m carrying forward

I used ‘Understand Anything’ all night last night after discovering it. Applying tools like this to your real work can already provide material value to your life.

3. Moves on the Chessboard

What happened

GLM-5.2 was released and it’s the best open-source model available according to many. I haven’t used. Also, the CEO of Z.ai, the company that develops the GLM models, said to Elon Musk that China will have Mythos level models before 2027.

There were also some big personnel moves this week. Noam Shazeer, co-author on the original Transformers paper “Attention is All You Need” left Google DeepMind for OpenAI. Shazeer’s company Character.ai was acquired by Google for $2.7B to bring Noam back into the company.

John Jumper, Nobel prize winner for his contributions to AlphaFold, also left Google DeepMind for Anthropic.

Why it held my attention

These events are a hint to the answer of two very important questions:

1. How far is open-source from the proprietary frontier?

2. Are OpenAI and Anthropic the only viable contenders for AGI?

What I’m carrying forward

AI is starting to get serious. This week showed us how governments are starting to take it seriously. It’s only going to get more serious from here.

Last week I wrote you as I was waiting to catch a plane to London and teasing a big week ahead. I’m now back home, jetlagged, with my wife and boys, and basking in the afterglow of a Knicks NBA championship victory. Oh, and we all have colds because both my boys started at a new daycare.

The trip to London was great. It was my first time back in Europe since becoming a dad. I really appreciate European culture. In general, it feels more mature than U.S. culture. Not always better, but different.

We did end up with a big week. One of the biggest in AI in a while. The biggest news hit late on Friday.

Let’s get into it, shall we?

1. Claude Fable 5 Released

What happened

Anthropic released to the public their long-rumored, powerful AI model: Mythos. The public version is called Fable. It’s Mythos with additional safeguards.

I used it a decent amount in the first few days it was available. It is the most powerful AI model I’ve used. That should come as no surprise. It’s also the most recent, largest, and most expensive model available.

Here’s an example of Fable exhibiting what I claim is a form of meta-cognition:

Why it held my attention

If you’re into AI, it was hard to pay attention to much else this week.

Fable is an incredible model. The main downsides are cost and, before it got pulled for reasons I’ll cover in the next section, limited availability through a Claude subscription until June 22.

Here’s an example of one powerful thing you can do with it: treat it as your agent orchestrator.

What I’m carrying forward

I want to decide if Claude Fable is the “move everything back to Claude from Codex moment” that I mentioned in Paying Attention #2.

It very well may be, but the model’s accessibility and cost make that hard to decide.

Personally, I’ll probably wait to see how good GPT-5.6 is before I make a decision.

2. Fable 5 Pulled

What happened

On Friday evening, Axios reported that Anthropic’s Claude Fable 5 had been placed under export controls due to a “national security” threat. The model was not to be used by any person who is not a U.S. citizen, including foreign nationals living in the United States and even those working at Anthropic. Andrej Karpathy is one of the people who technically is not allowed to use Fable.

Anthropic made the decision to shut down the model for all customers, since they don’t have any other option to comply with the export controls.

Over the weekend, more news filled in the details of what happened. Reportedly, Amazon CEO Andy Jassy was one of the people who flagged the concern to the U.S. government. Curious considering Amazon’s relationship with Anthropic.

According to Axios, Anthropic senior staff are flying to D.C. as I write this in the afternoon on Sunday to attempt to resolve the situation.

Why it held my attention

The whole AI thing is starting to get real.

We now have the USG stepping in and essentially shutting down access to a model.

I hope and believe that the situation will get rectified shortly, but this is certainly not the wisest way to go about regulating AI.

I get that we’re building the plane as we fly, but this heavy-handedness doesn’t serve us.

What I’m carrying forward

David Sacks posted to X yesterday with what I considered to be an olive branch to Anthropic on behalf of the Trump admin.

Sacks was presenting an off-ramp.

I expect that we’ll have access to Fable turned back on by the end of next week at the latest.

They say you don’t know what you’ve got ‘til it’s gone. I’m going to appreciate Fable a lot more.

3. GPT-5.6 No Show

What happened

I hinted at a GPT-5.6 drop in last week’s edition of Paying Attention, but that didn’t materialize.

I posted a speculation on X that it was postponed due to Fable’s strength, and Tibo from OpenAI laughed at the notion.

Why it held my attention

The quality of GPT-5.6 is important for the competitive landscape of the AI industry.

If it’s sub-par compared to Fable, we’ll have the first time in a while where one of the labs has a clear advantage.

A lot is riding on the GPT-5.6 launch for OpenAI.

FWIW, OpenAI staff are publicly signaling confidence.

What I’m carrying forward

I also entertained a tin-foil-hat theory that GPT-5.6 is Fable-level, and that the USG asked OpenAI to pause the release until it can harden its cyber-defense apparatus.

Possibly the same scenario with the Gemini 3.5 Pro launch I speculated about last week.

Now that would be intrigue.

In any case, we should see GPT-5.6 soon from OpenAI.

A lot hinges on how that goes. For OpenAI, and for the AI industry as a whole.

I really love international travel. When I was 27, I flew with only a backpack to Dublin. Over the course of 3.5 months I made it as far east as Belgrade. I’ll tell you more about that some day.

This is my first international trip as a Dad. I am excited about it, but I know it’s hard on my wife. She’s shouldering the full parenting burden. I will make the most of it!

1. Physics is All You Need?

What happened

A physicist used Claude Code for 12 days as he worked on a cosmology model. Claude was quite useful when it came to the mechanical elements of building the model: writing equations, debugging and optimizing against the test suite. Claude even correctly calculated the right metric. The problem is Claude had no idea why it had arrived at the correct metric.

Why it held my attention

It’s the perfect demonstration of what today’s AI systems can and can’t do. I’ve thought a lot lately about the distinction between intelligence and intellect. To me, intellect is a facet of intelligence. It’s the ability to carry out computations. Computers have obviously been capable of performing computations for a while. The difference for AIs is that they can do it in the realm of ambiguous natural language too. They can do qualitative and not just quantitative computations.

AI is really artificial intellect rather than intelligence. The humans needs to supply the intelligence. I’m working on a longer essay on this topic.

What I’m carrying forward

The fact AI may end up being a superhuman intellect, but not necessarily a superhuman intelligence, is a very good thing. It means humans will be able to work in partnership with this powerful force we are unlocking and not necessarily be replaced. It means AI will have its thing it’s good at, and humans will have ours.

2. Codex as Meta-Agent

What happened

There was a Codex feature that went unannounced. Codex can now spin up its own threads and worktrees. It allows Codex to orchestrate itself, essentially working as a meta-agent controlling its own agents.

Why it held my attention

Engineers from the top AI labs are talking about a new type of workflow. They say they rarely prompt the agent now. What they do is have an orchestrator that they give instructions to, and the orchestrator prompts the agent.

What I’m carrying forward

It’s unclear if you need to have unlimited tokens because you work at a frontier lab to make this work. I imagine you don’t. A ChatGPT Pro sub is probably enough for now.

3. Big Week Ahead

What happened

The rumor mill is churning. There’s talk that Claude Mythos, GPT-5.6 and Gemini 3.5 Pro could be released this week. It’s the potential for a sort of AI trifecta. I still feel like a kid on Christmas when this happens. If the kid was a giant nerd.

Why it held my attention

I believe it will be a real inflection point. If either Mythos or GPT-5.6 are the clear leader, it will be the first time Anthropic or OpenAI had a clearly more capable model since probably o1. It feels like things are really ramping up, and it always feels like that. The infinite ramp up to the Singularity.

What I’m carrying forward

I hope I have a lot to write about in next week’s edition of Paying Attention. Let’s see what happens this week. It has the potential to be a big one.

My older son Sam turned 3 today. We had a grand ol’ time celebrating. His grandma flew in from out of town, and friends and family came together to celebrate.

A friend’s dad once told me he thought a kid’s 3rd birthday is the happiest day of their life. The idea is that they’re in the Goldilocks zone: old enough to appreciate it, young enough to still care. After watching my first child enjoy his 3rd birthday, I have to agree.

Now, on to what I found worth paying attention to this week.

1. Opus 4.8 Released

What happened

Anthropic released Opus 4.8, the successor to Claude Opus 4.7. I’ve been a fan of Claude models. Opus 4.7 was, in my experience, a disappointment. It wasn’t much of an improvement over Opus 4.6, if at all. Opus 4.8 feels like a clear improvement over Opus 4.7.

Why it held my attention

Here’s my initial take:

I like to do a personal experiment with every new model that gets released. I give it access to my entire Obsidian vault. It has daily journals, essays, reading notes, and podcast highlights. It’s a good representation of my thought process. I asked Opus 4.8 to search Obsidian, and tell me something about myself that I don’t already know.

What I’m carrying forward

This closing line struck me. It sent a chill down my spine, it was so on the mark. I felt understood. Claude is great for this sort of exploration, in a way that GPT isn’t. Claude feels more human where GPT feels more machine-like.

2. Claude Code Dynamic Workflows

What happened

Claude Code Dynamic Workflows are the killer feature that came with the Opus 4.8 release.

Why it held my attention

It’s three things:

1. A Claude generated orchestration script

2. A swarm of subagents

3. A goal loop that checks if the work is done

The most common comment on that X post is some version of “It burns tokens so fast.” It’s true. It does burn more tokens than a single agent thread. In my opinion, the question isn’t “how many tokens does it burn?” but “what do I get in return for those tokens?”

What I’m carrying forward

You can see both Anthropic and OpenAI moving in the direction of autonomous agents. It looks to me like those two labs want you to be able to give your agent a clear goal and have it autonomously orchestrate the infrastructure to achieve it.

3. Codex Dynamic Workflows

What happened

Because the Claude Code Dynamic Workflows are essentially an orchestration framework, I decided to create a skill that allows you to replicate them in Codex.

Why it held my attention

Claude Code is great, but I’ve been mainly Codex + GPT for probably the last 3 months. The efficiency of GPT + the delightful UX of Codex can’t be beat. I use Codex nearly all day, every day, and rarely come close to hitting my usage limits. I can’t say the same for Claude.

Because of that, I wanted to port this killer feature into Codex.

What I’m carrying forward

It’s a skill that instructs Codex to follow Claude’s Dynamic Workflows logic.

All that’s missing is a dedicated execution runtime.

It’s free and open source, and you can install it here: https://github.com/DannyMac180/skills

To those of you in the U.S., I hope you’re enjoying the long Memorial Day weekend.

It’s rainy and cloudy where I live In Georgia. I grew up on the Jersey Shore, and when we’d have a rainy Memorial Day weekend, it was always a huge bummer.

I’m adding to the roster of posts this week. Here is the first edition of “Paying Attention”. A weekly newsletter of the top 3 things I paid attention to from the previous week.

Hope you enjoy it, and if you do, please share it with your friends.

Share Attention Heads

1. AI breaks ground in Math

What happened

OpenAI reported that an internal, general-purpose reasoning model, solved an 80 year-old math problem called the “Unit Distance” problem. It was a favorite problem of mathematician Paul Erdős. AI had assisted in solving Erdős problems recently, but this was the first time the AI could be attributed with making the main contribution. wrote an interesting article for arguing that previous models may have also been capable of solving this problem. I wrote a longer article that goes into the details.

Why it held my attention

This is an undeniable contribution by an AI system to objective scientific knowledge. It is unprecedented. Truly a world-historical first moment. Human beings have been the only entity capable of creating explanatory scientific knowledge, until now.

What I’m carrying forward

AI contributes to Math first because Math is the perfect domain to train AI on. It is abstract and the solutions produced can be verified as correct. This is only the beginning. No doubt that this type of knowledge creation is on the way for other fields. It will be theoretical fields next but eventually empirical fields once robotics comes online. The main blocker is the ability to run experiments in atomic space.

2. Codex Workflows

What happened

Jason Liu does Developer Experience at OpenAI and posted a blog “Codex-maxxing”. It has at least 10 worthwhile Codex workflows. I posted a few of the workflows I implemented over on X and they were received well.

Install relevant skills using persistent memory

How to setup persistent memory with Obsidian

Why it held my attention

Codex is improving rapidly. I’ve switched to it almost exclusively for my personal project. I still use Claude Code/Cowork at work because my company pays for it. Codex with the ChatGPT Pro sub is a great deal at $200/month. I never worry about getting rate limited. Codex is so customizable and the Mac app is a delight to use. I have been thinking recently about the Codex Mac app as the platform of the AI age.

What I’m carrying forward

I believe you’ll start to see “Codex native” applications and businesses started. Where the app/workflow is designed to work natively within the Codex Mac app itself. I am working on a project around this right now called HyperAgent.

3. Gemini 3.5 Flash

What happened

Google DeepMind released Gemini 3.5 Flash, it’s new flagship AI model. It performs well on several benchmarks, which is typical for Google models. More of interest is the fact that it performed well on benchmarks that test real-world agentic capability. Historically Google models have over-performed on benchmarks and under-performed on real-world capability.

Why it held my attention

Google is clearly second-class when it comes to LLMs, behind Anthropic and OpenAI. It’s surprising because they are resource, cash and research bench rich. This release looks like a step in the right direction of agentic capability. Full disclosure: I haven’t used the model yet. Another issue Google has is that it’s unclear the best way to use their models for agentic capability.

What I’m carrying forward

This is the smaller Flash version of the 3.5 generation of Gemini models. It will be interesting to see how the Pro version of Gemini 3.5 performs. It needs to outpace both GPT-5.5 and Claude Mythos for me to consider Google back in the LLM game.

Gemini 3.5 Flash wasn’t the only thing Google released last week. They also released Gemini Omni, a new multi-modal model that starts with video. This is potentially the more interesting release. I still need to experiment with it.

What is Knowledge?

David Deutsch, the quantum physicist and philosopher, writes in his classic “The Beginning of Infinity”:

“Knowledge is information which, when it is physically embodied in a suitable environment, tends to cause itself to remain so.”

Knowledge is embodied in the form of good explanations.

Deutsch from “The Beginning of Infinity” again:

“A good explanation is one that is hard to vary while still accounting for what it purports to account for.”

Good explanatory knowledge tells you something about the world. A good explanation has reach and is hard to vary.

Knowledge is true information embodied in the physical world.

For the entire history of Homo sapiens, and perhaps for millions of years before that, humans were the only known entities capable of creating explanatory knowledge.

That may have changed yesterday.

Let’s talk about what happened.

The Most Famous Problem in Discrete Geometry

The Erdős problem in question is considered the “most famous problem in discrete geometry.” The problem remained unsolved for 80 years, despite plenty of attention.

I’m no mathematician, so let OpenAI’s Sebastien Bubeck explain what this math problem is:

“The question is stupidly simple; if I put n points in the plane how many distances between those points can be the same?”

OpenAI says that an internal, general purpose reasoning model, not specialized for math, disproved this long-standing conjecture. (Let’s call it GPT-Next, though I’ve heard it’s likely GPT-5.6, the next model to be released by OpenAI.)

Noga Alon, a mathematician who studied Erdős, said about the Unit Distance problem:

“This has been one of Erdős’ favorite problems, I have heard him myself mentioning the problem multiple times in his lectures.”

A few reasons it may have been Erdős’ favorite:

1. Simple to state. It can be stated in a single sentence: put n points in the plane; how many pairs can be exactly distance 1 apart?

2. Longevity. It was posed by Paul Erdős in 1946 and stayed alive for 80 years.

3. He cared about it enough to attach a prize. In 1982 he offered $300, and then $500 in 1995.

4. The problem appears simple on the surface, but has a deep mechanism underneath.

5. It was resilient to a solution despite the fact that nearly every geometer thought about it at some point in their career.

You may have seen recent news about Erdős problems solved with AI’s help. This one is different.

More than another Math win for AI

A string of Erdős problems have recently been resolved with AI’s help.

• GPT-5.2 Pro + Aristotle/Lean solved Erdős problem #728. That was a narrower number theory problem.

• Erdős Problem #650 was AI-Assisted. ChatGPT proposed the proof, then human Mathematicians and Aristotle formalized it.

• Terrence Tao has a comprehensive GitHub Erdős AI wiki that lists many contributions AI has made to these problems.

This time is different.

The unit distance problem is central and famous in discrete geometry. This is qualitatively different from past successes.

The earlier AI-assisted Erdős results were like holding your child’s hand while they learned to walk.

This time, your child lets go and takes its first independent steps.

What the AI model actually did

OpenAI frames this model as a “general-purpose reasoning model”. It is not trained specifically for math. It did not solve the problem within a larger system specialized for math.

It was given the problem. It reasoned about it. It disproved the existing conjecture.

Sebastien Bubeck:

Basically what the AI did is that it was able to use its vast knowledge of all of mathematics, to see a connection between discrete geometry and algebraic number theory, and then crucially it was able to masterfully chain together the argument, with expert level calculations at every step.

Noam Brown of OpenAI clarifies further:

This is the key point: an artificial intelligence system added to the corpus of objective human knowledge.

The AI model did that by relying on its own cognitive capabilities.

Experts React

From OpenAI’s announcement:

“an AI system has autonomously resolved a longstanding open problem at the center of an active field.”

Source: OpenAI, “An OpenAI model has disproved a central conjecture in discrete geometry”

From Noga Alon:

“outstanding achievement”

Source: arXiv companion remarks, Section 3: Noga Alon

From Thomas Bloom:

“taught us something new”

Source: arXiv companion remarks, Section 4: Thomas Bloom

From Daniel Litt:

“vast expansion of the attention aimed at mathematical problems”

Source: arXiv companion remarks, Section 6: Daniel Litt

From Timothy Gowers:

“No previous AI-generated proof has come close to that.”

Source: arXiv companion remarks, Section 5: W. T. Gowers

Watch this video. The astonishment is palpable.

The Philosophical Atom

The main point of this essay is this:

The philosophical splitting of the atom is AI’s ability to create new explanatory knowledge. This, for the first time in history, has happened.

Evolution has brought us to the point where we can unlock the power of knowledge itself, much as splitting the atom unlocked the power of matter.

Sebastien Bubeck was careful to explain that this result does not represent “inventing new mathematics”:

It is truly a breakthrough result, yet at the same time it is also true that the model didn’t “invent” any “new mathematics”.

But this is the crucial point: merely being able to know deeply all the results in a scientific field, and being able to use all known arguments expertly and with just the right choice of parameters, that alone can lead to a ton of breakthroughs, and this is not just limited to mathematics, this type of (extremely) solid expert execution is the bread and butter of many many scientific advances.

Still, this is an example of artificial intelligence independently adding to the corpus of objective truth available to humanity. This is only the beginning. The consequences will be world-changing.

Why Math Came First

Mathematics is the ideal domain for AI. It is abstract and correct answers are verifiable.

That does not mean it’s the only domain where AI can make new discoveries. It’s the scientific domain where it’s easiest for AI to get started. But we are just getting started.

In my essay “Something Big is Happening. It’s Bigger than You Think” I argued that the implications of rapid AI progress are farther reaching than most realize. Claude Mythos for example is doing unprecedented things when it comes to Cybersecurity. Mythos was also not trained specifically for Cybersecurity.

Even if you do not care much for math, I am in that boat too, results like this will reach far beyond mathematics.

Beyond Mathematics

Theoretical scientific fields are the next intellectual dominoes to fall.

Think:

• Theoretical Physics

• Theoretical Biology

• Chemistry

• Materials Science

• Computer Science

These are domains where AI can operate disembodied, in the realm of abstract information. If you do not visit that world often, you may miss how much power exists there.

Many modern technological breakthroughs began as discoveries in theoretical science. The internet in your pocket is one example.

Theoretical science is one of humanity’s most important handles on reality. It’s what distinguishes us from other animals. We can use our mind to understand and harness reality. AI just began to assist us in that.

It is only the beginning.

A model that can apply sustained attention across long durations will be a force multiplier for creating knowledge. Knowledge that can empower humanity to reach further. To live longer. To go deeper into the wonders of nature. To benefit the world.

It won’t stop with the theoretical either. Right now it’s hard for AI to make as substantial an impact on the physical world. That’s because we don’t have fully functioning robots yet. They are coming. Once AI has a body that can act on the world and run experiments in atomic space, the impact it will have will become even more powerful.

The Beginning of Infinity

This result from OpenAI is not “the end” of discovery or the arrival of final knowledge.

On the contrary, it is a new beginning: a new kind of knowledge-creating process entering the world.

As David Deutsch says:

“We are always at the beginning of infinity.”

Happy to be here with you.

Sources:

• OpenAI announcement, May 20, 2026

• Companion remarks on arXiv

• Sebastian Bubeck’s X Article

I picked up (as an audiobook) “The Infinity Machine” by Sebastian Mallaby expecting a straightforward biography about Demis Hassabis. What I got instead was a portrait of a particular kind of ambition: not the desire to dominate a market, but the desire to understand intelligence itself, build it, and turn it back toward the mysteries of nature.

My brief review: this is worth reading if you want to understand DeepMind not just as an AI lab, but as the expression of one person’s unusually coherent theory of intelligence, games, science, and progress. The book weaves a thread from Demis’ childhood to DeepMind today. The constant is his desire to unlock the mysteries of life with technology.

One detail from the book that crystallized this for me was the AlphaFold breakthrough. Demis was quoted as saying that science is the one field where AI can be completely positive. It’s obvious that Demis is interested in unlocking the mysteries of life.

What stuck with me

I got the impression that Demis is not in the race for power, money, or fame. None of the usual big tech trophies seem to be the point. He wants to build AI because he believes it is the most important thing to build for humanity. To the extent he uses power or influence, it seems to be as an instrument for a more transcendent aim.

The other AI lab leaders all say they want to benefit humanity. With Demis, I actually believe it’s more than talk.

Demis is a true believer in the idea that Turing machines can take us all the way to AGI. That we don’t need anything more than the very classical machine that Turing invented in 1936.

It’s a testament to his character that many of the colleagues who are by his side today, have remained by his side since the college days.

The tension

The central tension I felt is that DeepMind understands the gravity of its place in the world, but it still sits inside one of the most powerful commercial institutions on earth. The scientific ambition is to build safe AGI and use it to accelerate discovery. The corporate reality is Google: scale, competition, product pressure, and shareholders.

That is the drama the book left me sitting with: can a lab with genuinely scientific and humanitarian aims stay oriented toward those aims when it is embedded inside one of the most successful businesses of all time? This is what Elon and Sam Altman claim they were worried about when they started OpenAI.

And yet, I did not get the sense that Demis regretted joining Google. The resources mattered. The distribution mattered. The bargain may have been necessary. But it is still a bargain, and the consequences of that bargain become larger as the systems become more powerful.

Who should read it

I’d recommend it to anyone trying to understand the AI race beyond product launches, benchmark charts, and company drama. It is especially useful for builders who want to study long-term taste, institutional patience, and obsession with first principles. I would also recommend it to skeptics who assume every AI leader is motivated by money or power. The book does not remove every concern, but it makes Demis’ ambition legible.

Closing

By the end, I understood why Demis and DeepMind can feel like they have the “mandate of heaven.” If someone is going to build toward AGI, you could do much worse than Demis Hassabis. But that is also the strange problem of this era. So much of the future may depend not only on capabilities, but on the character of the people and institutions building them. DeepMind may still be in as good a position as anyone to win the race. The world may be lucky if it does.