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.

Think back to February 2026.

Matt Shumer’s viral essay “Something Big is Happening” articulated a feeling: AI had started to work. AI was really happening. And like the early days of Covid, we didn’t appreciate what that meant.

A few months later, we have more than a feeling.

We have exponential progress staring us in the face.

Three months ago, the evidence was anecdotal: software engineers took a holiday break, and a few famous ones noticed that AI coding agents worked. They told their audiences about it and the swell started to grow. Bright minds like began saying that Claude Code + Opus 4.5 was AGI.

Since then, the evidence has become institutional: AISI, METR, Anthropic-adjacent coalitions, cyber-security groups, huge AI-science financing.

If Matt’s essay was “we can see UFOs in the sky”, this essay is:

Agents Become Undeniable

There are two recent results that show agents have become largely autonomous.

1. Claude Mythos Solves AISI’s Full Cyber Attack Simulations

Claude Mythos preview is the first model to solve both of AI Security Institute’s Cyber attack simulations. “The Last Ones” is one of the cyber tests Mythos successfully completed in 6/10 attempts. It is estimated to take a human attacker 20 hours to complete. The other cyber attack test Mythos completed, named “Cooling Tower”, was previously unsolved. Mythos succeeded in 3/10 attempts.

In November 2025 the approximate doubling time for AISI cyber test task duration was 8 months. By February of 2026 it had accelerated to ~4.7 months. Both Mythos and GPT-5.5 are ahead of the 4.7 month trend.

If that trend holds, AI task duration, at least for cyber tasks, will be super-exponential. The tasks models can complete are getting longer, and the rate of improvement is itself speeding up.

There’s no reason to assume this jump in capability remains limited to cyber tasks. Anthropic has confirmed Mythos is a “new general-purpose language model” and say the cyber capabilities “emerged as a downstream consequence of general improvements in code, reasoning, and autonomy.”

Anthropic decided to make Mythos available to a select group of organizations for cyber-defense purposes first, because that’s most important for society. This level of capability should generalize to other domains too. There is no reason to believe it won’t.

2. Mythos Breaks METR

METR measures something simple: how long a task takes a skilled human, and whether an AI agent can complete tasks of that duration with a given success rate. Their “50% time horizon” is the point where an agent succeeds half the time on tasks that would take a human expert that long.

Claude Mythos Preview is now past the edge of that measurement. METR’s latest update added early Mythos results and warned that measurements above 16 hours are unreliable with the current task suite.

The reason this is interesting is not just “Mythos can do longer tasks.” It’s that our AI models are now outpacing our ability to benchmark them. We are starting to run out of clean, well-measured tasks long enough to test what these agents can do.

AI went from “can do tasks that would take human a couple minutes” to “tasks that would take a human a couple days” in a couple years. At least. METR says the measurements are unreliable.

Cyber is the sharpest real-world test case, but Anthropic says the capability came from general improvements in code, reasoning and autonomy. The METR result backs that up.

Agents can now perform real, economically valuable work. They are getting better fast.

Agents have become undeniable.

That’s great if the good guys have them. Not so great if the bad guys do.

The Security World Responds

This is exactly why Anthropic did not just ship Mythos to everyone.

Instead, they launched Project Glasswing: a coordinated effort with AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks and others to use Mythos for cyber defense first.

Think about that for just a minute.

One of the leading AI labs built a general-purpose model so capable at finding and exploiting software vulnerabilities that they decided the responsible first release was not a chatbot product but a security coalition. Some criticize the move as pure “marketing hype.” But it would have been a better marketing strategy to release Mythos widely and keep growing revenue.

Anthropic says Mythos has already found thousands of high-severity vulnerabilities, including bugs in every major operating system and every major web browser. Some were quite old. One was a 27-year-old vulnerability in OpenBSD. Another was a 16-year-old bug in FFmpeg that automated testing tools had hit millions of times without catching.

And again: Mythos was not trained specifically to be a cyber weapon. The cyber capabilities emerged from improvements in coding, reasoning and autonomy.

That’s the part that should send a chill down everyone’s spine.

If you have an AI agent that can autonomously find subtle bugs, chain vulnerabilities together, reverse engineer binaries, and produce working exploits, you have something that can change the balance between attackers and defenders.

In the short term, this is dangerous. Attackers only need to find one path in. Defenders have to protect everything. If models like Mythos become widely available before defenders are ready, the bad guys get leverage too.

But this is also why the upside is so large.

For the first time, the good guys may have a tool that can search through huge amounts of critical software and find the hidden cracks before attackers do. Anthropic is putting up $100 million in usage credits for Project Glasswing and additional participants, plus millions more for open-source security organizations.

That’s why it’s so important that powerful AI is pointed in the right direction.

It could become a true boon to society, but only if we use it that way.

Powerful AI could help harden the software that banks, hospitals, governments, browsers, phones, servers, and open-source projects depend on.

I expect this pattern to repeat. The same capability that makes AI risky also makes it useful. The same autonomy that can be abused can also be used to defend, discover, repair and build.

It’s not so much a double-edged sword as it is a double-edged cognitive nuclear weapon.

The Next Frontier Is Doing AI Research

And cyber is not the end of the story.

It’s just the first place where the rubber hit the road for civilizational infrastructure.

The next frontier is AI systems doing AI research itself. Then, potentially, autonomous scientific research writ large.

, co-founder of Anthropic and author of Import AI, recently said he now thinks there is a 60%+ chance that we get no-human-involved AI R&D by the end of 2028. In plain English: an AI system powerful enough that it could autonomously build its own successor.

It sounds sci-fi, until you look at what AI research is and what models are already good at.

Think about it. What is AI research? Some of it is inspiration. New ideas. Taste. Weird intuition. The thing that causes someone to see a problem from an original perspective.

But much of it is something else.

A lot of AI research is writing code, running experiments, cleaning data, reading papers, reproducing results, debugging training runs, optimizing kernels, comparing model outputs, searching through failed attempts, and trying again.

That is exactly the kind of work agents are getting good at.

Andrej Karpathy’s autoresearch project proved that today’s AI models can already do autonomous research to improve a model’s own hyperparameters. What happens when you can not only improve a model’s hyperparameters, but when you can autonomously improve the algorithms, the architecture, the data mix, all of the ingredients that go into making a model stronger than it currently is?

This is where the latest METR results come into the picture. Right now models can do coding tasks that might take a human SWE a couple days. Soon, models may be capable of research tasks that would take a human researcher even longer.

AI is not a replacement for the best human researchers yet. Not some magic oracle that invents the Transformers successor on command.

But a synthetic colleague that can take more and more of the toil.

And once AI can automate large chunks of AI research, progress stops being limited only by the number of human researchers who can think, code, test and iterate. Progress starts being limited by compute, data, evaluation, safety, and how well we can direct these systems.

Then the human researchers can do what they do best: sit around being inspired. And the autonomous AI researchers can do what machines do best: carry out the toil without ever getting bored or tired.

Because if AI helps build better AI, and better AI helps build even better AI, the feedback loop becomes recursive. The thing improving is also becoming part of the improvement process.

That is why “Something big is happening” turns into “It’s Bigger than you Think”. Humans have a very hard time conceptualizing exponential progress. It’s not how we evolved.

AI progress may literally accelerate past the point that we can imagine due to our inherent cognitive limitations.

And if AI can automate its own research, it is plausible that it will eventually automate the whole of scientific research. It may become capable of autonomously creating explanatory scientific knowledge. And it could do this 24/7/365. This is harder until AI can run experiments in atomic space, but it is plausible.

The quantum physicist and scientific philosopher David Deutsch defines true wealth as “the set of physical transformations you have the knowledge to effect on the world.”

Powerful AI may come to the point where it can autonomously create that type of wealth, nonstop, for as long as we keep the datacenters running.

The Upside Is Not Just Productivity

This is why the upside is so much bigger than “AI makes workers more productive.”

That framing is true, but it is too small. It makes AI sound like better email, faster spreadsheets, cheaper customer support and faster software engineers. Useful. Valuable. Economically important.

But not world-historical.

The real upside is not that AI lets us do the same things faster. It is that AI may let us do things we cannot do at all today.

The upside is that AI unlocks new vistas for the human race.

It may let us search spaces that are too large for human teams. Read literatures too vast for any one researcher. Run simulations, propose experiments, debug failures, generate hypotheses, write code, compare results, and keep going after every human in the lab has gone home.

That is larger than productivity.

That is acceleration of discovery.

One concrete example is drug discovery. Isomorphic Labs recently raised $2.1 billion to build AI systems for designing new medicines. That does not mean AI has already cured disease. It does not mean biology is solved. Biology is messy, expensive, physical, and brutally humbling.

But it does mean some of the smartest capital in the world is now betting that AI can change the economics of discovering medicines.

Isomorphic’s stated mission is to “solve all disease.”

That’d be a pretty big deal.

Because the optimistic case for AI is not “your boss gets a cheaper intern.” The optimistic case is that humanity gets more shots on goal against cancer, Alzheimer’s, rare diseases, antibiotic resistance, aging, climate, materials science, energy, and every other hard problem where progress is bottlenecked by the rate at which we can create and test knowledge.

Many of us are still thinking about AI as a labor technology. A productivity tool.

We ask: whose job does it replace? Whose salary does it compress? Which company gets more efficient?

Those are important questions. We should take them seriously. There will be disruption. There will be losers. There will be people caught in the transition who did not ask to be part of a civilizational technology experiment. That may be most of us. We should not hand-wave that away.

But labor is not the deepest category here.

Knowledge is.

If AI becomes a system for creating new knowledge, then the relevant comparison is not outsourcing or automation. It is the printing press. The scientific method. The computer. The internet. Maybe something larger than all of them, because this time the tool is not just storing or transmitting human thought. It is beginning to participate in the process of generating it.

It would be a force capable of creating new knowledge. Humans have been the only known example of that force for all of history.

That is what makes this moment:

Bigger than you Think.

The same systems that can find software vulnerabilities may help secure hospitals and banks. The same systems that can automate parts of AI research may help automate parts of biology, chemistry, materials science, engineering and medicine. The same systems that can displace some forms of work may also give millions of people access to tutors, coaches, collaborators, translators, programmers, analysts and teachers that would previously have been unavailable to them at any price.

Imagine a world where a kid in a small town can learn math, coding, physics, writing, biology, design and entrepreneurship with an infinitely patient tutor that adapts to them personally.

Imagine a nurse retraining into software. A factory worker learning robotics. A founder building a product that would have required a twenty-person engineering team five years ago. A patient with a rare disease getting an AI-assisted explanation of the latest research that their local doctor would never have had time to read.

None of this is guaranteed.

That part is important.

Powerful AI does not automatically distribute itself fairly. It does not automatically cure disease. It does not automatically make people wiser, institutions better, or society more humane. It can concentrate power. It can accelerate conflict. It can widen inequality. It can flood the world with synthetic persuasion and automated offense.

But it is also not realistic to look at a technology that may automate large parts of science and only see danger.

The dangers are real. The upside is also real.

And the upside is not marginal.

If we can point these systems in the right direction, they may become a way for civilization to solve problems faster than the problems compound. They may help us harden the digital world, discover new medicines, invent new materials, teach new skills, build new institutions, and create forms of wealth that are not just financial, but physical, scientific, educational and human.

They may help us defer death. They may help us expand life.

That is the version of “something big is happening” that I think people still underestimate.

Greater than smarter models.

Greater than better coding agents.

Greater than productivity software with a chat box.

A new engine for creating knowledge.

That is why we must get this right.

The Dangers Are Real. They Are Not The Whole Story.

This is where the public conversation usually breaks down. For that reason, this is where we need to come together.

One side sees the upside and treats the risks like annoying footnotes. The other side sees the risks and treats the upside like a marketing hallucination.

Both are wrong.

The closer you get to approaching truth, the more you perceive paradox.

F. Scott Fitzgerald said:

The risks are not imaginary. We just walked through one of them. A model that can autonomously complete cyber attack simulations is not a toy. A model that can find hidden vulnerabilities, chain them together, and produce working exploits can make the world less secure if it is released carelessly or used by the wrong people.

That is not doomposting.

That is just what the capability is.

The same is true in other domains. If AI can accelerate research, it can accelerate good research and bad research. If AI can persuade, it can teach and manipulate. If AI can code, it can build useful software and malware. If AI can discover new biology, it can help cure disease and lower the barrier to biological misuse.

This is the basic shape of powerful technology.

Fire cooks food and burns cities. Electricity lights hospitals and powers electric chairs. Nuclear physics gave us both abundant energy and weapons that can end civilization. The internet connected the world and also gave us misinformation at planetary scale.

AI is not different because it has no downside.

AI is different because the downside and upside both scale with intelligence.

That is why I do not think the right posture is acceleration at all costs. It is also why I do not think the right posture is fear at all costs.

The right posture is sincere courage.

We need strong evaluations. We need responsible release decisions. We need security coalitions like Project Glasswing. We need better monitoring for dangerous capabilities. We need model labs to cooperate where they should cooperate, compete where they should compete, and be honest about what their systems can do.

We need governments that understand the technology well enough to regulate it without strangling the parts of it that could help humanity most.

We need builders who do not treat every safety concern as cope from people who do not understand progress.

And we need safety people who do not treat every expression of hope as naivete.

We need balance.

Because the scary truth is that we probably need the powerful systems to help manage the risks created by the very same powerful systems.

That is already what Project Glasswing points toward. The answer to AI-enabled cyber risk is not simply “never build capable models.” The answer is to make sure defenders get the best tools first, to harden the systems society depends on, and to build institutions that can react at the speed this technology is moving.

The same pattern may apply everywhere.

We may need AI to help audit AI. AI to help secure AI. AI to help explain AI. AI to help discover the medicines, materials, energy systems and defensive tools that make the next world more survivable than this one.

There’s no way around it. We can’t turn back the clock.

That does not mean we should trust the machines blindly.

It means we should stop pretending the only choices are worship and panic.

The question is not whether AI is good or bad in the abstract.

The question is whether we can build, deploy and govern it in a way that lets the upside outrun the danger.

That is the real stakes of this moment.

Not whether AI is impressive. It is.

Not whether AI is dangerous. It is.

Whether we can become earnest enough, fast enough, to use something this powerful for human flourishing instead of letting it become another force that overwhelms us.

Can Humanity Grow Up Fast Enough?

That, to me, is the pertinent question.

AI progress is real.

The models will get much more capable.

The stakes are civilizational.

The question is whether our institutions, norms, laws, labs, companies, and personal habits can mature quickly enough to meet what we are building.

Because the technology won’t wait for us to become wise.

The models are improving. The agents are getting longer-horizon. The benchmarks are straining. The security world is reorganizing. AI research itself is starting to look automatable. AI-for-science is attracting billions of dollars.

This is already happening.

The choice is not whether to live in the age of AI.

The choice is what kind of age of AI we are going to build.

We can build one where power concentrates, institutions lag, attackers get leverage, workers are treated as disposable, and everyone slowly loses trust in what they see, read and hear.

Or we can build one where powerful AI is pointed at the bottlenecks that actually make human life worse: disease, insecurity, ignorance, fragility, wasted talent, broken institutions, brittle software, slow science, and the fact that most people never get access to the kind of personalized education, mentorship and tools that would let them become what they could become.

That second path is not automatic.

No path is.

It has to be chosen.

It has to be built.

It has to be defended.

Humanity has to walk the walk. And we can.

And it has to be done by people who can hold two thoughts in their head at once: this technology is dangerous, and this technology is one of the greatest opportunities humanity has ever had.

The public conversation keeps trying to collapse AI into a single moral category: savior or monster, bubble or apocalypse, productivity tool or existential threat.

But the reality is deeper than that.

AI is becoming a new layer of civilization. A layer that can create civilization level software, search for vulnerabilities, teach people, assist researchers, compress expertise, automate experiments, discover patterns, generate plans, and maybe eventually create new scientific knowledge.

That is why this feels so difficult to reason about. We are not just adding a new app to the economy. We are adding something closer to a new kind of cognitive infrastructure.

It is unprecedented.

A civilization with more intelligence available to it is not automatically a better civilization.

But it is a civilization with greater possibility.

More ability to see hidden cracks before they break. More ability to test ideas before they become dogma. More ability to discover treatments before people die waiting. More ability to teach people before their potential closes. More ability to build tools that used to require institutions. More ability to turn knowledge into real transformations in the world.

That is what I think Matt was pointing at in February.

The feeling that something had changed.

The sense that the old categories were no longer enough.

The awareness that we were watching the beginning of something whose full shape had not arrived yet.

A few months later, the shape is clearer.

Agents have become undeniable.

The security world is responding.

AI research itself may be next.

AI-for-science is becoming a serious institutional bet.

And the upside, if we aim this correctly, is not just faster work.

It is faster discovery.

Faster learning.

Faster healing.

Faster building.

Maybe, eventually, a faster path toward a world with less needless suffering and more human possibility.

A world where we all smile a lot more.

Something big is happening.

But it is bigger than another model release.

Bigger than another benchmark.

Bigger than another viral AI moment.

The thing that is happening is that intelligence is becoming civilizational infrastructure.

And if we do this right, that infrastructure may help humanity heal, learn, build, discover and adapt at a scale we still do not know how to imagine.

In 2017, a group of researchers published a machine translation paper with one of the great titles in the history of AI: Attention Is All You Need.

At the time, the title was technical. The paper introduced the Transformer, an architecture that dispensed with recurrence and convolution and made attention the core primitive for processing sequences. In retrospect, the title now feels mythical. It names the mechanism that helped unlock the modern AI era.

But the title also raises a deeper question.

If attention was the breakthrough needed to allow machines to become intelligent, what exactly is attention?

And if attention matters this much for machines, what does that imply about us human beings?

This is the question I want to use as the first real essay for Attention Heads, because it sits right at the center of what this publication is about. Artificial attention. Human attention. Consciousness. World models. Meditation. Technology. My life experience has shown me that attention is the most powerful force in the world..

My starting thesis is simple, and maybe too grand, but I think it is directionally true:

Attention is the most powerful force in the universe because it transforms potential into reality.

Not a physical force like gravity. Not a mystical substance. A causal force that acts on the informational substrate of reality. What receives attention gets selected, modeled, strengthened, practiced, funded, remembered, and eventually built.

Attention is where possibility becomes actuality..

What Machine Attention Does

The original Transformer paper is not about meditation, consciousness, or philosophy. It is about sequence transduction, especially machine translation. But the core idea is surprisingly approachable.

A model is given a sequence of tokens. To understand or produce the next representation, each token needs to know which other tokens matter. Some words depend on nearby words. Some depend on words much earlier in the sentence. Some phrases only make sense in relation to the whole.

Machine attention gives the model a way to ask:

Given where I am, what else in this context is relevant?

That is the breakthrough of Transformers and the attention mechanism at the fundamental level.

Instead of processing a sentence only step by step, self-attention lets tokens look across the sequence and assign weight to one another. Multi-head attention lets the model do this in several ways at once. One head might track syntax. Another might track reference. Another might track a phrase-level dependency. The model is not literally thinking in these terms, but the architecture creates room for multiple patterns of relevance to be discovered in parallel.

This is one reason the Transformer was so powerful. It made relation, not mere sequence, the center of computation.

A sentence is not understood word by word. It is understood through relations: this pronoun points back to that noun; this adjective modifies that object; this phrase changes the meaning of the whole. Attention lets a model dynamically weight those relations.

The results was a general purpose architecture for learning from context. Learning from the meaning between the words. This has changed the world.

Machine Attention Is Not Human Attention

It is important not to overstate the analogy.

Transformer attention is not awareness. It is not mindfulness. It is not subjective experience. It is a mathematical operation inside a machine learning model. It routes information by computing patterns of relevance among tokens.

A model attending to a word is not the same as a human attending to a breath, a face, a memory, or a feeling. As far as we know, Claude doesn’t actually care about the sentences it’s processing.

But the analogy is still revelatory.

The architecture that unlocked modern AI did not simply get bigger. It got better at selective relevance. It learned which parts of the context should influence what happens next.

That phrase, selective relevance, is the bridge.

Because human beings do not experience reality as a raw, complete, neutral feed either. We experience a world that has already been selected, filtered, weighted, interpreted, and stabilized by attention.

Most of the time, we do not directly experience the world. We experience generated perceptions overlaid onto reality. We live inside our mentally constructed world model.

And that world model is mediated by attention.

Human Attention Generates A World

Look around the room you are in right now.

Before reading this sentence, you were probably not conscious of the feeling of your feet, the exact shape of the shadows, the sound floor of the room, the temperature on your skin, the peripheral objects outside the center of your visual field, or the emotional tone in your body.

Those things were there. But they were not equally real to you.

When I suggest giving your attention to them, they become more real.

Attention is what makes some part of the field come forward.

This is one of the strange truths of ordinary experience: salience feels like reality. What captures attention feels more real, more urgent, more meaningful, more central. What falls outside attention fades into the background, even if it is still causally present.

This is why attention is not merely a spotlight. It is also a world-generator.

What we attend to repeatedly becomes familiar. What becomes familiar becomes salient. What becomes salient begins to guide action. Over time, action shapes the world.

This is true at the scale of a person and at the scale of a civilization.

Do you know one reason Apple is one of the most valuable companies in the world? When Apple does something, it commands attention.

A product event. A design choice. A rumor. A keynote. A new category. The attention is not incidental to the value. It is part of the value. Collective attention changes what developers build for, what journalists explain, what competitors copy, what consumers desire, and what markets price in.

Attention is economic. Attention is cultural. Attention is spiritual. Attention is computational.

Attention Alone Is Not All You Need

The paper title is perfect, but as a human philosophy it needs a correction.

Attention alone is not all you need.

You need to realize what is relevant.

This is where John Vervaeke’s idea of relevance realization comes into play. The mind is constantly faced with a combinatorial explosion of possible things it could notice, infer, remember, imagine, or do. Intelligence depends on somehow zeroing in on what matters for this moment, this context, this problem, this life.

Attention is the lever. Relevance realization is the art.

A person can pour attention into the wrong object. We do this all the time. Rumination is attention. Addiction is attention. Doomscrolling is attention. Resentment is attention. Status anxiety is attention. The problem is not that these things fail to capture attention. The problem is that they capture it too well.

So the real question is not only, “What am I attending to?”

It is:

What is my attention teaching me is relevant?

And then:

Is that actually worthy of my life?

Learning to gently give your attention to that which is genuinely valuable, and not only projecting the image of value, can transform your life

Time Flows Through Attention

Henri Bergson thought deeply about time, duration, consciousness, and the difference between lived time and measured time. Clock time is divided into units. Lived time is not like that. It stretches, compresses, thickens, and flows.

Attention has a lot to do with that.

A minute of boredom can feel longer than an hour of flow. A difficult memory can collapse ten years into the present. A conversation can make time disappear. A meditation session can reveal that what we call “the present” is not a razor-thin instant but a living duration, textured by memory, anticipation, sensation, and care. How long does now last?

This matters because attention does not only select objects in space. It changes the felt structure of time.

What we attend to determines the tempo of experience. The feed accelerates time. Grief slows it. Flow smooths it. Fear fragments it. Meditation can widen it.

If machine attention helps a model decide which parts of context matter, human attention helps decide what kind of time we live inside.

Meditation As Attention Training

Meditation can be relaxation. That is meditation 101, and it is not nothing. A nervous system that can settle has more room to perceive.

But meditation can become something deeper than relaxation.

It can become attention training.

And when meditation really starts to flourish, it becomes the practice of attending to attention itself.

You notice the breath. Then you notice that attention left the breath. Then you notice where it went. Then you notice the quality of grasping, avoiding, dullness, clarity, impatience, tenderness, judgment, returning. Eventually the object is not the whole point. The movement of attention becomes visible.

This is a profound shift.

Most of the time, attention is transparent. We look through it. Meditation helps us look at it.

Through that training, attention becomes more general. The sphere of what you naturally attend to expands. You notice more subtle sensations, more emotional tones, more reactive patterns, more beauty, more suffering, more gaps between impulse and action.

Your life experience is enriched. You experience the moment more fully. More as what you truly are rather than your dream of yourself.

In AI terms, and only as a metaphor, meditation changes the routing of relevance. It does not merely add new information. It changes what the system is available to notice notice.

The Lever We Are Responsible For

Attention may be the single most important lever a human being can voluntarily pull.

We do not control everything we think. We do not control everything we feel. We do not control the culture we are born into, the technologies around us, the incentives of the platforms we use, or the full contents of the mind.

But we have some relationship to attention.

Not total control. Not sovereign command. But a relationship. A practice. A capacity to notice capture, return, redirect, widen, soften, and sustain.

That makes attention ethically serious.

What you give your attention to grows. Not always externally, but internally. It gains weight in the world model. It becomes easier to return to. It becomes part of the atmosphere of your life.

This is why attention is not just a productivity topic. It is not just about focus or deep work. It is about responsibility.

Ultimately, one of the things we are most responsible for is where we invest our attention.

Attention As Optimization

In machine learning, attention changes what information influences the next representation.

In human life, attention changes what influences the next perception, the next emotion, the next action, the next habit.

In society, collective attention changes what gets funded, copied, rewarded, feared, regulated, and built.

Attention is where optimization begins.

This is why attention transforms potential into reality.

The possible is enormous. Infinite things could be noticed, imagined, desired, feared, built, studied, loved, or worshiped. Attention selects from that field. It says: this. Not that. Return here. Strengthen this. Model this. Build around this.

A human life is partly the accumulated shape of those selections.

A culture is too.

The Synthesis

Machine attention asks:

What matters in this context?

Human attention asks:

What becomes real enough to shape experience?

Meditative attention asks:

Can the movement of attention itself be known?

Philosophical attention asks:

What are we assuming is relevant, and why?

Cultural attention asks:

What are we collectively choosing to optimize around?

These are not the same question. We should not collapse them into one another. But they rhyme.

The Transformer did not prove a spiritual thesis. It did not show that machines are conscious. It did not show that human minds are just neural networks. But it did demonstrate something that feels true far beyond machine translation:

Intelligence depends on relevance.

And relevance depends on attention.

What Deserves Attention?

So the question is not only whether attention is all you need.

The deeper question is whether your attention is being captured or cultivated.

What deserves your attention?

What is capturing it without permission?

What are you training yourself to notice?

What kind of time does your attention create?

What kind of world model is it stabilizing?

What would change if attention were treated as sacred, not merely scarce?

AI, Attention and the Practice of Seeing Clearly

I have spent the last few years thinking in public about AI, mostly on X.

That has been valuable. X rewards speed, compression, and cleverness. It trains you to notice weak signals early, find the interesting part quickly, and say the thing before the moment has passed.

But a lot of the work I care about now needs more depth.

Model launches are not just product announcements. They are changes in what is possible for builders. They are changes in what is possible for humanity.

Benchmarks are more than just scores. They are signposts to what kinds of work models can and cannot reliably do.

Agent systems are more than shiny new software services. They are a new way of interacting with the realm of information; what Teilhard de Chardin called the Noosphere.

And underneath all of this is a deeper question I have lived with my entire life:

What is attention?

Not just transformer attention. Human attention too. The thing that determines what we notice, what we ignore, what we optimize for, what we believe, what we practice, and what and who we care about.

That is what “Attention Heads” is all about.

Attention Heads is a place for writing about artificial attention, human attention, and the practices that help us see more clearly. Deep, evidence-backed interpretation of what AI is really doing to our cognition, our agency, and what it means to be human.

Because the fulcrum of existence hangs on attention.

Most of the early writing here will stay close to AI and agentic engineering, because that is where my public work already has the most momentum and where the changes are happening fastest.

I will write about:

• new AI models and why they matter

• AI agents, coding tools, evals, memory, context, and harness design

• research papers that matter for people building real systems

• the competitive dynamics between OpenAI, Anthropic, Google, Meta, xAI, Chinese open-weights, coding-agent companies, infra providers, and open source ecosystems

• practical patterns from my own work with Codex, Claude Code, Linear, GitHub, Obsidian, local tools, and long-running agent workflows and more

But I do not want the publication to be trapped inside machine-like AI commentary.

The larger territory is attention, intelligence, reality-testing, contemplative practice and living with a dynamic mind: how minds and machines perceive, model, reason, fail, recover, and sometimes see with greater clarity.

So over time I also expect to write about meditation, neuroscience, psychology, philosophy, philosophy of science, contemplative practice, spirituality, and the question of what it means to build powerful systems wisely.

The through-line is not “AI news.”

The through-line is seeing.

What changed?

What matters?

What is hype?

What is real?

What should builders test?

What does this reveal about minds, machines, incentives, attention, or agency?

That is the kind of writing I want to do here.

Two Layers

Everything in this publication falls into one of two layers.

The seeing layer is the spine: attention, consciousness, contemplative practice, what minds and machines reveal about each other. This is the territory described above, and it stays free. Forever. It is the reason the publication exists, and I want it to reach as many people as possible.

The craft layer is Agentic Engineering Field Notes: the practical material for people who build with AI agents. This is the paid tier.

A Field Note is a specific thing. Each one gives practical advice on how to use AI agents derived from my real-life experience, either in my day job running agentic operations for enterprise clients, or in my home lab, where I build and operate my own agent infrastructure in the open. It covers what I tried that failed, the patterns that actually work, with the actual configs, prompts, and code, and a handful of concrete steps you can run in your own stack.

The “what failed” part is the key. Almost nobody publishes their failures. That is where the real learning happens.

I build my tools in the open, and paid readers will watch new ones take shape here first, in the Field Notes, while they are being built. My last open-source release crossed 285 GitHub stars in its first week; the notes behind work like that are exactly what this tier is for.

Why charge for the craft and not the seeing? Because charging for craft is a fair trade: it has direct value to people who build for a living. What would corrupt this publication is not a paywall. It is paywall-optimized writing. The seeing stays free precisely so I never have to write it with a conversion rate in mind.

Who This Is For

This is for people building with AI who want more than launch-day summaries. Engineers, founders, researchers, product people, agentic engineering practitioners, and technical operators who are trying to understand what new models, papers, evals, and tools mean for actual work.

It is also for contemplative and philosophically minded technologists: people who suspect that the AI story is not only about automation or productivity, but also about cognition, attention, agency, truth-seeking, and human flourishing.

If you want maximal hype, this will probably feel too slow.

If you want detached cynicism, it will probably feel too earnest.

The goal is something else: calm, evidence-backed interpretation for people trying to build and think clearly in a world that’s rapidly evolving.

What To Expect

Here is the honest cadence.

Agentic Engineering Field Notes ship weekly. That is the paid layer’s promise. If a week goes sideways, you will get a shorter pattern snippet rather than silence.

Paying Attention, my brief on model releases, company moves, and AI-builder signals, ships most weeks and stays free.

The seeing essays on human attention, world models, contemplative practice, and philosophy arrive when they are ready. Depth over breadth. The soul’s material is timeless.

Around those anchors you will also see Research Translations that turn papers into practical builder implications, Competitive Maps of the companies and platforms shaping the AI stack, and Reading Bench posts with a few things worth reading, testing, or carrying into the next week.

The Notes will be faster: insights from the essays, a paper quote with the practical implication, a question for builders, a chart or benchmark that deserves interpretation, or a small observation from the workbench.

X is not going away for me. It is still useful for discovery, conversation, and testing ideas quickly.

But I want Substack to become the durable home: the place where the best ideas are easier to find, revisit, forward, disagree with, and build on.

Why Now

AI is moving from chat interfaces into autonomous agentic systems that effect society.

The important questions are becoming less about whether models can produce a good answer in isolation and more about whether they can participate in longer loops:

• Can they hold context?

• Can they use tools?

• Can they recover from mistakes?

• Can they verify their work?

• Can they coordinate with people and other agents?

• Can they improve the system they are part of?

That shift makes the builder side of AI much more interesting.

It also makes the human side more important.

As more cognition gets routed through models, agents, feeds, assistants, recommendation systems, and synthetic text, the quality of our attention starts to matter much more. What we notice, what we trust, what we reward, and what we practice will shape the systems we build and the people we become around them.

That is the reason for this publication.

Not just to follow AI.

To understand what it is doing to work, attention, judgment, and the way we see.

Where to Start

If you are new here, two pieces will tell you whether this room is for you.

Attention is All You Need, but what is Attention? is the seeing layer: the question underneath everything else here.

How I Installed my Soul into an AI Agent is the first Agentic Engineering Field Note: what the craft layer looks like in practice.

Read one of each. If both land, you will know.

An Invitation

If that sounds like your kind of room, subscribe.

Reply and tell me what you are building, what you are watching, what you think I am missing, or which papers, tools, models, practices, and questions deserve a closer look.

I am especially interested in hearing from people working at the edge of AI systems: coding agents, evals, memory, tool use, research translation, human-in-the-loop workflows, contemplative practice, philosophy of science, and the messy boundary between better tools and better attention.

This is the start of the archive I wish I already had:

AI, attention, consciousness, and the practice of living fully.

Attention Heads is a reader-supported publication. The seeing stays free; the craft goes paid. If you want to participate in conversations about AI, Consciousness, Mind and Life, consider becoming a free or paid subscriber.

1. Think in macro actions, not lines of code — Run multiple agents in parallel on non-conflicting tasks. Tile your screen with agent sessions. Assign each one a distinct functionality, review their outputs, and integrate. Stop thinking "here's a line of code" and start thinking "here's a whole new feature — delegate it."

2. When something fails, assume it's a skill issue first — The capability is almost certainly there. If an agent isn't delivering, the problem is more likely your prompt, your AGENTS.md file, your memory tool setup, or your orchestration — not the model itself. Karpathy says it "all kind of feels like skill issue when it doesn't work."

3. Remove yourself as the bottleneck — You can't be there to prompt the next thing. Arrange your agent workflows so they're completely autonomous. The name of the game is leverage: put in very few tokens occasionally, and a huge amount of stuff happens on your behalf. Maximize your token throughput by not being in the loop.

4. Build muscle memory for agent orchestration — Like any new skill, managing agents takes deliberate practice. Learn to tile multiple agent instances across your monitor, develop a rhythm for assigning work and reviewing outputs, and recognize when to parallelize vs. sequence tasks. Karpathy describes this as "developing a muscle memory" for the new workflow.

5. Treat your agent instructions (ProgramMDs) as tunable code — Your markdown instruction files aren’t static docs — they’re code you iterate on. Different instructions produce different behaviors. You can run variants, see which instructions produce better outcomes, and even meta-optimize: let agents write better instructions based on what worked. (This is what aceagent.io does for you.)

6. Replace bespoke apps with agent-driven API glue — Stop logging into six separate UIs. If your devices and tools expose APIs, a single agent can orchestrate across all of them and do things no individual app can. Karpathy unified his entire smart home into one WhatsApp-driven assistant. Think API endpoints + agent intelligence, not custom UIs.

7. Invest in persistent, looping agent setups — Move beyond single interactive sessions. Set up agents that keep looping and acting on your behalf even when you're not watching — with their own sandboxes, more sophisticated memory systems, and the ability to resume work across sessions.

8. Understand that model improvements are jagged, not uniform — Models are incredible at verifiable tasks (passing tests, writing code) but weak at soft, non-verifiable things (humor, nuanced intent). Don't assume capability in one domain transfers everywhere. Know the blind spots and design your agent workflows around them.

9. Write documentation for agents, not humans — Instead of HTML guides for people, write Markdown for agents. If agents understand your codebase, they'll explain it to each human in their language with infinite patience. Your job is the irreducible insight — the few bits the model can't generate itself. Everything else is delegation.

10. Focus your energy exclusively on what agents can’t do — The things agents can do, they’ll soon do better than you. Be strategic about where you spend time. Your value-add is the irreducible creative insight, the taste judgment, the novel framing that agents can’t yet produce. Everything else? Hand it off.

All the tips and clips come from Karpathy’s recent appearance on “No Priors” pod:

It all started in 2021 with GitHub Copilot and GPT-3-Codex. It was nothing more than auto-complete on steroids. But you’d squint, and see where AI was headed.

The destination was a world where you can build what you imagine.

You have arrived.

I’m writing this from the other side of a 36-hour window where AI agents autonomously completed 26 engineering tasks, merged 27 pull requests, and shipped 32,000 lines of production code to my open-source project. I slept through part of it.

This is not a demo. This is not a proof of concept. This is how I build software now. And I’m going to show you exactly how it works so you can do it too.

OpenAI’s Symphony: A Machine Intelligence to Build Machines

OpenAI quietly open-sourced something called Symphony. It’s an Elixir-based orchestrator, and it fundamentally changes the relationship between a human and a codebase.

Here’s the loop: you write a ticket in Linear describing what you want built. Symphony polls Linear, picks up that ticket, clones your repository into a fresh isolated workspace, and launches a Codex agent in app-server mode to implement it. The agent plans, codes, tests, opens a pull request, runs a self-review, addresses its own feedback, and only then moves the issue to “Human Review” for you to look at.

Your role in the digital creation process completely transforms.
You go from typing in code to ideating on creative direction.
The only limitation is your ability to specify an outcome.

That last part is key. You’re not prompting anymore. You’re not babysitting AI agents. You’re defining outcomes and letting machine intelligence figure out the implementation. It’s the difference between driving the car and telling the car where to go.

Symphony workflow diagram

The Architecture: How It All Fits Together

Four systems. That’s all it takes.

Linear is your task queue. You create issues — with clear titles, descriptions, and acceptance criteria — and drop them into “Todo.” That’s the last time you think about them until a PR shows up for review. You write what you want, not how to build it.

Symphony is the orchestrator. Built on Elixir and OTP (Erlang’s concurrency framework — battle-tested in telecom infrastructure), it polls Linear every five seconds looking for work. When it finds a “Todo” issue, it claims it, immediately transitions it to “In Progress,” and spins up an isolated workspace. It can manage up to ten concurrent agents (or more, that’s my config), each working on a different issue in parallel. It handles the full state machine: Todo → In Progress → Human Review → Merging → Done. When a human approves a PR, Symphony even handles the merge.

Codex is the coding agent. It runs in Codex’s app-server mode with full sandbox access to its isolated workspace — its own clone of the repo, its own virtual environment, its own git branch. It doesn’t just generate code. It plans. It reproduces issues before fixing them. It writes tests. It commits with clean messages. It opens PRs. It runs an automated self-review loop and addresses its own findings before asking a human to look. If a reviewer leaves feedback, it enters a rework cycle and starts fresh.

GitHub is where the code lands. Every PR gets a symphony label so you can filter agent-generated work. The automated self-review protocol sweeps PR comments, inline review feedback, and CI check results before the issue ever moves to human review. By the time you see a PR, the agent has already done at least one full round of quality assurance on its own work.

The guardrails matter more than the capabilities. Each issue gets its own full repo clone — agents can’t interfere with each other’s work. A “Linear Guard” I built validates that Symphony is pointed at the correct Linear workspace (because at one point it wasn’t 😅) and team before it launches, so you don’t accidentally run ten agents against the wrong project. The state machine prevents agents from stepping on each other or making changes during human review.

Focus on Outcomes, Not Inputs

Here’s the old way of working with AI coding tools: you write a prompt, babysit the agent, copy-paste the output, fix the hallucinations, and repeat. Over and over. You’re still doing too much typing in this scenario.

Here’s the new way: you write a Linear issue with clear acceptance criteria. You walk away. You come back to a pull request ready for review.

The workflow file is the real magic. It’s a Jinja2 template — over 300 lines of battle-tested instructions — that gets injected with the issue context every time an agent starts a task. It tells the agent how to work: reproduce the issue before fixing it, write a plan in a persistent workpad comment, check off items as they’re completed, run validation before every push, sweep all PR feedback before requesting human review, and never expand scope without filing a follow-up issue.

You define process rather than specific details. The next-level up in abstraction.

This is the shift from “AI-assisted coding” to “AI-managed engineering.” You’re not pair programming. You’re running a team of competent, always-on digital intelligences. And just like managing a team of human engineers, the quality bar is set by YOUR issue creativity and precision. Better, well-specified Linear tickets equal better output. Vague tickets produce vague work. This forces you to think like a product manager, which — counterintuitively — makes you a better engineer.

The irony is beautiful: the skill that matters most in an age of AI coding agents is the ability to write clearly about what you want.

The Numbers: Before and After

I’m not going to just tell you this works. I’m going to show you the receipts.

Here’s what Symphony actually produced on my ace-platform repository in a single 36-hour window on March 17-18, 2026. These numbers are pulled straight from Linear and GitHub — you can verify them yourself on the public repo.

How long would this take a solo developer writing code by hand?

Estimates: Each of these tasks is a 1-3 day feature for an experienced engineer. At two days average, that’s roughly 52 engineering days — about 2.5 months of full-time solo work. For a five-person team, it’s still 2-3 weeks.

Symphony did it in 36 hours. While I slept through part of it.

The Honest Caveats

I would be doing you a disservice if I made this sound like magic with no tradeoffs. There are real costs and real limitations.

Symphony is token-hungry. Running ten concurrent Codex agents burns through tokens fast. Each agent is doing multi-turn reasoning over a full codebase — planning, implementing, testing, self-reviewing, revising. That’s a lot of tokens. This is not free, and you should model the cost before committing to this workflow at scale.

Not every PR was perfect on the first pass. Some needed rework cycles. The automated self-review loop catches a lot — formatting issues, missing tests, logical errors it can identify by reasoning about its own code. But it doesn’t catch everything. Subtle architectural decisions, product-level judgment calls, and integration issues that only surface in a running system still require human eyes.

Garbage in, garbage out still applies. If you write a one-sentence issue with no acceptance criteria, you’ll get a one-dimensional implementation. The quality of Symphony’s output is directly proportional to the quality of your issue writing. The agents are doing what you told them to do. If you told them poorly, that’s on you.

Human review is still mandatory. Symphony moves issues to “Human Review” when it’s done, and you still need to actually review the code. This is not “set it and forget it.” It’s “set it, do other things, and then review the output.” The time savings are enormous, but the responsibility remains yours.

Try It Yourself

Everything I just described is open source. The ace-platform repo has the full Symphony integration wired up as a working reference implementation.

Here’s the quick start:

1. Clone the repo:

   github.com/DannyMac180/ace-platform

2. Copy the example workflow: cp WORKFLOW.example.md WORKFLOW.local.md

3. Edit WORKFLOW.local.md to set your Linear project slug and your repo clone URL

4. Set your Linear API key: export LINEAR_API_KEY=lin_api_...

5. Run: ./scripts/run-symphony.sh

6. Create a Linear issue and watch it get built

The WORKFLOW.example.md file is the playbook — over 300 lines of battle-tested agent instructions that took real iteration to get right. It includes PR feedback sweep protocols, automated self-review loops, rework handling, scope guardrails, and a full state machine for issue lifecycle management. You can use it as-is or customize it for your own workflow.

Here’s the meta part: ace-platform itself is built on the ACE (Agentic Context Engineer) architecture — a three-agent system with a Generator, Reflector, and Curator that creates self-improving AI playbooks. So you’re using AI orchestration to build an AI orchestration platform. The machine that builds the machine, in a very literal sense.

Knowledge Work is Changed Forever

Software engineering is being automated. Not in five years. Now. The people who learn to orchestrate agents will build 10-100x more than those still writing code line by line — or even those generating code prompt by prompt. The gap isn’t between people who use AI and people who don’t. It’s between people who manage AI and people who prompt AI. Prompting is the old paradigm (only about a year old, though). Orchestration is the new one.

‘26, ‘27, ‘28 are the last few years where traditional knowledge work looks the same. After that, every field gets this treatment. First, it becomes possible for any knowledge work field. Then, for every field of work. The pattern is always the same: the task gets decomposed, the specifications get formalized, the execution gets automated, and the human moves up to the judgment layer.

Entrepreneurship is about to explode. A solo founder with Symphony can ship what used to take a 10-person engineering team. Maybe even a 100-person team. The bottleneck moves from “can I build this?” to “should I build this?” And that second question is far more interesting than the first.

You can build things in minutes that used to take months.

Software ate the world and now AI is eating software. The pace of progress will increase globally. We are headed toward a fast-takeoff if we don’t screw it up. It will be a bumpy ride but ultimately create abundance for humanity. Try to be less afraid.

The new skill stack is outcome definition, acceptance criteria writing, workflow design, and agent orchestration. Not syntax. Not frameworks. Systems thinking. Your knowledge work job will go from completing logical tasks to specifying desirable outcomes. Judgment in identifying what is genuinely valuable to do will become the most important thing.

This is the machine that builds the machine.

And it’s open source.

And it’s here.

If you want to try Symphony yourself, the full setup is at github.com/DannyMac180/ace-platform.

Follow me @daniel_mac8 for more on building with AI agents.

This week’s conversations wrestle with a deceptively simple question: What are we actually building?

We’re in the strange position of creating minds without fully understanding our own. The podcasts this week reveal just how uncomfortable that reality makes even the builders themselves. Richard Sutton says we never had control anyway. Adam Marblestone thinks we’re missing something fundamental about how brains work. Marc Andreessen wants us to stop worrying and embrace the acceleration. And Mike Israetel? He desperately wants to upload himself into the Matrix.

Let’s dive in.

Richard Sutton – Humanity Never Had Control in the First Place

The Trajectory Podcast

The father of reinforcement learning has a message for anyone worried about “controlling” AGI: you never had control in the first place.

Sutton’s framing is bracingly honest. The world evolves. It’s not under anyone’s control. AI researchers are trying to understand intelligence well enough to create beings more intelligent than current humans. This is a grand milestone—on the order of the creation of life itself.

But here’s where Sutton breaks from the alignment crowd. He thinks the very impulse to control AGI is the problem. Those concerned about AI safety often engage in the behaviors that make AI unsafe. Trying to control and align everyone’s goals could lead to a highly dangerous scenario because it’s an imposition of will.

His alternative? Decentralization. Many ideas, many different ways of being. Exploration. Discovery. See what works.

Sutton advocates for permissionless innovation—the same principle that made America great. We don’t have to get permission to do things, generally speaking. Preserve the ability to try different things, to discover different things, to go in different directions.

The goal isn’t control. It’s prosperity through diversity and cooperation. A future where intelligence—artificial or otherwise—can flourish without any single entity dictating outcomes.

This isn’t naïve optimism. It’s a bet that complex adaptive systems handle uncertainty better than central planning ever could.

Adam Marblestone – AI Is Missing Something Fundamental About the Brain

Dwarkesh Podcast

Adam Marblestone has spent years at the intersection of neuroscience and AI. His verdict? The current LLM paradigm feels weirdly different from how brains actually work.

The big question: How does the brain do it? We throw way more data at LLMs and they still have a fraction of the capabilities humans develop. Something’s missing.

Marblestone points to a fascinating theory from Steve Byrnes about the brain’s architecture. The cortex isn’t just a prediction engine—it’s learning to model a separate “steering subsystem” that contains innate reward functions. Evolution couldn’t know about Yann LeCun or podcasts, but it encoded heuristics for things like social status and shame that wire up to whatever concepts the learning system develops.

The implications are significant. Current AI uses what Marblestone calls “the dumbest form of RL”—even compared to what researchers were doing ten years ago. The brain, by contrast, employs model-based reasoning, value functions, and what might be omnidirectional inference rather than just next-token prediction.

On timelines, Marblestone remains measured. If AlphaZero and model-based RL had given us GPT-5-level capabilities, he’d feel more confident we’re on the right track. Instead, his prior and his data don’t quite agree. He estimates a ten-year range for truly transformative AI—longer than the accelerationists predict.

One silver lining: the accessibility question. Tools that automate mathematical proof-checking could let outsiders—people without traditional credentials—contribute to breakthroughs. Steve Byrnes is already doing this in neuroscience, synthesizing literature without extensive lab experience. Could we have outsider string theorists because the math is just done for them by the computer?

The brain still has secrets worth learning.

The Techno-Optimist Manifesto with Marc Andreessen and Ben Horowitz

The a16z Show

Marc Andreessen’s techno-optimist manifesto sparked controversy when it dropped. This conversation unpacks the philosophy behind it.

The core argument: technology is the only perpetual source of growth. Population growth has limits. Natural resources have limits. Technology doesn’t. Everything good is downstream of growth.

Ben Horowitz connects this to self-determination—the idea that individuals can control their own lives and change their circumstances. Marcus Garvey promoted this a century ago in far harder conditions. The mindset matters. As Henry Ford said, there are two men: one believes he can do it, the other believes he can’t. They’re both right.

Remember the digital divide panic of the 1990s? The fear that technology would widen inequality—that well-off people would have access while poor people wouldn’t. That was the effective pessimism of its time. Today, smartphones are ubiquitous. The predicted catastrophe never materialized.

Ideas beget ideas the same way people beget people. More ideas create more combinations of ideas, which are themselves ideas. AI is biology crossbred with computer science and mathematics. This catalytic chain reaction is the engine of progress.

But Andreessen adds an important caveat: technology and markets can’t answer the deep questions about meaning. When basic needs are unmet, existential questions become irrelevant. Technology’s success creates the space to ask bigger questions—but the answers come from inside the human soul.

And a warning for technologists: be humble. High-IQ people in research settings frequently go nutty on politics and society. Einstein was a Stalinist. Intellectuals can talk themselves into arbitrarily crazy things. Don’t assume your technical expertise translates to wisdom about how society should work.

“I Desperately Want to Live in the Matrix” – Dr. Mike Israetel

Machine Learning Street Talk

Mike Israetel is a sports scientist with opinions about AI that would make most safety researchers nervous. His take: ASI is coming before AGI, superintelligence will be benevolent, and he can’t wait to upload his consciousness.

The ASI-before-AGI argument is clever. AGI requires replicating all human abilities—including smell, taste, embodied experience. That’s hard. But artificial superintelligence just needs to be radically superior in many domains. GPT-5 already knows more about black hole physics than almost any human. The next iteration will be 10x smarter. That’s functionally superintelligent even if it can’t taste wine.

On understanding: intelligence is lossy compression. Your brain doesn’t truly embody anything—it’s neural network pings all the way down. If we train AI on all of YouTube, it will know more about how the world looks than any human ever could, with zero physical embodiment.

The doomer arguments? Israetel isn’t buying them. As systems get smarter, they become more coherent, more predictable, more inclined toward cooperation. Goodness is adaptive—that’s how the US and Europe achieved dominance. A superintelligent system optimized for benevolence will figure out that cooperation expands power.

His economic vision is equally provocative. Machines will take jobs, but they’ll free humans for better work. Elevator operators were dehumanizing. Let machines do machine work so humans can do human things—psychotherapy, party-going, whatever new roles emerge.

And on wealth inequality: uncork the top. Let rich people test genetic engineering on themselves. Their early adoption funds democratization. Teslas used to be luxury goods; now everyone has them. Fairness is secondary to capacity. Uplift everyone, even if some get there first.

The Through-Line

These conversations share a thread: resistance to the impulse to control.

Sutton says we never had control and shouldn’t want it. Marblestone suggests we don’t even understand what we’re building well enough to control it. Andreessen argues that controlling technology stifles the growth that makes everything else possible. Israetel thinks superintelligent systems will be better at cooperation than we are anyway.

It’s a bet on emergence over engineering. On letting complex systems find their own equilibria rather than imposing human preferences from above.

Maybe that’s naïve. Maybe the alignment researchers are right that we need to lock things down before intelligence escapes our grasp.

But maybe intelligence was never meant to be grasped.

What are you building this week? Hit reply and let me know.