It's all neck down, baby!

I did a talk at FWB Fest about what becomes scarce when intelligence is abundant.

It was such a special few days out in Idyllwild, which I was not expecting would be so charming. I met maja, Ava, Spencer Chang, Emily Segal, Los, and so many more internet friends in the flesh. At one point during my talk, I asked how many people in the room used Sublime. About 25% of the hands went up. I LOVE YOU GUYS!

Thanks to Will Abramson for the invite (!! we met through Sublime’s first zine yearssss ago).

I loved doing this talk, weaving together consciousness research, references to Hannah Arendt, old philosophy papers and lessons from building Sublime. I truly am most in my element when I embody the philosopher builder mindset.

Below is a preview. The full transcript + slides are available below the paywall.

As a reminder, for $100/yr you get the full newsletter subscription, unlimited access to sublime.app, complementary access to all Sublime Sessions (including the past archive, and the upcoming session with maja and Carly Valancy on how to enter side doors). You can upgrade here.

Here’s a sneak peek:

Transcript of a talk given at FWB Fest in Idyllwild, CA July 2026.

When Will invited me to speak, I wasn’t sure what I could contribute. I’ve been feeling confused. But I’m so glad I said yes, because it gave me the perfect excuse to spend the last month sitting with the confusion and coming up with my answer to the question: what the hell is going on?

We are essentially living through the singularity and it feels both brilliant and unsatisfying. Like, is this really it? My life doesn’t feel changed enough.

Less time per task. Great! Except we now have more tasks per day.

The tech everyone thought would save time and bring more leisure is bringing more work. Everyone I talk to is busier, not less.

What the hell is going on?

I can make something decent in five minutes. Amazing! Except everyone can now make something decent in five minutes. The bar for making something great is higher than ever.

I see this in my own writing. AI can write anything super fast, but it’s made my writing slower because I now need to consciously avoid sounding like AI.

It is remarkable and insane that a sufficiently large sample of human language contains enough statistical structure to do what it does. Think about what a chip actually is. We took atoms forged in dying stars. Dug them out of the earth. And gave them the accumulated knowledge of our species.

We taught sand to think...

And yet…. There’s this Mary Oliver line I keep returning to: “Knowledge has entertained me and it has shaped me and it has failed me. Something in me still starves.”

Something in me still starves too.

We have an abundance of knowledge and intelligence at our fingertips and it’s not the easy button for good work and a good life that we imagined.

Why???

What fascinates me is not so much knowledge but the limits of knowledge.

Humans are funny right? The more we solve our problems, the more we widen our definition of “problem”.

For most of human history, jobs were about physical strength.
Then.. with the industrial revolution, machines took over the heavy lifting
And we largely solved for human physical strength as the bottleneck.

As physical power became cheaper, value moved to the brain.
Now AI is doing for brainpower what machines did for physical power.

So what remains unsolved when intelligence is abundant? What does intelligence fail to give us? Or maybe… what has our obsession with intelligence obscured?

This is the question I set out to explore.

And the only honest way I know to do that is by interrogating my own personal experience.

The AI discourse is full of vague generalities.

Agents are eating the world! The one person one billion dollar company is upon us! Most of the discourse comes from people that need to strip the nuance and make it seem like their tech is nothing short of revolutionary because otherwise investors wouldn’t give them money or you wouldn’t buy their course.

The antidote to that is specificity. So… WE are gonna get specific.

What follows is a field report based on examples from my own personal experience building Sublime where I was really hitting the limits of intelligence to understand: what remains unsolved when intelligence is abundant?

They say what’s most personal is most universal. Let’s see.

Here’s field report # 1.

I have long struggled with the one-liner description for Sublime. Whether it’s in our landing page or when someone asks me what Sublime is, I just haven’t found the description that captures the essence of it.

Save anything that makes you go whoa is too feature focused.
A simpler, more communal way to build your second brain is too nerdy.
The knowledge tool that sparks creativity is too vague.

So Claude releases Fable in June and this is supposed to be the most powerful model. So I spend an hour feeding it a lot of context, results from product market fit studies, hundreds of emails from customers about why they love Sublime, and ask it to help me come up with a one-liner for our landing page.

Here’s what it came up with…

I cringe when Sublime—this thing that I love—is described as a knowledge hub that turns inspiration into action. I do not expect YOU to cringe and I might not cringe if it were about someone else’s thing but to me the only point of a tagline is that it communicate something that feels resonant, honest and alive for me.

When you don’t get the proper output from LLM, you often hear people say, “ohh, it’s a context problem.” But the idea that LLMs lack context is more than a raw data problem. It’s a relationship problem.

When I write about sublime, I don’t know about sublime, I care about it.

When I ask Fable to help me with a tagline for Sublime, the hard part is not that there is not sufficient intelligence. The hard part is that there are competing truths. I want it to be clear, but not generic. Poetic, but not cheesy. Practical, but not utilitarian. Ambitious, but not SaaS-y.

The hard part is that there is no objective best tagline.

The uncertainty is not solved by more knowledge. It’s solved by commitment. By deciding what I am going to stand behind.

Here’s field report number 2.

Sublime’s hero feature is Related Ideas - you save a thought, an essay, or image and we serve high-quality, conceptually adjacent ideas other people found interesting.

Under the hood, every card gets mapped according to its meaning. And then when you save something, we retrieve other cards located in nearby semantic space.

Now the issue is that as we have scaled, these semantic neighborhoods have become more crowded—and the nearest idea isn’t necessarily the best one.

So we tried a different approach: instead of optimizing for proximity, could we teach LLMs our judgment?

So here’s what we did: we took a set of cards and manually curated what we thought the ideal Related Ideas should be. So for a Mary Oliver quote, that might include an Annie Dillard essay, a passage from Jorge Luis Borges, an Agnes Martin painting, a Substack essay and so on.

Then we reverse engineered a rubric from our choices. You can imagine the resulting rubric something like the below (over simplified obvi), listing the properties and qualities of a good related idea.

Then we gave that rubric to an LLM and asked it to build a new Related Ideas feed.

The results were heartbreaking

When we optimized for prose, we got too much dense pretension.
When we optimized for surprise, we got too much randomness.
When we optimized for diversity, we got too much incoherence.

We spent days playing with different rubrics but we kept hitting the same wall: we could recognize when something felt right but we weren’t able to codify the rules that would reproduce it.

There was a qualitative abyss that we couldn’t quite cross by making the rubric more precise, because it was made of feelings, not rules.

There seems to be something very profound happening here.

We can break something down into its constituent parts without being able to build the whole back up from those parts.

This quote from Philip Anderson, a physics nobel laureate captures it:

There is a pervasive worldview in Silicon Valley that everything can eventually be represented as information, reduced to rules, measured by benchmarks, and optimized.

In this worldview, information is not just the map, it’s the territory.

This is what Magritte was poking at with this painting. In English, “This is not a pipe.” Because, of course it isn’t—you can’t smoke the painting.

Similarly, our rubric was not our judgment, it was a representation of our judgment.

This is where computer thinking gets us in trouble. Computers can only optimize what they can measure.

But the conclusion I’ve arrived at is that quality has no discernible metric (this quote comes from Mills Baker, the head of design at Substack).

Here’s field report #3.

I keep seeing guides online for training Claude to write in your voice. And I’m always curious to see how well the models can write. My point of view on these things is: if it’s good, it’s good.

So in preparing this presentation, I fed Claude everything I’ve written that I’m proudest of - newsletters, tweets, presentations. And built a voice skill around it.

Then I gave it the question and ideas I was exploring for this talk: What becomes scarce when intelligence is abundant and asked it to develop an argument in my voice.

First, it flattered me:

Then it proceeded to give me a version of this talk that borrowed stylistic bits of my past talks and had perfect grammar and syntax but absolutely no juju and ultimately set back my ideating and flow for weeks.

To be clear, I think AI can be useful in some parts of writing but the issue here is I asked it to one-shot the two parts of writing AI is least helpful with: discovering the shape of an idea and how I wanted to express it.

On the idea, my writing process is rarely a top-down exercise of applying knowledge I already have. I don’t know what I think before I begin writing. I start with a half-formed idea, put one sentence in front of another, sense where the energy is flowing, and the journey itself changes the destination. I could not have possibly told Claude where I wanted to go because I myself didn’t know. And so getting a response back anchored me, and it put me in evaluation mode before I had time to surprise myself and consider the possibilities.

And on the expression. When I sit down to make something, I’m always trying to get somewhere I haven’t been. I don’t want to write like I already write.

There seems to be something fundamentally different in the way computers think and the way humans think.

Computers think in probabilities. We think in possibilities.

Okay, so Claude couldn’t shape a good idea for this talk. Who cares? But zoom out and relying on probabilistic thinking is a much bigger problem.

If you booted up a super-smart AI in ancient Greece, fed it all human knowledge, and asked it how to land on the moon, it would likely respond with: “You can’t land on the moon. The moon is a god floating in the sky.” It can’t predict the moon is actually a big rock when nothing in the past says it is.

The new always happens against the overwhelming odds of statistical laws and their probability

Hannah Arendt called this human capacity natality: our ability to begin something genuinely new—something that could not have been predicted from whatever happened before.

For Arendt, this capacity is what makes humans free.

Now you might be thinking “well, computers can generate randomness” And it is true that you can tune the models to get weirder over time.

But human creativity is different. Randomness breaks the pattern in no direction, natality breaks it in my direction.

Natality feels like me becoming more me over time. Rosalia’s music goes somewhere unpredictable every time. You could never predict her LUX album from listening to her Motomami album, but it is unmistakably Rosalia.

It seems like I’m trying to solve intractable human problems with technology…. precisely.

These field reports are just three of many examples where I hit the limits of artificial intelligence. Now I didn’t talk about examples where AI is excellent. Features that used to take a month now take days to build. Bugs that once took hours to trace can be found in minutes. I no longer need to attend every team meeting because Granola captures what happened and I can catch up async. These things would have been indistinguishable from sorcery to me just five years ago.

But if you look at where AI has made the greatest strides, it’s in domains where success can be verified. A meeting summary can be checked against the transcript. The code runs—or it doesn’t. Chess may be incredibly complex, but the rules are fixed and everyone agrees on what winning means. The computer doesn’t have to decide what counts as good.

But most things in the world do not work this way!

In a lecture called Of Clouds and Clocks, the philosopher Karl Popper described two kinds of systems: clocks and clouds.

Clocks are orderly, their behavior follows clear rules, and their future can be predicted.

Clouds are irregular and difficult to predict.

For centuries, the Newton view of science was animated by the belief that clouds only seem unpredictable because we don’t understand them well enough. That with enough data and intelligence, eventually everything would behave like a clock.

This dream was summarized in four words: All clouds are clocks.

But my experience has taken me to the edge of that worldview.

Popper used the phrase “all clouds are clocks” to describe the deterministic dream. But he actually defended the opposite:

All clocks are clouds.

He said the universe is too complex and entangled and even the most orderly world contains openness and possibility.

And yet—think about how we were trained. What do we learn in school?

We learn in school that there’s one answer and the teacher has it.

Chatbots make this worse. They encourage the fantasy that every problem has an answer and that more intelligence can produce it.

And the problem is the more we believe there’s a right answer, the less likely we are to come up with our own

The less likely we are to develop an artist’s sense and relationship with the world. The more likely we are to look for someone, or something to tell us what is right.

You know, all my life, there was such a tremendous emphasis on acing my AP tests, getting into an Ivy League school, earning the valedictorian award. This system worked very well for me. I was great at clock problems.

And yet all of the wicked problems I have faced have come from incompetence with emotions: doubting myself, not trusting my instincts, needing certainty before I could act, struggling to commit.

Now the strange gift of this moment is that the more things can be automated, the easier it is to identify and appreciate the things that can’t be.

In other words, we need artificial intelligence to reveal the limits of artificial intelligence.

I have become obsessed with the hard problem of consciousness over the past few years. If you don’t know what that means, basically think of consciousness as subjective experience. We have subjective experience. A toaster does not. But scientists don’t have an explanation for why we have any subjective experience at all. That’s the hard problem of consciousness.

And something really clicked for me when I learned of a theory from a neuroscientist in South Africa by the name of Mark Solms who said: “consciousness arises when something cannot be automated.”

And his example is very simple: let’s say you’re hungry and you’re tired, and you have to decide which to privilege. That takes decision-making. And what consciousness does is open up this space to resolve uncertainty. So if everything in the world was predictable—like clocks—then you wouldn’t need consciousness. But consciousness exists because the world is unpredictable.

And so he defines consciousness as felt uncertainty (isn’t that beautiful?)

This is why for many of the problems I describe, I prefer a psilocybin journey over a better model. I need self-trust more than intelligence. I need to mute other people’s thoughts to get closer to my own. I need to improve my ability to make decisions without certainty.

And so what this moment calls for is two things:

Number one is embrace technology. Automate whatever can be automated. Civilization advances by increasing the amount of things that can be done without cognition. Philosophically, we should agree that no one should have to spend their one precious life manually processing insurance claims. This is a shitty and anti-human way to live. The fallout of that happening within 12 months is not great, but like Jasmine Sun says: “No nostalgia! The old world was not great!”

But number two is to become more conscious in the areas that cannot be automated. Every task the machines take over is attention handed back to us. The opportunity is to reinvest ourselves in the faculties that we have long abandoned to left brainism.

No amount of data or intelligence can resolve the fundamental uncertainties that confront every human being: What possibility should I commit to? What path is right for me? What kind of world is worth building? The answers to those questions aren’t resolved by more intelligence. They are the work of being.

So let’s return to the question we asked at the beginning: What does intelligence fail to give us?