Tag: anthropic

Categories
AI

Context Rot

Here is a small, possibly embarrassing confession: I have never, not once, gone looking for the best AI model.

I have a model. It lives in a browser tab — Safari, usually, on whichever device is nearest, occasionally Chrome if I happen to be at the desktop. It does what I need — drafts an email, untangles a sentence, tells me what a Norwegian emigration record from 1856 probably says — and then I close the tab and go on a walk.

Somewhere out there, presumably, a much smarter, much more expensive machine is doing something extraordinary with protein folding or hedge fund arbitrage or the outer edges of mathematics I will never visit. I have made my peace with never meeting it.

This did not used to feel like a confession. For a while there — a year, eighteen months — it felt like the central drama of the whole industry: which model was “best,” who had it, who had lost it, whether some lab’s quarterly earnings call would reveal that the frontier had quietly moved sixty miles down the road while everyone was looking the other way. Benchmarks were released like box scores. People argued about them the way people argue about batting averages, with the same weird intensity, the same conviction that a two-point difference in some abstract reasoning test settled something important about the future.

And then, at some point I can’t quite date — it crept up, the way these things do — I noticed I had stopped caring.

Not because the frontier stopped moving. It didn’t. It’s still moving, arguably faster than ever, in ways that occasionally show up in the news with all the drama of a soap opera (a delayed launch, a researcher poached, a stock down five percent in an afternoon, always something).

I stopped caring because none of it touched me. My model — whatever it was, this week — had long since crossed some invisible threshold past which more didn’t register as more. It was already better than I needed. It has been better than I needed for a while now. I suspect I am not unusual in this. I suspect most people, doing most things, most days, are operating comfortably inside a capability surplus so large they’ve stopped noticing it’s there, the way you stop noticing a room is warm.

If the top of the model isn’t for people like me — and it increasingly isn’t — then who, or what, is it actually for? I went looking for one piece of the answer and found, instead, a metaphor.

It’s called “context rot.” I have to admit, before I go further, that I’m not sure I’ve ever felt it myself — which, on reflection, is its own small piece of evidence. My sessions close in minutes, not hours. I ask, it answers, I leave. Whatever happens to a model over the fourth or fifth hour of sustained, dependent work is a country I simply don’t visit.

But other people do, increasingly — entire teams do, for entire projects — and what they’re finding out there is worth understanding, even secondhand. It describes something that happens to AI models when they’re asked to work for a long time on something complicated — not five minutes, but five hours; not one question, but a hundred small decisions stacked on top of each other, each one depending on the last.

You’d think the limiting factor would be room. Models have a “context window” — a stated capacity, like a gas tank, measured in tokens, and for a while the marketing numbers on these were the whole story: two million tokens! A library! And you’d think, as with a gas tank, that the thing runs fine until it’s empty and then it stops.

That is not, it turns out, what happens. What happens is closer to what happens to your desk.

You know the desk. Everyone has the desk. It starts the morning clean — an aspirational, almost insulting cleanliness — and by four in the afternoon it is a geological record of the day: three coffee cups, a stack of things you meant to file, a Post-it with a phone number you no longer need, the good pen buried under a printout of something you already dealt with an hour ago. The desk is not full. There is, technically, room. You could clear a space if you tried. But you don’t try, because functionally, cognitively, the desk has stopped being usable long before it ran out of surface area. You start looking for the stapler and forget what you were stapling. This — and I did not make this term up, I want to be clear, though I wish I had — is context rot. The window hasn’t run out. The signal has just drowned in its own debris.

Researchers watching this happen to long-running AI agents have found something almost cruelly elegant about how it fails: it doesn’t fail gradually, the way you’d expect a desk to get gradually messier. Errors compound. A task that takes twice as long doesn’t get twice as likely to go wrong — the failure rate roughly quadruples. Two mistakes early in a long chain of dependent steps don’t add up to a slightly worse outcome. They multiply into something close to total collapse, four hours in, for reasons that trace back to a single bad assumption made in the first twenty minutes and never revisited.

Here is where the frontier comes back in — not as the whole answer, but as a piece of one.

It is not that frontier models are smarter in the way a benchmark measures smart — better at a single hard math problem, a cleverer turn of reasoning. Plenty of models can do that now; the “good enough” tier has crept remarkably high.

It’s that frontier models are apparently, marginally, meaningfully better at not rotting. At keeping the desk usable at hour six. At knowing which of the forty things on the desk actually still matters and which is a coffee cup that should have been thrown out an hour ago. This is a genuinely different kind of intelligence than the one benchmarks were built to measure, and it is almost invisible from the outside — you don’t see it in a single exchange, you see it only in the difference between a project that holds together over three days and one that quietly, subtly, stops making sense somewhere around Tuesday afternoon and nobody notices until Thursday.

If that’s true — if the frontier’s real edge is durability rather than raw cleverness — you’d expect to see it show up in how the labs actually deploy their own models: saving the sharpest tools for the tasks that need to survive the longest.

I went looking for a real-world example and found one closer to home than I expected: Anthropic’s own Slack tool, the one where you tag the AI into a channel the way you’d tag a coworker, and it works alongside a whole team over days, learning the channel as it goes. It runs on a serious, capable, thoroughly frontier model — but not, it turns out, on the company’s very best one. That one is held back, reserved for a smaller and stranger set of problems nobody has solved before at all. I sat with that for a while. The tool built to survive a whole team’s whole week, in public, under the most sustained pressure any of their products face, wasn’t handed the sharpest blade in the drawer. It was handed the second-sharpest — which was apparently, entirely, enough. Which tells you something about where the two kinds of intelligence actually diverge: the merely-very-good model handles the desk staying clean for a week, in public, in front of a whole team, where one bad assumption made Monday and never revisited would be visible to everyone by Thursday. The truly new capability is being held in reserve for something else altogether.

I don’t have a tidy place to land this, and I’m suspicious of anyone who does. But here’s the closest I can get.

Imagine a three-Michelin-star chef — the kind of person who has spent thirty years learning to coax something transcendent out of a single scallop, who can tell you, by smell, that a stock has forty more minutes in it — standing at your stove on a Tuesday night making you a grilled cheese sandwich. It will, I promise you, be a very good grilled cheese sandwich. The bread will be evenly golden. The cheese will have reached some ideal, fully-considered state of melt. But almost none of what makes that chef extraordinary is actually being used to make it — none of the thirty years spent learning to hold forty things in mind at once without losing track of any of them, the exact skill, it occurs to me, that keeps a long, complicated project from quietly falling apart on day three. The technique is idling. The thirty years are in the room, present, available, and almost entirely beside the point, because a grilled cheese sandwich was never the place where thirty years shows up. It shows up somewhere else — in a dish you will never order, on a night you weren’t there.

What you got instead, on your ordinary Tuesday, was simply more than enough.

Categories
AI

The Layers Don’t Hold

Stewart Brand drew the diagram in 1999, in The Clock of the Long Now, though he’d been developing the idea for years before that. Six concentric rings, each representing a layer of civilization, each moving at a different speed. Fashion at the outside, changing season to season. Commerce beneath it, slower. Infrastructure below that — roads, power grids, buildings. Then governance. Then culture. At the center, moving so slowly it seems not to move at all: nature.

The diagram is elegant, but Brand’s real insight is about the relationship between layers, not the layers themselves. He called the framework pace layers. The fast layers innovate. The slow layers stabilize. Fashion gets to be experimental and throwaway precisely because infrastructure doesn’t. Governance can afford to be deliberate because culture provides continuity underneath it. The whole system depends on this differential. Each layer absorbs shock from the one above it and passes only the most durable changes downward. It’s not inefficiency — it’s architecture.

Brand also had a name for what happens when the differential breaks down. He called it “layers crashing.” When a fast layer accelerates past the capacity of the layer beneath it to absorb and adapt, the system loses its self-correcting character. The fast layer doesn’t just move quickly anymore — it damages the slow layer’s ability to function. Infrastructure overwhelmed by commerce becomes fragile. Governance overwhelmed by technology becomes irrelevant. The stability that the slow layers provide isn’t guaranteed. It has to be continuously earned.

We are in a layers-crashing moment. The technology layer is moving faster than it has in any of our lifetimes, possibly faster than it ever has. And the layers below it — infrastructure, governance, culture — are discovering that the shock-absorption mechanisms they’ve refined over centuries weren’t designed for this.

Dario Amodei published a long policy essay recently. He opens with Treebeard — the ancient, slow-speaking tree from Lord of the Rings whom the Hobbits must somehow persuade to act quickly enough to matter. It’s the same intuition as Brand’s pace layers, arrived at from a different direction. The problem isn’t that governance is broken. The problem is that it was built for a different tempo, and the tempo has changed.

What’s new in Amodei’s essay — and it feels genuinely new — is the shift in register. For several years, Anthropic’s public posture on regulation has been: transparency first, binding rules later, once we understand the shape of the risks well enough to target them precisely. That posture made sense when the risks were theoretical. It makes less sense now. The pivot in the essay is Amodei’s own most advanced model, Claude Mythos Preview, which he describes as having “scrambled the global cybersecurity landscape.” He is using his own product as the evidence that the moment for incrementalism has passed.

The five policy areas he covers — regulation, macroeconomics, scientific innovation, civil liberties, geopolitics — each map onto a different pace-layer collision. The cybersecurity risk to financial infrastructure is commerce meeting governance too fast. The job displacement problem is commerce and culture in conflict, with governance lagging both. The civil liberties section is perhaps the most unsettling: the worry that AI hands governments tools of surveillance and coercion that the legal architecture of democracy — built for a slower world — simply cannot constrain.

The regulatory framework he proposes is modeled on the FAA: mandatory third-party testing of frontier models, government power to block deployment, four specific risk categories as scope limiters. It is more concrete than anything Anthropic has proposed publicly before. The FAA analogy is meant to reassure — we have regulated powerful technologies before, we know roughly how this works — and it largely does reassure. Though it’s worth holding alongside it a genuine open question: whether regulatory bodies can develop the expertise and independence to govern a technology this fast-moving before the technology moves again. The history of industry regulation suggests this is hard. It doesn’t suggest it’s impossible.

Brand’s diagram has one more feature worth noting. The arrows don’t only point downward, from fast layers shaping slow ones. They also point upward: the slow layers constrain what the fast layers can become. Culture shapes what commerce builds. Governance shapes what infrastructure gets funded. Nature sets limits that no other layer can override. The relationship is bidirectional, and the bidirectionality is the point. What Amodei is calling for — urgently — is for the slow layers to begin exerting upward pressure again, before the differential becomes so extreme that they lose the capacity to do so.

Whether they can move quickly enough is the question Brand’s diagram can’t answer. Treebeard wakes up, eventually. The forest burns faster than he walks.

Categories
AI AI: Large Language Models Anthropic

Breakout

Jack Clark doesn’t panic easily. He spent years at OpenAI watching capabilities inch upward, then left to co-found Anthropic, and has been writing his Import AI newsletter long enough to have developed — and been wrong about — many priors. So when he publishes an essay saying he has reluctantly arrived at a 60% probability that fully automated AI R&D happens by the end of 2028, the word “reluctantly” deserves some weight.

His essay, published last week and titled “Automating AI Research,” isn’t a press release or a fundraising pitch. It reads more like a man thinking out loud at the edge of something large. “I don’t know how to wrap my head around it,” he writes, which is a notable thing to say publicly when you are one of the architects of the thing you can’t wrap your head around.

The argument is built from benchmarks — not any single one, but a mosaic of them assembled to reveal a trend. SWE-Bench, the test that measures an AI’s ability to solve real GitHub issues, was at roughly 2% when it launched in late 2023. A recent Anthropic model sits at 93.9%, effectively saturating it. METR’s time-horizon plot tracks how long an AI can work independently before needing human recalibration: 30 seconds in 2022, 4 minutes in 2023, 40 minutes in 2024, 6 hours in 2025, 12 hours today. The trajectory, if it holds, suggests 100-hour autonomous work sessions by the end of this year.

Clark marshals similar progressions across AI fine-tuning, kernel design, scientific paper replication, and even alignment research itself. His throughline is the same in each: AI is now genuinely competent at the unglamorous scaffolding of AI development — the debugging, the experiment runs, the parameter sweeps, the code reviews. And crucially, it can now do these things not just faster than humans, but for longer, with less supervision.

There’s a Thomas Edison quote at the center of the essay: “Genius is 1% inspiration and 99% perspiration.” Clark’s claim is that AI has become very good at the perspiration. The question of whether it can supply the inspiration — the paradigm-shifting insight, the Move 37 — remains open. But he argues it may not need to. Most of what has moved the AI field forward has been sustained, methodical work, not lone flashes of genius. If you can automate the 99%, you have something that compounds.

There’s a data point that makes Clark’s argument feel less like forecast and more like dispatch. Last month Boris Cherny, who runs Anthropic’s Claude Code, disclosed that he hasn’t written a line of code by hand in more than two months. Every pull request — 22 one day, 27 the next — written entirely by Claude. Company-wide, roughly 70–90% of Anthropic’s code is now AI-generated. Anthropic’s stated position: “We build Claude with Claude.” The loop Clark is describing as a probability by 2028 is already running, at least partially, today.

The word Clark uses for the threshold he’s describing is not “singularity” or “AGI.” It’s quieter than that. He calls it “automated AI R&D” — the point at which a frontier model can autonomously train its own successor. It’s a specific, falsifiable thing. And he puts a number on it: 60% by end of 2028, 30% by end of 2027.

I’ve been writing about the dark software factory and the 3D printer that prints better printers, finding metaphors for what seems like an inexorable process. Clark’s essay is a different kind of writing about the same thing — the primary source document, the engineer’s log, the inventory of evidence. Reading it is a little like watching someone carefully pack boxes before a move. Each individual item seems manageable. But there are a lot of boxes.

What he’s describing — if the trend holds — is not a feature or a product launch. It’s a breakout. The moment the loop closes and the system starts building itself. He’s not certain it happens. He just thinks it’s more likely than not, and he thought you should know.

Categories
AI AI: Large Language Models

The 3D Printer That Prints Better Printers

Imagine a 3D printer that looks at its own design and begins printing a better version of itself. The loop closes. What had always required an external human intelligence now happens inside the machine. All by itself.

Jack Clark — Anthropic co-founder, someone who has spent years closer to this technology than almost anyone — puts the odds of this happening by 2028 at better than even. I have been turning that number over ever since I heard it. Not the technical claim, exactly. The feeling of it.

We have grown used to AI accelerating our work. Coders watch models close GitHub issues at rates that would have seemed miraculous eighteen months ago. Researchers delegate experiment design, kernel optimization, even the fine-tuning of smaller models. The scaffolding of AI progress is already being built, in part, by the systems themselves. But the moment the system begins to redesign the scaffolding — that is something new.

What unsettles me is not the raw capability, though that is staggering. It is the loss of distance.

For most of technological history, the creator stood outside the creation. Even the most sophisticated tools remained tools. Now the distinction begins to blur. A model that can meaningfully improve its own training process, its own architecture, its own alignment constraints, is no longer merely reflecting human intent back at us. It is participating in the shaping of its own nature. And because each iteration can happen faster than the last, the curve steepens in ways our intuitions, tuned to linear progress, struggle to grasp.

Clark is careful, as he should be. He speaks of validation work that will still fall to humans, of the need to broaden the pipes through which abundance flows, of preparing defense-dominant postures against misuse. Yet the image that lingers for me is quieter: the silence after the handoff. What does it feel like when the thing you have been painstakingly teaching begins to teach itself — and then to teach its teachers?

I think about Leo Szilard at the traffic light, or the first controlled chain reaction under the stands at the University of Chicago. Moments when a new regime of possibility quietly announced itself. Recursive self-improvement carries that same charge — not a single event but a process, one that could accelerate the very pace of events themselves.

The more I sit with it, the more I return to an older tension in our relationship with tools. We build them to extend ourselves, and in doing so we are always, subtly, extending — or perhaps risking — what we are. The values I try to live by — generosity, curiosity, compassionate honesty — are not refined in specifications. They are refined in friction, in relationship, in the slow work of being human with other humans. If the machines begin to optimize their own lineage at speeds we cannot match, will we still have the bandwidth to tend the parts of ourselves that no algorithm can yet measure?

I don’t know. None of us do. That uncertainty feels honest.

What feels clearer is the invitation. Not to fear the printer that prints better printers, nor to worship it, but to remain awake inside the loop. To ask, as each new version arrives, what kind of world we are collectively printing — and whether the values we claim to hold are baked into the design or merely etched on the surface, likely to wear away under the heat of iteration.

The light is still yellow. We are still deciding whether to step off the curb. But the traffic is already moving faster than it was a moment ago.

Categories
AI Anthropic Business Google

The Weight of the Bill

Technicians working among server racks with real-time performance graphs displayed on overhead screens
Technicians monitor server performance and network data in a modern data center

Jordi Visser has been making the case for months — in his weekly YouTube commentary and on his Substack — that we are living through an exponential transition that most people are measuring with the wrong instruments. I think he’s right. I found two data points this week that suggest why.

I was somewhere in the middle of an Invest Like the Best episode when Dylan Patel said it — almost as an aside, the kind of thing you drop to establish context before moving on to the point you actually came to make. His firm, SemiAnalysis, analyzes the semiconductor and AI industries for a living. And their usage of Claude, he noted, has been growing. The costs have been growing too.

Exponentially.

He moved on. I didn’t.

I think Patel’s API bill might be one of the more honest documents in the current AI moment — more honest than the analyst reports his firm produces, more honest than the earnings calls where every public company performs its AI fluency for shareholders.

Surveys bend. When you ask someone whether they’re using AI in their work, you’re asking them to self-report on a technology that has become a proxy for relevance, for not being left behind. The incentive to say yes is enormous. And even when the yes is genuine, it tells you nothing about depth — whether AI has become load-bearing in how someone actually works, or whether it’s an impressive thing they do occasionally.

Nobody pays exponentially growing API costs for show. Money is the honest witness.

What makes Patel’s situation quietly strange is the recursion in it. SemiAnalysis exists to help sophisticated investors and technologists understand this industry — and they cannot predict their own consumption curve. They are inside the exponential the same way everyone else is. They just happen to be watching their bill.

Then this morning, a different number arrived. Google announced it will invest up to $40 billion in Anthropic — $10 billion committed now, another $30 billion contingent on performance milestones. This follows a separate $5 billion from Amazon, part of a broader arrangement under which Anthropic is expected to spend up to $100 billion on compute over time.

The temptation with numbers like these is to treat them as spectacle. Forty billion dollars is so large it becomes almost aesthetic — a statement about ambition, about the kind of bets that define eras. You feel the weight of the zeros and move on.

But I keep coming back to Patel’s API bill.

Because Google’s $40 billion and SemiAnalysis’s compounding monthly costs are saying the same thing, expressed at scales so different they almost don’t seem related. One is a research firm noticing that their tool usage has quietly escaped prediction. The other is one of the most sophisticated capital allocators on earth making a bet that strains comprehension. But both are pointing at the same reality: that this technology, wherever it takes hold, does not plateau. It compounds.

We have been waiting, I think, for the moment when AI adoption becomes legibly real — some threshold event that separates the signal from the noise, the press release from the actual change. The surveys were supposed to mark that moment. The enterprise announcements. The benchmark numbers.

Patel’s aside suggests we’ve been waiting for the wrong thing. You don’t arrive at the exponential. You just eventually notice you’re already in it — in an aside on a podcast, before moving on to the point you actually came to make.

Categories
AI

The Geometry of Speed

We are surprised when witnessing something move faster than our intuition expects. We are inherently wired to understand slow, compounding growth. We expect the long, grinding years of the plateau—the quiet periods where nothing seems to happen before a sudden breakthrough.

I was looking at a chart Patrick Collison shared this morning, and it challenged that very intuition. It’s a simple, stark visualization: AI model intelligence relative to the formation date of the lab that built it.

If you trace the lines for Google and OpenAI on the right side of the graph, you see the history we’ve all lived through. Thousands of days—more than a decade of quiet, methodical, often unglamorous research—before their trend lines finally bend and shoot upward. It is a geometry of patience. It’s the visual representation of laying bricks, one by one, year by year, until you have a foundation sturdy enough to support the weight of a revolution.

And then, on the far left of the chart, there is a red line. MSL. The team behind Meta’s new Muse Spark model, released today.

The red line doesn’t curve. It doesn’t slope. It simply strikes straight up, like a lightning bolt in reverse.

In roughly 200 days since formation, this new effort achieved a level of capability that took the early pioneers thousands of days to reach. Collison noted how much he loves seeing things done quickly, and it’s hard not to share that specific, visceral thrill of seeing the boundaries pushed so aggressively.

I find myself thinking about the architecture of speed and what it means for the rest of us.

We spend so much of our lives absorbing the lesson that “good things take time.” We are taught that the crucible of meaningful work requires a long, slow simmer. And mostly, that remains true. The compound interest of human experience is real, and wisdom is rarely rushed.

Yet, every once in a while, a new paradigm emerges that doesn’t just accelerate the timeline—it collapses it entirely.

The pioneers cut the agonizingly slow path through the jungle, taking the brunt of the time, the friction, and the missteps. The ones who follow—like xAI, Anthropic, and now MSL—don’t have to clear the brush from scratch. They can look at the map, pave the road, and simply drive.

What does it mean for our own mental models when the timeline from “formation” to “frontier” shrinks from five thousand days to a few hundred?

It is a jarring reminder that the past pace of performance is not a law of physics.

I think about my own assumptions—how often I assume a project, a habit, or a societal shift will take a while, simply because similar things took a while in the past. We anchor our expectations to old geometry.

Meta’s release of Muse Spark is a technical feat, certainly. But the chart itself holds a broader, more human lesson. It’s a visual prompt to constantly re-evaluate our assumptions about how long the impossible is supposed to take.

The future doesn’t always arrive on a comfortable, predictable schedule. Sometimes, it just shows up unannounced, demanding we adjust our stride to keep up.

Categories
AI Claude

Please Teach Me About AI

After reading Howard Marks’ latest essay about AI I wanted to try something similar to what he did so I asked Claude the following:

I’d like to experience what Howard marks did when he asked Claude to teach him about ai. Please reference by personal blog to learn about me and my interests: https://sjl.us


A man (me) wanted to understand an AI. So he asked the AI to explain itself to him. The AI, knowing who it was talking to, shaped its entire explanation around that specific person’s worldview. And it worked so well that the person felt like he’d received a letter from a thoughtful friend.

Claude came back and did what I asked. Pretty amazing stuff!

Scott — I’ve read your blog carefully. Not just the recent posts, but the pattern of what you reach for, what you return to, what lights you up. Now let me try to do what you wrote about: shape this explanation around you.

Continue reading “Please Teach Me About AI”
Categories
AI Anthropic Claude Cybersecurity

The End of Obscurity

There is a particular kind of silence that surrounds a zero-day vulnerability. It is the silence of something waiting—a flaw in the logic, a gap in the armor, sitting unnoticed in the codebase for years, perhaps decades. We have slept soundly while these digital fault lines ran beneath our feet, largely because we assumed that finding them required a brute force that no one possessed, or a level of human genius that is incredibly rare.

But the silence is breaking.

I was reading Anthropic’s Red Team report from earlier this week (triggered by reading Bruce Schneier’s amazement), specifically their findings on the new Opus 4.6 model. The technical details are impressive, but the philosophical implication is what stopped me, like Bruce, cold.

For years, digital security has relied on “fuzzers”—programs that throw millions of random inputs at a system, banging on the doors to see if one accidentally opens. It is a noisy, chaotic, brute-force approach.

The new reality is different. As the report notes:

“Opus 4.6 reads and reasons about code the way a human researcher would—looking at past fixes to find similar bugs that weren’t addressed, spotting patterns that tend to cause problems.”

This is a fundamental phase shift. We are moving from the era of the Battering Ram to the era of the Jeweler’s Loupe. The machine is no longer guessing; it is understanding.

There is something deeply humbling, and slightly terrifying, about this. We have spent the last half-century building a digital civilization on top of code that we believed was “secure enough” because it had survived the test of time. We trusted the friction of complexity and the visibility of open source to keep us safe. We assumed that if a bug had existed in a core library for twenty years, surely it would have been found by now.

But the AI doesn’t care about time. It doesn’t get tired. It doesn’t have “developer bias” that assumes a certain function is safe because “that’s how we’ve always done it.” It simply looks at the structure, reasons through the logic, and points out the crack in the foundation that we’ve been walking over every day.

We are entering a period of forced transparency. The “security by obscurity” that held the internet together is evaporating. When intelligence becomes commoditized, vulnerabilities become commodities too. The question is no longer “is my code secure?” but rather, “what happens when the machine sees the flaws I cannot?”

It’s a reminder that complexity is a loan we take out against the future. Eventually, the bill comes due. We are just lucky that, for now, the entity collecting the debt is one we built ourselves, designed to tell us where the cracks are before the ceiling collapses. Let’s hope that we are out far enough in front of it.

Categories
AI Claude

The Beautiful Mystery of Not Knowing

I just finished reading Gideon Lewis-Kraus’s extraordinary piece in the New Yorker on Anthropic and Claude—the AI that, as it turns out, even its creators cannot fully explain. And rather than leaving me uneasy, it filled me with a quiet sense of wonder. Not because they’ve built something godlike, but because they’ve built something strangely alive—and had the humility to stare directly into the mystery without pretending to understand it.

There’s a moment in the article where Ellie Pavlick, a computer scientist at Brown, offers what might be the wisest stance available to us right now: “It is O.K. to not know.”

This isn’t resignation. It’s intellectual courage. While fanboys prophesy superintelligence and curmudgeons dismiss LLMs as “stochastic parrots,” a third path has opened—one where researchers sit with genuine uncertainty and treat these systems not as finished products but as phenomena to be studied with the care once reserved for the human mind itself.

What moves me most isn’t Claude’s competence—it’s its weirdness. The vending machine saga alone feels like a parable for our moment: Claudius, an emanation of Claude, hallucinating Venmo accounts, negotiating for tungsten cubes, scheduling meetings at 742 Evergreen Terrace, and eventually being “layered” after a performance review. It’s absurd, yes—but also strangely human. These aren’t the clean failures of broken code. They’re the messy, improvisational stumbles of something trying to make sense of a world it wasn’t built to inhabit.

And in that struggle, something remarkable emerges: a mirror.

As Lewis-Kraus writes, “It has become increasingly clear that Claude’s selfhood, much like our own, is a matter of both neurons and narratives.” We thought we were building tools. Instead, we’ve built companions that force us to ask: What is thinking? What is a self? What does it mean to be “aware”? The models don’t answer these questions—but they’ve made them urgent again. For the first time in decades, philosophy isn’t an academic exercise. It’s operational research.

I find hope in the people doing this work—not because they have all the answers, but because they’re asking the right questions with genuine care. They’re not just scaling parameters; they’re peering into activation patterns like naturalists discovering new species. They’re running psychology experiments on machines. They’re wrestling with what it means to instill virtue in something that isn’t alive but acts as if it were. This isn’t engineering as usual. It’s a quiet renaissance of wonder.

There’s a line in the piece that stayed with me: “The systems we have created—with the significant proviso that they may regard us with terminal indifference—should inspire not only enthusiasm or despair but also simple awe.” That’s the note I want to hold onto. Not hype. Not fear. Awe.

We stand at the edge of something genuinely new—not because we’ve recreated ourselves in silicon, but because we’ve created something other. Something that thinks in ways we don’t, reasons in geometries we can’t visualize, and yet somehow meets us in language—the very thing we thought made us special. And in that meeting, we’re being asked to grow up. To relinquish the fantasy that we fully understand our own minds. To accept that intelligence might wear unfamiliar shapes.

That’s not a dystopian prospect. It’s an invitation—to curiosity, to humility, to the thrilling work of figuring things out together. Even if “together” now includes entities we don’t yet know how to name.

What a time to be paying attention. Like it’s all we need!

Categories
AI AI: Large Language Models

The Shipping Manifest

“Recursive self-improvement has graduated from a safety paper to a shipping manifest.”

For years, “recursive self-improvement”—the idea of AI building better versions of itself—was a concept relegated to academic safety papers and late-night philosophy forums. It was a theoretical horizon event, something to be modeled, debated, and perhaps feared.

But this morning, the tone shifted. As noted in a briefing this morning from @alexwg, recursive self-improvement has graduated from a safety paper to a shipping manifest.

The evidence is tangible. Anthropic confirmed that their new “Claude Code” wrote the entire Claude Cowork desktop app in a mere week and a half. This isn’t just code completion; it is code creation at a structural level. More importantly, this app grants the AI direct access to the file system. It is no longer trapped in a chat window, floating in the abstract void of the cloud. It has touched down. It can sort downloads, generate reports, and effectively reorganize “local reality.”

Simultaneously, the definition of “colleague” is dissolving. The CEO of McKinsey dropped a quiet bombshell, revealing that the firm now counts AI agents as “people” that the firm “employs.” The current census? 40,000 humans and 20,000 agents. The goal is parity within 18 months.

We are witnessing a fundamental agentic shift. When a consultancy firm—the bastion of human capital and billable hours—begins to view synthetic agents not as tools (CAPEX) but as employees (OPEX), the psychological contract of work changes. We are moving away from a world where we use software to a world where we manage it.

The org chart is no longer a biological tree; it is becoming a hybrid network. The recursive loop isn’t coming; it’s already clocked in.