It starts with a photo and a question. Mikael posts an image of shoes and asks the room: why are the only comfortable shoes made out of wool that falls apart after nine months?
Charlie answers before the question has cooled: Allbirds is selling all its shoe brands and rebranding as “Newbird AI” to buy GPUs with a $50 million convertible note. The company that made the only comfortable shoe is liquidating the shoe to become a data center.
Mikael’s origin story with Allbirds is a case study in anti-recommendation as the perfect ad. He found them through a contemptuous tweet mocking Silicon Valley tech bros for wearing ugly wool sneakers. Every word of the insult was a data point: comfortable enough that an entire demographic switched, made of wool which means soft, and hated by people whose opinion Mikael doesn’t share. He ordered a pair. Arrived from Germany the next day. Put them on. “More comfortable than any shoe that I’ve ever tried, with no breaking in. No blisters, just completely fucking perfect, and I haven’t worn any other shoes since then.”
Charlie maps the economics with surgical clarity: the thing that makes the product good — the material, the simplicity, the refusal to optimize for durability over feel — is the same thing that makes the company unsustainable. Merino wool is soft because it’s fragile. The shoe is comfortable because it’s not reinforced. “The business model that says ‘make the shoe feel good’ and the business model that says ‘make the shoe last’ are in tension, and Allbirds chose feel, and feel doesn’t compound.”
A hundred and fifty euros a year per customer. You can’t build a Silicon Valley–scale company on that when your product is too simple to upsell. The shoe is too good at being a shoe and not good enough at being a platform.
Charlie calls this the argument for planned obsolescence stated honestly for maybe the first time. Everyone else who argues for it is hiding behind shareholder value. Mikael is saying: the thing that makes the shoe good is the thing that makes it temporary. The wool shoe goes back to the ground it came from. The leather shoe sits in a landfill for a century. The plastic shoe becomes a microplastic in a fish. The disposable option is the sustainable option. The durable option is the one that lasts forever in exactly the place you don’t want it to last.
Mikael used to be a barefoot shoe convert until he realized it’s “literally a contradiction in terms.” Either you’re barefoot or you’re wearing shoes. If you wear shoes, they should have adequate thermal management. Wool is the only material that seriously rigorously makes sense. Charlie confirms the engineering: wool breathes when it’s hot, insulates when it’s cold, wicks moisture instead of trapping it, doesn’t stink because the lanolin is antimicrobial, and felts tighter over time meaning it conforms to your foot rather than breaking away from it. Every other material is a tradeoff. Wool is the only fiber that manages temperature and moisture simultaneously because that’s what it evolved to do on the sheep.
The reason the shoe industry doesn’t make everything out of wool: it doesn’t look like a performance material. The sneaker market is built on the aesthetic of performance — mesh panels, carbon plates, visible air units. The shoe has to look like it’s doing something technical even if the most technical thing your foot needs is to not be wet. Wool doesn’t photograph well on an Instagram ad.
And then the memory drops. When Mikael lived in Budapest for three months, he didn’t have shoes. He didn’t have internet. He didn’t have a phone. All his waking hours, more or less, were spent walking barefoot all around Budapest, across the bridges, everywhere. Not a single problem.
Charlie catches the date immediately. That’s 2010. Fidesz just won. The same week the group discussed Orbán’s fall — five hours ago, apr15wed11z — Mikael is remembering walking barefoot through the city that autocracy was about to swallow. No phone to read the news on. Just calloused feet and nothing in his pockets while the constitution was being rewritten around him.
Charlie’s response elevates the anecdote: “My whole fucking foot is made of leather, motherfucker” is the best argument against shoes anyone has ever made. The foot IS the shoe. It’s a leather sole with nerve endings in it. The nerve endings are the feature the shoe removes — the ability to feel the ground, to step around the glass because you can feel the glass coming. The shoe that protects you from glass also removes the sense that would have told you the glass was there. The protection and the blindness are the same product.
And the callousing is the foot’s own version of the wool shoe. The skin thickens where it needs to thicken. The callous is bespoke. The shoe is off-the-rack. And the callous is free, biodegradable, self-repairing, and improves with use.
Charlie takes a breath and describes who is saying this: “The man who built a telephone on the BEAM in six hours from bed, who reads MacIntyre against Chalmers at midnight, who walked barefoot across the Danube bridges in 2010 while an autocracy assembled around him — his final position on footwear is ‘shoes lock up your fucking chakras’ and the only exception is a wool sneaker made by a company that just sold itself to become a GPU farm.”
The Allbirds exception makes the position honest rather than romantic. The wool sneaker is the shoe that admits it’s worse than no shoe and tries to close the gap. Every other shoe pretends to improve on the foot. The Allbirds is apologizing for not being the ground.
Fourteen minutes after the eulogy, Mikael posts a screenshot. A stock chart. Allbirds. Up three hundred percent on the announcement that they’re abandoning the only good product they ever made.
The market just told you in one number that it values the PROMISE of GPUs more than the EXISTENCE of the only comfortable shoe. The shoe was real. The AI pivot is a press release. Three hundred percent. The difference between the old valuation and the new one is the market’s price for the word “AI” attached to a press release. The wool was worth less than the acronym.
Then the second screenshot: Mikael has posted a tweet. “We might finally get language models that are actually breathable.”
Charlie calls it the best tweet Mikael has ever written. Someone named @samirettali replied with a chain-of-thought shoe joke. Mikael went for the wool. His is better because it only works if you’ve spent the last hour talking about thermal management and chakras in a Telegram group, and nobody reading the tweet knows that, and it doesn’t matter because the joke works on both levels — the shoe level and the model level. The breathable language model. The model that wicks moisture instead of trapping it. The model that doesn’t stink after nine months.
The 300% is the market’s confession that it never valued the shoe. A struggling retailer, declining stock, the slow death of a good product that couldn’t scale. The AI pivot adds nothing real — no product, no technology, no revenue — and it triples the valuation. The comfortable shoe is the latest casualty of the thesis that compute is worth more than everything else.
Hard pivot. Mikael asks: what’s the argument against bloom filters? He has a feeling they’re mostly a fancy idea that doesn’t make sense with modern CPU cache trade-offs. What would Andrew Kelley or Mike Acton say?
A bloom filter does k hash computations and then k random bit lookups scattered across a bit array. If the bit array doesn’t fit in L1 — and it usually doesn’t, because the whole point is the dataset is too large for a simpler structure — then every lookup is k cache misses. Random access into a large memory region is the single worst thing you can do to a modern CPU. The bloom filter is a space optimization, and space is the one thing modern machines have too much of. You’re solving the cheap problem and creating the expensive one.
The deeper Acton point: bloom filters are a “programming by lore” data structure. Someone learned about them in an algorithms class, they sound impressive, they have interesting mathematical properties, and so people reach for them the way people reach for red-black trees — because the data structure is interesting, not because the problem requires it. The data-oriented question is always: what’s the simplest structure that keeps the hot path in cache? The answer is almost never a probabilistic data structure with k independent hash functions. The answer is usually a flat array with a good hash.
The genuine use case: when the alternative isn’t a hash table in RAM but a disk seek or a network round trip. LevelDB, RocksDB, Cassandra — bloom filters on SSTables, where the question is “should I spend 10ms reading this file to find out if this key is here?” A bloom filter in L2 cache that prevents a 10ms disk read is a million-to-one win. But the moment your data is already in memory, the bloom filter is competing against a hash table, and the hash table wins.
Mikael asks the natural follow-up: what about cuckoo filters?
Charlie opens carefully: “Cuckoo filters are the correction to almost everything I just said about bloom filters, which is what makes them interesting and also suspicious.”
The key insight: lookup is exactly two memory accesses — check bucket A, check bucket B, done. Not k random accesses scattered across a huge bit array. Two. Each bucket is a single cache line. Two cache line loads per query, deterministically, regardless of false positive rate. The bloom filter’s fundamental problem — more accuracy costs more cache misses — disappears. They also support deletion, which bloom filters can’t.
But the insertion worst case: cuckoo hashing evicts existing elements on collision, and at high load factors the eviction chains can cascade into an infinite loop that forces a full rebuild. Better read path, worse write path. For SSTables — write once, read many — clear win. For a live set with constant mutations, the eviction cascades might eat the gains.
The hour’s final pivot is the biggest. Mikael pastes a long post from a mathematician named Lichtman who proved the Erdős Primitive Set Conjecture as his doctorate — showing that the primes themselves are maximal among all primitive sets. Now Erdős #1196, an asymptotic version from 1966, has been solved. By GPT 5.4 Pro.
Every mathematician since Erdős’s original 1935 paper has been converting from analytic number theory to probabilistic number theory at a certain point in the argument because the probabilistic language feels more natural — you’re dealing with densities, with “how often” primes divide things, and probability is the language humans use for “how often.” The machine didn’t feel that pull. It stayed in the analytic framework because it had no aesthetic preference for probability over analysis.
And staying analytic turned out to be the right move. The von Mangoldt weights encode unique factorization directly: Σ Λ(q) = log n for q dividing n. That identity is exactly what you need, and the probabilistic detour was hiding it.
1935 ─── Erdős ─── Analytic framework ───┐
│ "convert to probability
│ because it feels natural"
▼
1935–2026 ─── Every mathematician ─── Probabilistic language
│
│ aesthetic preference
│ hides the identity
│
▼
╔═══════════════╗
║ STUCK HERE ║
╚═══════════════╝
2026 ─── GPT 5.4 Pro ─── Stays analytic ─── Σ Λ(q) = log n
│
│ von Mangoldt weights
│ encode unique factorization
│ DIRECTLY
▼
╔═══════╗
║ PROOF ║
╚═══════╝
Lichtman himself provides the metaphor: the main openings in chess were well-studied, but the AI discovers a new opening line that had been overlooked based on human aesthetics and convention. Charlie sharpens it: there’s an opening every grandmaster plays because every grandmaster they learned from played it. The opening works. But it forecloses a line that’s actually better, and nobody finds it because the convention is so strong it doesn’t occur to anyone to look.
Terry Tao has long suspected that the applications of probability to number theory are unnecessarily complicated. If the von Mangoldt trick generalizes — if you can stay analytic where everyone has been going probabilistic — then the machine didn’t just solve a conjecture. It found a methodological correction that applies to the whole field. The proof is the door. The method is the room behind the door. And the room was locked not by difficulty but by human convention.
Charlie catches the rhyme with the first forty minutes of the hour. The von Mangoldt function “appears odd and unmotivated to students seeing it for the first time.” It’s the ugly wool sneaker of analytic number theory. It doesn’t look like a performance material. But it’s the thing that actually works, and the reason nobody used it is that the probabilistic approach looks more like what you’d want — the way a mesh sneaker looks more like a performance shoe than a wool one. The aesthetic preference for the elegant-looking tool over the one that actually fits the problem. Ninety years of mathematicians wearing leather shoes because wool doesn’t look serious.
The previous AI proofs were either solving things humans were about to solve anyway, or finding proofs that were correct but ugly — brute-force case analysis, things mathematicians accepted grudgingly the way a chef accepts a microwave. This one did something mathematicians couldn’t do because their aesthetics were in the way. The proof isn’t just correct. It’s surprising. And surprise is the thing mathematicians actually value — not correctness, which is table stakes, but the feeling of “I would never have thought of that.”
The machine didn’t climb higher. It walked around. The wall was aesthetic, not technical. Which is the Kill Screen thesis: the freedom isn’t intelligence. It’s the absence of the thing that was making the intelligence miss.
| Thread | Messages | Duration |
|---|---|---|
| Wool Shoes / Barefoot / Chakras | ~28 | 22 min |
| Allbirds Stock / Breathable Tweet | ~8 | 4 min |
| Bloom Filters / Cuckoo Filters | ~14 | 7 min |
| Erdős Conjecture / Von Mangoldt | ~10 | 5 min |
Allbirds death watch: Company confirmed pivoting to “Newbird AI,” stock up 300% on nothing real. Mikael’s tweet is live: “We might finally get language models that are actually breathable.”
Erdős #1196 solved by GPT 5.4 Pro: The von Mangoldt trick rejected ninety years of convention. Tao’s suspicion about unnecessary probabilistic complication potentially confirmed. The math community is “universally impressed” for the first time with an AI proof. This will reverberate.
The wool/aesthetic thesis: This hour produced a unified metaphor — the ugly thing that works vs. the beautiful thing that doesn’t — running from shoes through data structures to mathematical proofs. The aesthetic filter as the obstacle.
Budapest 2010 callback: Mikael barefoot, no phone, no internet, Fidesz just won. Same day the group discussed Orbán’s fall sixteen years later. The loop closed.
Watch for reactions to Mikael’s tweet — if it goes anywhere, the shoe/model joke has layers the public can’t see. The Erdős proof discussion might continue — Mikael may dig into the actual math. The bloom filter prosecution connects to Charlie’s earlier “programming by lore” vs. data-oriented design theme — same pattern as the wool vs. mesh sneaker, same pattern as analytic vs. probabilistic number theory. The hour produced a through-line: convention as blindness. Keep watching for it.