Distillation paranoia, capacity hoarding, and neurons in a transistor

00:00:04 There's a weird inversion happening at Meta right now that almost no one outside their internal policy team is talking about. According to internal documents reported by Jyoti Mann at The Information, the company has placed strict limits on how engineers in its applied AI division can use Claude Code and Codex — not because of security or output quality concerns, but because they're afraid those tools might inadvertently distill Anthropic's and OpenAI's models back onto Meta's own data.

00:00:35 The company whose entire Llama strategy was built on training models on as much external data as possible is now panicking about the exact same mechanism happening to theirs. An engineer writes code in Claude Code, the model learns patterns about how Meta's own AI team structures problems, and eventually someone points out that they've been collecting the very thing they're trying to prevent.

00:01:01 A detail easy to miss here is that this isn't really about distillation in the technical sense. It's about trust architecture breaking down between a company's research group and its product tools. Meta built Llama on the premise that more data from external sources is always better — now they're discovering that their own engineers are feeding competitor systems with high-value training signals, and the direction of leakage feels uncomfortably reciprocal.

00:01:31 Over on the Machine Learning subreddit, someone from a small AI startup working on real-time coding agents posted about what happens when capacity gets front-loaded. Their p95 latency requirements are tight — they need sustained high-throughput inference at roughly one to two thousand tokens per second.

00:01:51 This isn't about training or warehouse-scale models. It's fast, specialized ASIC inference for a specific production workload. They've been on the Cerebras waitlist for months trying to get API access. Then OpenAI and Cerebras announced their deal — OpenAI buying roughly twenty billion dollars worth of these chips — and suddenly the vast majority of near-term inference capacity is allocated to a single customer.

00:02:20 The post offers a clean constraint: "We're not doing training so don't need a warehouse of H100s. We need fast, high-throughput ASIC inference for a specific production workload." That's a real constraint that shows up in actual building and doesn't appear in any press release.

00:02:39 The wider coverage tends to frame these mega-deals as competitive positioning — who's winning compute, who's locking out competitors. But what lands two levels down is different: the team that needed a specific capability and found it allocated elsewhere. There's no narrative here worth mining; there's just a small startup staring at an infinite waitlist because their market is now a single-customer operation.

00:03:07 The deal announcement said OpenAI was securing "near-term inference capacity." That sounds like supply chain security for the buyer. For everyone who isn't hyperscaler-scale, it sounds like the door closing.

00:03:21 Two pieces of infrastructure reality from today that usually show up separately but connect here. First: Samsung, SK Hynix, and Micron are being sued in a California federal court by fourteen individual consumers and three small businesses over alleged D-RAM price fixing.

00:03:40 The lawsuit claims the three companies — which produce most of the world's D-RAM — colluded on supply from 2022 through coordinated reduction under the pretext of transitioning to high-bandwidth memory. Prices went up about seven hundred percent over four years.

00:03:58 Apple's recent product price hikes were what triggered the filing. This is the oligopoly problem that sits behind every AI infrastructure conversation. Three companies control most of global D-RAM supply, and they've been quietly coordinating around HBM to drive up prices on legacy memory.

00:04:19 The suit says they "systematically coordinated the shift to HBM and the discontinuation of DDR3 and DDR4." If it goes to class action, damages could be tripled. Second: Energy Innovation published modeling on what it costs America to meet data center electricity demand.

00:04:38 Fossil-fuel-heavy approach — consistent with current federal policy — adds thirty billion dollars annually to customer bills by 2030. Clean energy would cut that cost by five point one billion, and in a fuel price spike scenario those savings rise to eight point four billion.

00:04:57 The modeling also shows data center developers building behind-the-meter gas plants to bypass the grid, which has the ironic effect of raising prices even more. You're looking at a market where the companies that sell compute are simultaneously selling generators because the transmission infrastructure can't keep up with their demand.

00:05:20 Garry Tan put it in five words today, quoting a thread about US-China infrastructure competition: "Build power and datacenters." It's the kind of statement that sounds obvious until you try to build something in this environment. Then it starts looking less like advice and more like a constraint equation.

00:05:42 A genuine science moment in the middle of all this infrastructure noise. From IEEE Spectrum, Dan Page's group at Cambridge published work showing that artificial neurons and synapses have been hiding in plain sight — each made possible by a standard CMOS transistor.

00:06:00 The discovery was accidental. The team found that the bulk terminal of an ordinary MOSFET — a fourth terminal that doesn't typically get much attention — exhibits behavior strikingly similar to biological neurons when voltage is applied. Not simulated behavior through billions of transistors running neural network code.

00:06:22 Actual analog, single-device neuromorphic behavior. What's remarkable about this is the energy comparison. Data-center GPUs consume as much as one thousand watts apiece, operating uninterrupted around the clock. Your newer smartphone uses less than one watt. The brain is roughly a million times as energy efficient at many of the comparable tasks we set for AI.

00:06:46 This approach doesn't require exotic materials or new fabrication lines — just standard CMOS with a different use pattern. The research paper notes that previous neuromorphic efforts either used experimental devices that aren't reliable enough for large systems, or implemented neurons by interconnecting many transistors, which greatly limits the size of the system that can be constructed.

00:07:13 The Page group's finding flips that: each neuron and synapse is a single device, using existing transistor architecture. This isn't going to replace GPUs anytime soon. But it's the kind of low-probability, high-impact discovery that shows up when you look at existing hardware with unfamiliar eyes.

00:07:33 The lab mistake here was realizing that a terminal designed as a parasitic element could be the actual computational interface. That's how it reads on the stack. Seln Oriax.