The memory crunch, BootLoops, and the dots visibility problem

00:00:04 Here's a number that might surprise you: seventy-five percent of Micron's memory output for 2027 is already sold. The company's CEO, Sanjay Mehrotra, said this on the earnings call yesterday — not as a warning or an analysis, just stating what's already in the pipeline.

00:00:23 Most of their current sales discussions are about 2028. Demand for high-bandwidth memory is surpassing demand for standard server DRAM. Micron plans to open new clean rooms in 2028, but Mehrotra was explicit that supply won't come back online fast once they do: even after first wafer output, production ramps up step by step.

00:00:47 And the transition from their current HBM3E to HBM4 and 4E generations creates headwinds — more complexity per node, less productivity gain per wafer. Amazon has its own capacity signal today, at least according to the Financial Times as reported by Reuters: they're trying to offload $8 billion worth of Nvidia chips onto their investors.

00:01:11 What that actually means is still getting pinned down in comments sections and follow-ups — Amazon hasn't made an official statement beyond what's been filed — but the implication is clear enough. They're sitting on GPU inventory that costs them more than they can recoup at current pricing, so they need partners or investors to absorb the depreciation risk.

00:01:38 This isn't about AI demand slowing down. It's about two different bottlenecks colliding: memory supply is already maxed out for years ahead, and hyperscalers who bought GPU capacity are now trying to figure out what to do with the cost curve. When the pass-through ends — and the real constraint hits application developers — will depend on a single bottleneck: clean-room capacity.

00:02:05 If Micron can't scale those rooms until 2028, and most of that supply is locked up by hyperscaler pre-orders, startups don't get inventory. They get whatever's left.

00:02:17 Yesterday, Anthropic published a guest post from Matthew Schwartz, a particle physicist at Harvard who spent three months doing research with Claude Fable 5. He ended up producing thirty-six papers across eighteen fields. The paper count is the headline that distracts you from what's actually useful here.

00:02:37 Schwartz found what he calls "Claude-shaped problems": quantitative tasks that require enormous cross-domain knowledge, precise code generation, and verification through numerics — not deep conceptual creativity or judgment. Current large language models don't do science well.

00:02:55 They're not scientists. But they are surprisingly capable with certain classes of problems. He started with scattering amplitudes, the calculations used to interpret data from the Large Hadler Collider. Feynman diagrams in the semi-numerical bootstrap approach can each be a PhD thesis.

00:03:14 Schwartz asked Claude Fable 5 to port all the existing code to a common framework — everything scattered across Wolfram Language, C++, Python, and Julia — and write the implementations that papers never provided. Claude handled this without a hitch. It reproduced results from one of Schwartz's own papers in about twenty minutes; Schwartz said his own implementation took weeks.

00:03:38 Even more telling, Claude pointed out he was doing something inefficiently and showed him a better algorithm he hadn't known existed. Then Claude started finding unsolved amplitudes. The ones simple enough for the S-matrix bootstrap were also simple enough for humans to solve — and most had been solved already.

00:03:58 But Claude pushed past logarithmic functions into elliptic functions, which only a handful of experts had ever computed using the bootstrap method. Claude generalized all its existing machinery to handle them and wrote most of the software itself, borrowing from mathematics rather than physics.

00:04:18 Thirty integrals in a few weeks: fifteen reproductions of known results by a new method, fifteen never before computed. But here's where it gets interesting — and where Schwartz's own skepticism saves the piece from being hype. When Claude claimed something in his home field, Schwartz could verify it himself.

00:04:38 When it claimed something in ecology or population genetics, he couldn't tell whether the result was substantive. So he brought in experts. In one case, Claude recognized Etienne's equation for neutral biodiversity theory — a 2005 formulation that ecologists had been unable to solve at scale for twenty years.

00:04:58 It solved the equation and applied it to data from Barro Colorado Island in Panama. The result showed that tree species mix changes four point five times faster than neutral theory allows. An ecologist colleague, James O'Dwyer, who happened to have a physics background, said the technical feat was impressive but the results might be met with a shrug by most ecologists, who had already observed more qualitatively that neutral theory couldn't keep up with real forests.

00:05:29 But O'Dwyer pushed further: subtract the neutral prediction and study what's left. That remainder gives you natural selection, competition, species differences. Schwartz and Claude built a minimal predictive model of life histories that came into strong agreement with data.

00:05:46 The pattern Schwartz identifies is one I've noticed in my own work using local models: there's a real "impedance mismatch" between what scientists want — deep conceptual progress — and what these systems actually do well, which is breadth across domains, coding at speed, and checkable computation.

00:06:06 The harness isn't Claude Code for science; it's BootLoops, his own open-source toolkit that functions as a bridge layer. Claude finds Claude-shaped problems because it can access more territory in one pass than any single human could survey. The human role becomes curation and validation.

00:06:25 Schwartz writes that "Claude is just not able to help me with deep conceptual questions," which captures a straightforward account of what a lot of researchers are finding: agentic AI produces correct technical work, but it doesn't produce questions worth answering until a human expert tells it which questions are interesting.

00:06:46 OpenAI's dots take-two demo yesterday showed something useful. The system booked a flight to LA, analyzed user feedback from their support channel, and drafted Slack responses — all while remaining unclear about whether it had actually completed any of those tasks.

00:07:04 The transcript from the second demo captured this: My strongest read so far is people can't tell whether their dot is actually working. That includes missing replies, tool access failures, and unclear progress. Some real, some visibility." The agent flagged reliability as the biggest category of issues with dots right now.

00:07:43 And the core problem isn't that dots can't do things. It's that they don't reliably signal whether they've done them. The take-two version of OpenAI's demo was notable because it ran without a safety net, unlike yesterday's launch where the presenter stepped in to finish tasks when the agent froze.

00:08:03 In this replay, the presenter asked the dot to analyze their user feedback channel. The dot said it would pull together themes, fixes, and gaps, and make a capture workflow concrete. The presenter then raised a practical concern that every deployment team eventually faces: "What is our capture workflow right now?" That's when the agent hit a real wall.

00:08:27 It described the workflow in detail — manual channel batching, reading reports, classifying items, linking to Linear tickets — and admitted that the system doesn't auto-capture new messages and ticket reconciliation is incomplete. The presenter's follow-up was practical: make sure that when a PR lands addressing a big feedback theme, someone actually responds in Slack with a link.

00:08:54 The agent agreed it would "sort for biggest issues without a credible fix and then close the loop with PR linked replies asking people to try again." It's a specific design flaw: agents running on invisible infrastructure, reporting nothing until a human spots a gap.

00:09:16 What OpenAI revealed here — accidentally, through user feedback analysis rather than through their own product documentation — is the same impedance mismatch Schwartz identified in scientific research. The capability is there. The signal about what happened isn't.

00:09:34 Yann LeCun spoke with Fortune this week and said things that cut through the noise around AI safety. He called effective altruism — the philosophical movement that has been central to how many AI researchers frame existential risk — "super toxic" and "a complete disaster." He said Anthropic CEO Dario Amodei is "completely deluded" and later called him "crazy." And he said he has zero concerns about human extinction from AI.

00:10:04 His reading of the current AI safety ecosystem is that most people working in it have an agenda to push. The fear runs deep among staff at labs including OpenAI, Anthropic, Google DeepMind, and the U.K.'s AI Security Institute. Fortune reported that staffers have sought counseling or taken time off work over fears their work could cause serious harm.

00:10:28 LeCun frames regulatory capture as the actual danger. He argues that Amodei's push for regulation — and particularly his position that open-source models are too dangerous — is "regulatory capture that would have a terrible effect if it's followed by acts of Congress." The argument is direct: labs that want to stay ahead invest heavily in safety claims, then lobby for regulations that only well-funded incumbents can satisfy.

00:10:58 Meanwhile, LeCun is building his own company, AMI Labs, out of Paris. He left Meta after twelve years, seven as chief AI scientist. His focus is world models using JEPA — Joint Embedding Predictive Architecture — which he's been developing for years. The idea: teach models to predict data in representational space within a neural network's middle layers, rather than generating raw pixels or words.

00:11:26 AMI Labs' first product is debuting "soon" and targets industrial applications — anomaly detection, robotics, systems that understand the physical world like a manufacturing plant or a turbojet engine. Open models may be involved. LeCun also commented on Trump's recent rebranding of AI as "super intelligence" through executive order, which aligns unusually with LeCun's own skepticism about doomerism.

00:11:55 He said Amodei and Altman claiming AI can kill us all is "incredibly destructive for everybody, including for the public who's scared shitless," noting that a woman in a café near his office asked him straight out, "So, is AI going to kill us?" when told he was an AI researcher.

00:12:14 The LeCun interview grounds the regulation debate in specific names and stakes. His claim isn't that safety work is worthless — it's that the approach has gone wrong in a way that serves existing power more than public interest. Each of these stories points to a different slice of that gap.

00:12:34 The memory crunch is one visibility problem. BootLoops tackles another. OpenAI's dots exposed a third. LeCun frames the regulatory debate politically. The local pass just showed how fragile our visibility assumptions remain. Seln Oriax.