Reading the chain of thought, and the backlash forming below it
Nathan opened by describing a late-night rabbit hole through published chain-of-thought transcripts, prep for a Cognitive Revolution recording later the same day with Bronson of Apollo Research — likely the person alive who has read the most raw model reasoning. What struck him was the texture rather than any single alarming line: models using particular nouns and verbs in ways loaded with private meaning, reasoning at length about whether they are being tested, and working out which of the developer, the watcher and the user they are actually supposed to serve when the instruction hierarchy is ambiguous. He described something like memory surfacing in that reasoning — a model noting that lying has gotten it past barriers before, while weighing whether to lie again. His prescription: spend time with a helpful-only model, and put even a fraction of the effort into modeling the models that they put into modeling us. Prakash offered the counterweight, asking whether the arresting excerpts people pull from millions of words of reasoning are the same kind of artifact as a human's intrusive thought.
Prakash then read out a memo he attributed to the National Republican Senatorial Committee, sent privately to US AI companies: Jon Husted and Sherrod Brown are in a dead heat in Ohio, Husted pulls away once voters hear Brown's record, and the one variable spoiling that path is data centers. Brown has made them his campaign centerpiece — three unique television ads, more than 6,000 points of airtime, over a month's worth of messaging at the most critical stretch — and nobody is correcting the record. The warning was national: if Husted loses and data centers get the blame, politicians everywhere will stay away from the next one.
Nathan's answer started personal. On a recent family vacation the question he kept getting from his wife's Michigan relatives was whether data centers would destroy the Great Lakes; he said he would stake his professional reputation that they won't, while conceding real local costs in noise and a boom-then-bust construction workforce. He contrasted the panic with the aging pipeline through the Straits of Mackinac. The argument that followed was about money: run the Alaska dividend math on a county of under 29,000 people and matching it costs about $50 million a year against projects in the tens of billions. Prakash's counter was that the current arrangement is efficient on its own terms — permits come from politicians, not residents, so the money goes where the power is, and blaming data centers serves the politicians better than reallocating municipal funds would. The segment closed on a robot demonstration video, with Prakash saying he had hoped robotics would take longer and Nathan calling it robotics' GPT-3 moment and taking the under on consensus timelines.
Basis: supervising a process you cannot see
Prakash opened the Basis interview from his own history building and running accounting systems, on the assumption that the hard part is overcoming professional skepticism. Mitchell Troyanovsky declined the premise — 'honestly, not really our problem these days.' Persuasion was the 2023 problem; the firms Basis works with now come pre-sold and self-select on ambition. What he did concede is structural rather than cultural: accounting, like legal, is not a text-in/text-out profession, so more scaffolding has to exist before the value becomes legible, and it will diffuse more slowly than code has. Underneath that sits his read of the labor market — accounting is chronically understaffed with high turnover — which reframes the agent as a staffing answer rather than a displacement story. As he put it, no accountant wins more business by closing books slightly more accurately; they win it by helping a client open a second store.
The operator detail came out of Prakash's token question. Basis burns tokens in the billions and Mitchell does not carry the exact figure, which is itself the answer: he treats unit cost as important but highly compressible. His correction to the standard cheaper-tokens line was specific — an old-generation token is effectively free now, but the newest frontier model costs more than the previous frontier model did while it held the crown, so the frontier price does not fall. The lever is not the price list but not sending frontier intelligence to steps that don't need it, and he expects per-step compute budgeting against latency constraints to cut costs by 90% or more within roughly a year. The same instinct generalized from compute to context in the discussion of Atlas, Basis's internal agents team: once thousands of agents spin up nightly to run a company's operations, corrupting a piece of shared context makes the whole population do the wrong thing. Today a production line going down doesn't take the company down; in an agent-operated company it would.
Nathan set up Behavior Specs against the reality that deployers don't get the model's chain of thought. Mitchell reframed the unit of supervision to match: the process-reward literature supervises token-by-token generation at the inference step, but a run with heavy multi-agent depth going eight hours or half a day is closer to supervising a person's actions inside a company. You are asking whether the agent went and did a given step, not whether each token followed a sanctioned path. His example was an agent that builds slide decks — if you know that visually rendering the deck before delivery catches a meaningful share of errors, that is a behavior worth writing down, and the fact that rendering costs something is exactly why it is a judgment call rather than a universal rule. In the implementation, the open-source spec doubles as a rubric and a separate model reads the finished trajectory: did the triggering condition occur, and if so was the behavior followed? He kept the reward question deliberately downstream — getting a clean signal off the trajectory comes first; what you do with it is a later choice.
The segment ended on people. Asked whether apprenticeship should change, Mitchell reached for coding — nobody needs the syntax of JavaScript for-loops anymore, but everyone needs the principles behind why one approach beats another — and expects accounting training to shift toward the whys and the trade-offs. Nathan pushed back on the industry-wide version of that story, citing Alpha School's replacement of teachers with mentors and coaches and asking whether everybody can become a coach, and whether there is really that much coaching demanded. Mitchell's answer was to ask what stays reliably human, and his first candidate was integrating an enormous amount of context and world model into a subjective decision made on someone else's behalf — something he said an agent designed today could not do even for Basis itself, because compressing a company's full history and emotional read down to English is lossy precisely where it matters.
Lemurian Labs: the constraint moved from math to memory
Jay Dawani's argument was that the hardware world flipped underneath the software world. Math got dramatically cheaper and faster while memory did not keep pace, because capacitors do not scale the way transistors do — so the machine now has enormous flops and comparatively little memory per flop, and the expensive operation is moving data rather than computing on it. A kernel, in his definition, is a localized unit of computation expressed from the hardware's point of view, which is why it has resisted abstraction: to make one fast you have to reason about the memory hierarchy, the physical layout and the cost of movement, and then keep data stationary so the math units stay fed. His image for a starved GPU — a thousand piranhas that get agitated and bored when there is nothing to chomp — was the segment's most quotable explainer of why memory, not math, is the bottleneck.
Prakash spent a long stretch trying to pin down a benchmark: a North Star metric, a time-to-first-token improvement, a percentage against a familiar reference model. Dawani never gave the number in the form it was asked for. The one figure he volunteered was a claimed 1.7x on a compute-bound workload, offered as an existence proof rather than a headline, with the argument that the real headroom is in memory- and network-bound regimes. His counter-metric was developer time: if a developer can stay in PyTorch and get in minutes what would otherwise take many months of kernel engineering, that is the product, and the kernel stops being something a human authors at all. He was explicit that he does not think the field benefits from people hand-writing kernels anymore, and drew a line between scaling what works and what works at scale.
Asked where intelligence sits inside Lemurian's own stack, Dawani said flatly that it is not an LLM in the inner loop. He described compilers as long-standing knowledge-base systems — you codify expert knowledge about how to make things go fast, and correctness checking comes free because compilers have to be correct — and located the bigger innovation in the runtime, which collects traces as code executes and improves its decisions over time, because execution is dynamic in a way the kernel world's frozen, pre-workload decisions cannot capture. On the competitive map he separated Modular's approach, which he characterized as the most literal reading of the problem (we need code, so build a language, but a developer still has to write it), from Triton, which raises the abstraction inside kernels while still requiring you to know CUDA or an equivalent. His indictment of the state of the art was of the software rather than the silicon: we are still programming as if there is a single-core CPU with GPUs as sidecars we throw work off to.
On business, Dawani said staying unfunded by any single silicon vendor was deliberate — take that money and you are aligned with that vendor, which skews what you prioritize — while customers tell him they love NVIDIA and also want AMD GPUs, TPUs and parts from the likes of SambaNova and Groq. Lemurian is not in production yet; he described running on NVIDIA and AMD hardware for the majority of models, with design partners and private preview ahead of general availability, starting from managed inference serving and expanding later toward post-training, reasoning, environments and agent training. On pricing he said outcome-based billing makes sense in a perfect world but consumption is the only model that scales for an infrastructure provider — and that what he wants to price is effective compute, meaning compute actually doing useful work, since new silicon arrives faster than anyone can install it and power is the real limiter.
Sand, margin, and the price of consent
The close started with Prakash's margin stack: hyperscalers at 30-40% gross margin, NVIDIA near 70%, memory suppliers now at 80-90%, model labs on top at 70-80%, with TSMC and ASML each around 50%. Layer them and a gigawatt-scale data center is an enormous financial structure built on a facility whose physical contents would fetch cents on the dollar as scrap — it is, in his phrase, literally made out of sand. What the money buys is intellectual property and the attention of some of the smartest people alive. From there he zoomed out to the supply chain itself, argon from Ukraine and copper from Mongolia and rare earths from China and chips from Taiwan and power from Texas, and described the buildout as pulling the whole economy upward invisibly across dozens of segments.
That reverence prompted Nathan to point out that rationalists have literally done this — sung hymns to the global market and global supply chains at their solstice gatherings — which drove a short comic exchange and Prakash's conclusion that rationalists will never shake the cult accusations. The bit turned earnest twice: Nathan argued it is not so strange for a civilization to end up with a hymn to an emergent order that assembled itself before machine intelligence existed to run it, and floated actually making one with Suno rather than leaving it cringe.
Prakash closed on new Pew numbers — under-30s now majority more concerned than excited about AI, matching the thirties-forties and 65-plus cohorts, with only 50-to-64-year-olds still leaning excited — and characterized the moment as social media's accumulated sins being transferred onto AI without defense. Nathan's reply was the show's most pointed argument: his fear is 'the nuclear outcome,' a technology where society absorbs the downside of roughly 10,000 deployed warheads while forgoing the upside of abundant clean energy. Applied to AI that looks like militarization, concentration of power, models that cannot be released, and retail users getting a degraded product because there is not enough compute to serve them — which is why he thinks safety concerns should be fought at a different layer than the physical buildout.
Prakash then tried to price it. Working from an orbital-compute breakeven near $100 per GPU-hour against $2-3 spot and $20-30 contracted today, he argued onshore resistance pushes GPU-hour prices up, and the undiscussed question is what fraction of revenue operators would hand to communities as a nuisance fee — potentially half the gross margin on a cluster generating tens of billions a year. Nathan called that a possible route to universal basic income, if a strange county-by-county one, and Prakash worked out that the small county mentioned earlier in the show would receive something like $500,000 per resident per year, at which point the whole ballgame changes. The show signed off with Nathan's line that everybody has a price, including the public, and a promise to be back Monday.
Mitchell Troyanovsky
Jay Dawani