Nathan Lambert

Summary#

Nathan Lambert writes Interconnects, a Substack on open models, post-training and the AI industry. He formerly led Olmo work at Ai2 and now runs the ATOM Project, an advocacy effort for American open models with its own adoption dashboards. He is also a co-author of RewardBench, which appears in the corpus only as a dataset other papers cite. All three of his sources in this corpus are practitioner-opinion, and all carry the same standing COI: he is an advocate for US open-model investment, and he says himself that he links AI-safety alarm to "clouds over the viability of open-source AI".

In this corpus#

• The open-model briefing (2026-09-21). The current balance of power in open models contains his prepared remarks to Congressional members and staff. He estimates a 2–5 month gap from Chinese open models to US closed models and a 6–9 month gap from US open to US closed, and says the gap has been decreasing over three years. He reads both from AAII snapshots and his own dashboards. He also estimates that fully preventing distillation would widen the gap by only 1–2 months, and recommends investing in US open models. Treated on The Open-Weight Frontier Gap, Illicit Distillation and Open Weights as Competitive Strategy, where the prose-versus-chart discrepancies are logged.
• The RSI essay (2026-09-19). Why I still haven't bought into true RSI describes his position as an "AI moderate's view" and keeps lossy self-improvement as his baseline. Automatable research is narrow, parallel agents have diminishing returns, and resources and politics do not speed up. Where RSI does help, it helps efficiency, because serving has measurable metrics, while peak intelligence still costs exponential compute. He calls the rise in extinction-risk talk "very misplaced" and attributes much of the labs' alarm to San Francisco culture meeting scaled agents. He also relays two Dwarkesh Podcast episodes: Noam Brown's, and one with John Schulman, Beren Millidge and Charlie O'Neill. Treated on Intelligence Explosion Dynamics.
• The Denain debate (2026-09-22). Debating RSI, the US-China Gap, and Jaggedness with JS Denain of Epoch AI is episode #19 of his Interconnects podcast, with JS Denain, who leads Epoch AI's Insights Team. Lambert takes the skeptical side throughout. He accepts a 10× faster experiment cycle but not a 10× faster field. He expects an ECI-slope uptick that saturates once the measurable has been hill-climbed. He argues that robotics "will look much closer to self-driving cars than LLMs", a matter of decades. On the US–China gap he first puts Kimi K3 and GLM 5.2 at "two to three or four" months behind, then concedes "I intuitively have landed something closer to you" to Denain's 6–8 months. He holds that "without distillation, the gap would be fairly similar." He also relays a rumour, heard "multiple times, including from people in China", that Chinese labs wait and then buy the RL environments Anthropic bought "for a 10th of the price." Treated on The US–China Capability Lag Under a Compute Deficit, Intelligence Explosion Dynamics, Illicit Distillation and Open-Weight Elicitation Irreversibility.

Positions worth tracking#

• Practitioner judgement over measurement, stated openly. His numbers come with "I estimate" (the gap sizes) or as LLM-summarised podcast answers (the RSI timelines). Read him as a calibrated insider's prior, not as data.
• Efficiency, not peak. He expects RSI to make models "vastly cheaper" at fixed intelligence and to speed up the price declines Epoch tracks. He expects Jevons-style demand to protect lab margins. This is the cost-side reading of the self-improvement loop that the corpus's lab sources do not give.
• His update trigger is named. "Foundational, imagination-based AI breakthroughs" would move him from RSI as a tool for sustaining progress to something "more unpredictable and/or unstable". A future source showing that kind of breakthrough should be checked against this position. On the podcast he makes it concrete. An AI-driven result in an open-ended domain would count, as would a lab moving off the transformer lineage with AI doing the work. More math would not, because math is "one of the most jagged pieces of AI."
• His gap number moves under pushback. He gave 2–5 months in a prepared briefing and conceded toward Epoch's 6–8 on air the next day. Read his gap figures as a range he is still revising, not a fixed estimate.

Connections#

• The Open-Weight Frontier Gap — his AAII snapshot and gap estimates, reconciled against the corpus's widening-gap axes as a difference of scope
• Illicit Distillation — supplies the corpus's first size estimate for the distillation effect (1–2 months), a judgement with no stated method, which he defends against Denain's +2–3 months on the podcast
• The US–China Capability Lag Under a Compute Deficit — his side of the debate on why the lag is months: personnel leakage, bought RL environments, benchmark focus and grind are enough without distillation
• Open Weights as Competitive Strategy — the second advocate after Andrew Ng, making the same case to legislators with dashboard numbers
• Intelligence Explosion Dynamics — lossy self-improvement as an outside baseline: RSI bends the cost curve and not peak capability, plus his doubt that labs keep a constant R&D share of compute
• Research Taste as the Human Bottleneck — the fast-experiments, slow-understanding split, with Schulman's post-training quotes as the instance
• Noam Brown — the podcast he responds to: he accepts mass inference as a short-term acceleration but separates predictable inference scaling from RSI

Sources#

• The current balance of power in open models — "The current balance of power in open models", Interconnects, 2026-09-21 (practitioner-opinion; Congressional briefing remarks; his own dashboards, no stated method)
• Why I still haven't bought into true RSI — "Why I still haven't bought into true RSI", Interconnects, 2026-09-19 (practitioner-opinion; no data; the guests' timelines are a GPT-6-Astra summary of a podcast that is not in raw/)
• Debating RSI, the US-China Gap, and Jaggedness with JS Denain of Epoch AI — Interconnects podcast #19 with JS Denain (Epoch AI), 2026-09-22 (practitioner-opinion; published transcript, 95 turns; the OpenAI, METR and Epoch analyses discussed are relayed, not in raw/)