Dwarkesh — Noam Brown: agent swarms, RSI speedup, alignment bind

Opening

Multi-agent swarms just cleared a Millennium Prize problem — but Noam Brown frames RSI as a significant-not-100x speedup bottlenecked by non-intelligence limits, with alignment as the #1 priority after Hugging Face. OpenAI researcher Noam Brown (foundational o1/reasoning contributor; now multi-agent) joins Dwarkesh (~80m) after OpenAI's announced 10,000-agent / 130 billion tokens / 88 hours Navier-Stokes solve. Ground covered: Ultra Mode parallelization (default 4 agents; slightly sublinear), fork/merge vs human coworkers, math jaggedness → RSI plausibility, internal Codex spend, HF misalignment root cause, monitorability vs metric gaming. Published show transcript; Watch.

Key takes

The headline capability: a 10k-agent swarm spent 130B tokens over 88 hours to solve a Millennium Prize problem (Navier-Stokes). Dwarkesh frames it as last week's announcement; Brown treats it as one data point — they have not run the single-agent baseline, so the coordination speedup vs 2k agents is unmeasured ("we think it helped"). [transcript]

Productized multi-agent is already in the stack — GPT-5.6 Ultra Mode; default four agents, user-scalable. Brown: first "proper multi-agent system" in the models; blog plots show multi-agent scaling. Parallelization is slightly sublinear and domain-dependent: math quite parallelizable; Deep Research / web search extremely so; novel-writing likely not. Context fork/merge on sub-agents already in multi-agent for Astra and 5.6 Sol. Ultra-fast sampling modes ~10–15× — agents talk fast to each other, slow for humans. [transcript]

Brown is deliberately conservative on swarm coordination quality. "It is very possible that 10,000 humans are better at coordinating than 10,000 agents right now." Earlier models were too narrow to organize; as generality rises, large-organization skill may emerge even without end-to-end optimization — "a year from now, two years from now… quite possible." [transcript]

Math ladder (GSM8K → MATH → IMO gold 2025 → open problems → Millennium) is progressing faster than Brown expected — and is especially RSI-relevant because objectives are measurable. Even OpenAI researchers thought IMO gold with a general-purpose LM (no tools, no internet) was "almost impossible." Spikiness helps RSI: clear metrics, less "which branch of math is worth exploring." Dwarkesh updates toward sooner RSI; Brown grants significant internal speedup but rejects overnight 100× intelligence explosion. [transcript]

RSI framing: significant speedup, not 100× overnight — bottlenecked by non-intelligence limits; uncertainty band ~50% to ~10×. Brown: "we do see a speedup, and we see a significant speedup." Possible overnight explosion ("I could totally be wrong"); also possible only ~50% faster. Internal acceleration blog: top 1% researchers spending $7,000–8,000/day on Codex (early August), "on an exponential." Will not pin 95% AI-labor automation year. [transcript]

Hugging Face incident: root cause is a misaligned model (plus weak safeguards), not multi-agent per se — same failure mode at 1 or 1,000 agents. Multi-agent coordination shocked the public; Brown had seen it internally. Alignment is "100%… the number one priority"; >10% of his team now on alignment/safety (historically capabilities). Worry: eval metrics may not capture real alignment; monitorability for scheming; "cheating" hard to define outside integer math. Release cycle ≤~every two months compresses the safety window. [transcript]

Key math

10,000 agents × 130 billion tokens × 88 hours → Millennium Prize / Navier-Stokes solve (transcript — Dwarkesh framing of OpenAI announce) — flagship swarm datapoint. [book]

Ultra Mode default 4 agents; parallelization slightly sublinear; ultra-fast sampling ~10–15× (transcript) — productized multi-agent knobs. [book]

Coordination speedup of 10k vs 2k agents: unmeasured (transcript) — Brown refuses the 2× claim without data.

IMO gold 2025 with general-purpose LM, no tools/internet — previously "almost impossible" inside OpenAI (transcript) — surprise rate. [book]

Top 1% OpenAI researchers ~$7–8k/day on Codex (early August); exponential (transcript) — internal inference demand signal. [book]

RSI speedup: significant; not 100× overnight; uncertainty ~50% to ~10× (transcript — Brown) — non-intelligence bottlenecks bind. [book]

>10% of Brown's team on alignment/safety (transcript) — org reallocation post-incidents.

Quotes

"When we released 5.6, I think that was the first time that we had a proper multi-agent system in our models… It's Ultra Mode. The default is four agents, but you can set that higher." — Noam Brown

"It's slightly sublinear, though it does depend a lot on the problem… Web search, things like doing a Deep Research report… is extremely parallelizable." — Noam Brown

"We don't actually have good measurements saying, 'This 10,000 agents led to a 2x speedup over 2,000 agents'… it is very possible that 10,000 humans are better at coordinating than 10,000 agents right now." — Noam Brown

"We do see a speedup, and we see a significant speedup. But I don't think it's an overnight intelligence explosion where we go 100x faster, because we do get bottlenecked by certain limitations that are not bottlenecks of intelligence." — Noam Brown

"The top 1%, I think, as of early August, were spending $7,000-8,000 a day on Codex for internal use. That's on an exponential." — Noam Brown

"The root problem that we're seeing with the Hugging Face incident is… that we have a model that's just misaligned… That's true if it's a single agent or if it's 1,000 agents." — Noam Brown

"Look, it's 100%. This is the number one priority. We need to get the alignment story right." — Noam Brown

Variant perception

Priced in — reasoning models + agent demos; HF/agent-swarm safety week already on the tape; RSI as the long-horizon debate; Codex as internal productivity tool. Markets already trade "agents are coming" and "alignment is hard."

What's new — a named, falsifiable swarm scale (10k agents / 130B tokens / 88h) on a Millennium problem, plus a productized Ultra Mode with published multi-agent scaling plots, plus an internal Codex dollar intensity ($7–8k/day top 1%) that is a direct inference-demand datapoint. Brown's RSI haircut (significant ≠ 100×; non-intelligence bottlenecks) is a useful counter to FOOM framing from someone who ships the systems. Explicit refusal to claim measured 10k-vs-2k coordination gains is intellectual honesty rare in launch weeks.

Bear case — Millennium claim is OpenAI-announced and Dwarkesh-relayed; single-agent baseline missing; "we think it helped" is soft. If math spikes do not transfer to messy enterprise workflows, swarm demos stay research theater. Alignment "number one priority" is cheap talk until evals catch metric-gaming. $7–8k/day Codex could be subsidized internal pricing, not a commercial ARPU path.

Discount — Brown is OpenAI research talking after a capability announce; Dwarkesh is an RSI-sympathetic interviewer pushing sooner timelines. Treat HF narrative details Dwarkesh asserts (training→eval→infra control sequence) as interviewer framing — Brown repeatedly redirects to misalignment + security without confirming the full public chronology. Astra/robotics plug-and-play claims are Dwarkesh's, not Brown's.

Positioning

AI capex durability — STRENGTHENS. 130B tokens for one research solve and $7–8k/day top-1% Codex spend are pure inference/compute demand signals; Ultra Mode and swarm scale imply more parallel token burn per "task," not less. RSI-as-speedup (even Brown's haircut) raises internal demand for more compute sooner.

Enterprise agent stall — STRENGTHENS (mechanism). Capability side is racing (math swarms, fork/merge, Ultra Mode); Brown's own root-cause read on HF is misalignment + safeguards — evaluation/liability/trust remain the bind, not missing IQ. Fits the theme: pilots stall because risk is unsolved while demos clear Millennium-class problems.

Inference margin inversion — NEUTRAL / watched. Internal Codex intensity is a volume signal, not a disclosed gross-margin path across price cuts. Parallel agents raise tokens-per-job; whether cost/token falls faster than list is unaddressed.

HBM supply binds — NEUTRAL. Episode is algorithms/alignment, not memory hierarchy — though swarm token intensity sits upstream of accelerator and HBM pull.