ResearchPapers, evidence & method
Nathan Lambert argues AI will make labs much faster at engineering without putting them on a path to general superintelligence
Interconnects author Nathan Lambert expects AI to speed up lab engineering and parts of pretraining research. In his view, that lets labs try more model variants without moving them toward general superintelligence, while the cost of intelligence keeps falling sharply.
StrongHow it’s measured
YouTube video · 1.6M views · 35.5K likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
AT Pulse research deskPublished · 5 min read · AI-assisted analysisLambert expects models to become superhuman at distributed GPU engineering within a few years. He also thinks research on pretraining architecture and data selection could be automated in two to three years. He calls the result parallelized, AI-assisted language modeling rather than recursive self-improvement: labs can test more variants, but the models stay fundamentally the same. [1]
The post builds on Lambert's earlier idea of 'lossy self-improvement.' He sees three limits. Automated work targets narrow single metrics. Parallel agents run into Amdahl's law and run out of human-written tasks. Compute is still handed out through internal politics. Training and inference throughput suit agents because they have measurable targets. [1] [6]
Lambert's main practical prediction is that the cost of intelligence will fall almost exponentially and that cheaper agents will raise total demand. Epoch AI's price data shows the cost of reaching a fixed performance level fell between 9x and 900x per year, depending on the benchmark. Epoch warns that the steepest drops came in the most recent year and may not last. [1] [8]
Lambert argues that advantage is moving toward how agents are packaged and delivered, not frontier benchmark scores. He cites Meta's Muse, a personal agent that runs in a dedicated virtual machine and has since expanded into small-business integrations. Whether users trust it, given Meta's privacy record, is still an open question. [1] [5] [9]
The evidence: Epoch's data shows steep, recent drops in the price of fixed AI capability. Lambert's inference is that cheaper engineering, not a fast takeoff, will drive change.
Read the full assessment
If he is right, planning should assume cheaper agents, more demand and competition on delivery.
Executive brief
Nathan Lambert, a post-training researcher who writes Interconnects, argues that AI will make labs dramatically faster at engineering without putting them on a path to general superintelligence. He expects models to become "superhuman distributed GPU engineers" within a few years. He also thinks pretraining research on architecture and data selection could be automated in 2–3 years. In his view, though, faster engineering lets labs try more model variants without changing what the models fundamentally are (Interconnects). His main practical prediction is that the cost of intelligence will fall near-exponentially, possibly faster than recent trends, and that cheaper agents will raise total demand for them. The post is commentary, and its second half is paywalled.
What changed and event timeline
Labs bet on RL environments
Surge (which reported $1.2B in revenue for the prior year), Mercor, Mechanize and Prime Intellect all pushed into building RL environments. Ross Taylor and Andrej Karpathy questioned whether the approach scales ().
Sutskever declares an "age of research."
On Dwarkesh Patel's podcast, Sutskever argued that scaling has run its course as the main recipe and that new ideas are now the bottleneck (;).
Meta launches Muse
Muse is a personal agent built on Meta's Muse Spark model. It is free, with paid tiers at $20 and $100 a month ().
Lambert proposes "lossy self-improvement."
AI Weekly's summary of the post lists three limits on recursive self-improvement: automated work targets narrow single metrics, parallel agents hit Amdahl's law and run out of human-written tasks, and compute is still allocated through internal politics ().
AIDE² paper on self-improving research agents
The authors report seven self-discovered improvements over an 8-day run ().
Lambert's post
He calls the coming change "parallelized, AI-assisted language modeling" rather than recursive self-improvement ().
Capabilities and access
- The post announces no new model. It is an argument about where progress is heading.
- Lambert names Meta's Muse as an early example of value coming from how agents are packaged and delivered. The US launch is limited to adults (Wikipedia)).
Read the full section
- The post announces no new model. It is an argument about where progress is heading.
- Lambert names Meta's Muse as an early example of value coming from how agents are packaged and delivered. Muse runs on Muse Spark inside a dedicated virtual machine and is available on web, iOS, Android and WhatsApp (TechCrunch). The US launch is limited to adults (Wikipedia)).
Technical analysis for researchers and developers
- Lambert argues that much of the training and inference stack can be optimized by agents because it has measurable targets: training tokens per second per GPU, FLOPs per token, cost per answer (Interconnects).
- Near-term gains come from inference-time compute and GPU flexibility. Co-designing chips and models comes later.
- Sutskever attributes the gap between benchmark scores and real-world performance to RL environments built with evals in mind (video) and to models that "generalize dramatically worse than people" (video).
Read the full section
- Verifiable targets. Lambert argues that much of the training and inference stack can be optimized by agents because it has measurable targets: training tokens per second per GPU, FLOPs per token, cost per answer (Interconnects). This matches his earlier point that automated research works best on single-metric tasks (AI Weekly).
- Sequencing. Near-term gains come from inference-time compute and GPU flexibility. Co-designing chips and models comes later.
- The interview's diagnosis. Sutskever attributes the gap between benchmark scores and real-world performance to RL environments built with evals in mind (video) and to models that "generalize dramatically worse than people" (video). He expects value functions to make RL more efficient (video).
Claims and evidence
- Cost of intelligence will fall near-exponentially. Epoch cautions that the fastest drops happened in the most recent year and may not last (Epoch AI).
- Labs already save 10–30% on serving costs after launch. This is Lambert's own figure, with no source given.
- Self-improving research agents are feasible. These results come from the authors and have not been independently reproduced.
Read the full section
- Cost of intelligence will fall near-exponentially. Epoch AI's independent analysis found that the price of reaching a fixed performance level fell 9x to 900x per year depending on the benchmark. Epoch cautions that the fastest drops happened in the most recent year and may not last (Epoch AI).
- Labs already save 10–30% on serving costs after launch. This is Lambert's own figure, with no source given.
- Self-improving research agents are feasible. AIDE² reports self-improved agents that matched or beat human-engineered baselines on four held-out benchmarks, and reward hacking fell from 55% to 32% (arXiv). These results come from the authors and have not been independently reproduced.
- RL data firms have passed $100M and $1B in revenue. Surge reported $1.2B for the prior year (TechCrunch). Lambert's claim that the output is often low-quality rests on what researchers have told him, not on any published study.
Context and prior work
Read the full section
- Lambert explicitly builds on Sutskever's "age of research" framing. Sutskever says scaling "sucked out all the air in the room," leaving "more companies than ideas" (video).
- Sutskever also expects current methods to "go some distance and then peter out" (video). He forecasts 5–20 years to a system that learns as well as a human (video).
- The new post continues Lambert's September argument for "lossy self-improvement" (AI Weekly).
Limitations, safety and contested findings
- Conflicting evidence on automated research. AIDE² shows agents improving their own code end to end (arXiv), which pushes against Lambert's skepticism.
- Diffusion is not guaranteed. Muse faces trust questions tied to Meta's privacy record, including a $5B FTC penalty in 2019.
- RL environments are hard to scale and prone to reward hacking (TechCrunch).
Read the full section
- Conflicting evidence on automated research. AIDE² shows agents improving their own code end to end (arXiv), which pushes against Lambert's skepticism. Its authors say the run was short and the results may not generalize beyond the tested domains.
- Diffusion is not guaranteed. Muse faces trust questions tied to Meta's privacy record, including a $5B FTC penalty in 2019. Meta says Muse will not access passwords or share data with ad systems, and TechCrunch notes security experts have yet to check that claim (TechCrunch).
- RL environments are hard to scale and prone to reward hacking (TechCrunch).
- Partial argument. The paywalled section of Lambert's post was not reviewed.
Business and practitioner implications
- Plan for falling unit costs.
- Lambert's thesis puts the advantage in how agents are oriented and delivered, not in frontier benchmark scores.
- Audit purchased RL data. Labs see a return on buying it, but quality varies widely, so check what you get.
Read the full section
- Plan for falling unit costs. If inference prices keep falling, as Epoch's data shows they have, products that are too expensive to run today may become viable soon (Epoch AI).
- Differentiate on delivery. Lambert's thesis puts the advantage in how agents are oriented and delivered, not in frontier benchmark scores. Muse's move into small-business integrations follows that path (CNBC).
- Audit purchased RL data. Labs see a return on buying it, but quality varies widely, so check what you get.
- Keep humans for direction-setting. Agents handle single-metric engineering well. Choosing research directions still needs people.
The source trail.
Sources (9)
I expect rapid progress but not towards general superintelligence
Article text retrieved; extracted text may omit tables or interactive elements. A transcript of the video linked from this page is supplied (linked video manual captions; https://www.youtube.com/watch?v=aR20FWCCjAs); automatic text may contain errors.
interconnects.aiLinked video: I expect rapid progress but not towards general superintelligence
linked video manual captions
www.youtube.com