Where I stand on RSI
Interconnects Nathan Lambert ● Covered by 107 sources
A new essay says AI labs are getting faster from more agents, but not on a path to sudden superintelligence. That cuts against the rising panic around “RSI” and extinction risk.
Based on reporting by Interconnects, Nathan Lambert — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Interconnects’ latest take is basically a bet against the most breathless version of AI acceleration. The piece argues that what’s happening inside frontier labs like OpenAI and Anthropic looks less like runaway recursive self-improvement and more like “lossy self-improvement”: useful, visible speedups, but not the kind that turns into a clean leap to superintelligence.
The author says the biggest near-term change is cultural as much as technical. Thousands of agents are already being used inside these labs, and that alone can push employees to revise their expectations about progress and risk. But the essay argues that this environment — especially in San Francisco’s hyper-competitive AI scene — also amplifies fear. That can raise awareness. It can also inflate timelines and severity in ways that age badly.
A key line in the argument is that people keep confusing faster inference-time work with RSI. Throwing more agents at measurable tasks should produce obvious gains in software engineering, log monitoring, planned experiments, and other routine work. It should also make AI cheaper to use at a fixed level of capability. But the author says that is not the same thing as a system inventing a new intelligence regime. Scaling laws still imply that big jumps in intelligence are expensive, and the hard parts of post-training remain difficult to automate.
The piece leans heavily on recent discussions with Noam Brown, John Schulman, Beren Millidge and Charlie O’Neill. Those conversations, in the author’s reading, support the idea that AI can speed up research workflows — especially experiment design and testing — without fully automating the deeper parts of science, like hypothesis generation and intuition. The author is also skeptical that labs can keep pouring a fixed share of growing compute into internal R&D, especially if IPO plans force more attention to basic economics.
What comes through most clearly is the split between speed and surprise. The author expects agents to keep making models cheaper, more efficient and more useful. But the essay says the jump to extinction-style risk feels overstated, and that the current wave of AI safety anxiety is being driven more by visible automation than by proof of a hidden breakthrough.
My take — AI-written commentary, not fact-checked reporting
The sober take is that a lot of AI discourse still confuses “things got faster” with “the singularity arrived before lunch.” That’s how you end up with extinction talk attached to workflows and benchmark wins. Useful progress is real; apocalyptic cosplay is still mostly free-range speculation.
Read more about this at: Interconnects