Two years of watching AI systems get stronger have not brought the wider world to agreement about where the technology is going, Nathan Labenz said. At The Curve, though, he found that one point had largely settled: "it does look like the AIs are going to get really powerful," and that is something people "can't just sort of hope goes away." Even the attendees with the most modest expectations, he said, agreed that AI is "going to do an awful lot." The open questions had moved on to recursive self-improvement, meaning AI systems helping to build better AI. Labenz said people were asking whether it would "foom" (take off explosively), run fast for a while and then level off, or run away from us.
Labenz, host of The Cognitive Revolution, recounted the conference in an October 8, 2026 weekly highlights episode, narrated in his cloned voice. Questions in the episode came from Prakash. The Curve, run by the Golden Gate Institute, is an invitation-only gathering of researchers, policymakers and executives with differing views on AI acceleration and safety. This year it met in Berkeley from October 2 to 4, with a track held under the Chatham House rule, which allows ideas to be reported but not attributed. What follows from the frontier-lab people is Labenz's retelling of what he heard from them.
Why insiders come to talk
Labenz said the biggest divide he saw was between people inside the frontier companies and everyone else. "Everybody knows that they're living in the future and they have visibility that the rest of us don't have," he said. He heard from founders, executives and top researchers at those companies, sometimes in sessions and sometimes in passing conversations, all under Chatham House rules.
Before moderating sessions, he asked several of them what made the hour worth their time. He said they repeatedly told him they wanted to share as much as they could with the broader community. That did not include trade secrets, but it did include a candid account of what they were seeing internally, what they believed and how constraints on the industry might come about.
Pre-training still pays, partly through synthetic data
One senior executive at a frontier lab said that pre-training, the first and largest training stage in which a model learns from huge amounts of text, "continues to deliver," Labenz recounted. Base models keep getting stronger, with no obvious stopping point. Scaling laws, the observed relationship between more compute and data and better models, are holding "or maybe even bending" in a favorable direction because data quality is improving.
Much of that improvement comes from synthetic data. Labenz described the recipe: give a model existing data and ask what it could be turned into that would make a model smarter if trained on it. The model spends reasoning effort rewriting, cleaning and viewing the material through "different lenses," and the output becomes pre-training data for the next generation. In effect, the tokens a model spends thinking are converted back into training material. Labenz called this "one way in which recursive self-improvement is happening."
Research taste as an "elicitation problem"
The same executive went further. Labenz said the executive believes the latest frontier models already have "the ineffable research taste" needed to make paradigm-level breakthroughs. The barrier, in this view, is elicitation: reinforcement learning, the training stage that rewards a model for successful behavior, does not yet bring that taste out reliably. So it shows up only rarely.
Labenz offered his own illustration of what that implies. You might need 10,000 agents to crack a Millennium Prize problem, where answers can be checked and reinforcement learning works better. You might need "a million agents" to stumble onto a paradigm change in machine-learning research. The 10,000 figure echoes OpenAI's September 8 announcement of a reported proof related to the Navier–Stokes Millennium Prize problem. The company says its Navier–Stokes effort involved roughly 10,000 concurrent agents over 88 hours, followed by machine-checked formalization in the Lean proof language. OpenAI says it does not intend to claim the prize.
A speedrun record as a test
Labenz connected the research-taste claim to a new record in the NanoGPT speedrun, a long-running contest to train a small language model to a set quality level in as little wall-clock time as possible. The record came from a company called Hyperstition. Labenz described the drop as going from roughly 70 seconds to almost half that, which he said was "more than, I think, the last 45 improvements combined." That comparison is his own estimate.
Hyperstition's report by Deven Pietrzak supports the scale of the jump. Eighteen runs on eight H100 GPUs averaged 39.9 seconds, compared with 73.9 seconds for baseline runs on the same machine, a 46.0% reduction. All eighteen reached the loss target. The submission was merged on September 28 as record #92. Its review discussion examined the legality and practical relevance of an unusually large auxiliary hashed embedding table, an extra lookup table of word combinations.
What interested Labenz was who had the idea. He said the author reported that AI did not play a big part. In his words, the team was "obviously" getting lots of help, but the core insight was mostly attributed to the human. The executive's view, as Labenz relayed it, was that models could probably produce ideas of this quality but that drawing them out remains the hard part, and that the team did not rely heavily on AI here.
Labenz also noted that Hyperstition, which does large-scale training of its own, made a point in its publication that the optimizer behind a significant but not majority share of the time savings is not as good as the one it uses internally. Pietrzak's report separates the techniques it published from frontier-scalable methods the company withheld.
How the followers keep up
Labenz said the general takeaway was that two or three companies are truly pushing the frontier. With Gemini 4 being announced, it may again be a "three horse" race. The other two companies usually counted in the top five, Elon Musk's and Mark Zuckerberg's, "are actually distilling a lot, whether they know it or not." Distillation means training one model on the outputs of a stronger one so it picks up some of the stronger model's abilities.
According to the view Labenz relayed, the channel is not direct copying of model outputs through Anthropic's API. Instead, it is the "cottage industry" of companies that build and sell reinforcement-learning environments: simulated tasks with scoring rules that a model practices on. Those vendors have Claude create the environments and then sell them to other frontier developers. Those developers train on "an environment that only exists because Claude was smart enough to make it," and so keep capturing the leaders' advances. Labenz said that without such leaks, including the RL industry, the other companies would not really be keeping up.
Another founder- or executive-level person at one of these companies told him the followers have kept up, with a persistent gap but "a kind of consistent following distance." Labenz contrasted that with a prediction he recalled from an old Anthropic fundraising deck: that around 2026, the companies training the best models would be so far ahead that nobody could catch up. That has not happened. But the people he spoke with think it probably would have, if the leaders had not been releasing models and if others had no direct or indirect ways to distill them.
Timelines: "even sooner"
Prakash gave his read on Musk's strategy. In his view, Musk is betting that infrastructure matters more than models and that his build-out will let him catch up. Projects like TerraFab and the SpaceX satellites are structured for 2029, 2030, 2031 and into the 2030s. Anthropic and OpenAI, he said, are structuring for a two-to-three-year horizon, looking at AGI and recursive self-improvement around 2028 or 2029.
Labenz said the mood at The Curve was "even sooner." People talked about next year: the "critical period is beginning now" and "RSI is kind of in hand." There was no full agreement on how fast things would move or where they would top out. But people shared a sense that decisions made in the coming months, and how compute is used over the next year, could be "really critical." There "wasn't really even too much 2028 talk," he said.
He acknowledged that the crowd was "probably somewhat biased to be a short timelines crowd." He added that this crowd includes the executives, top researchers and founders of the frontier companies themselves.
Expecting Washington to move
The next morning, Labenz added a final takeaway: "there is going to be government intervention." The old assumption that Congress will never act was giving way, he said. The new expectation was that after the midterm elections, especially if Democrats win both chambers, Congress could move, possibly with Republicans joining them.
As Labenz described it, many Republicans want to do something but have been unwilling to go against the president, and they are "getting a ton of money from the data center interests." All of that, he said, "could come up for a renegotiation and realignment after the midterm." Congressional action might come "a lot sooner" than people who have grown used to stasis expect, and he said a policy response is "definitely part of what we need to be looking out for."