The Harness

Karpathy's funnel of AI understanding

Show Notes

Andrej Karpathy sketches a funnel of understanding, from billions who barely use LLMs to roughly 5,000 people with access to frontier-grade systems inside labs. Sakana AI and UC Berkeley's MASS self-improvement loop lifts research-agent scores but slips on coding benchmarks, and every figure is self-reported. Voice AI executives at HumanX say the bottleneck is reasoning latency and keyword accuracy rather than voice quality.

What is The Harness?

A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.

Good morning, it's Monday, October twelfth.

In today's briefing we have Andrej Karpathy's funnel of who actually sees frontier AI, a self-improvement loop from Sakana AI and UC Berkeley that gains on research tasks but slips on coding benchmarks, and voice AI executives naming reasoning latency as the bottleneck.

In other news…

There's a continuing question about whether human comprehension can keep pace with AI capability. Andrej Karpathy has re-posted an earlier thread that puts rough numbers on it. His napkin math runs as a funnel. About six billion people have barely used large language models. One to two billion are casual free-tier users. About twenty million use frontier models professionally for code and math. And about five thousand have frontier-grade access inside the labs. He ties the funnel to three things: problem size, capability that is jagged and peaks in digital, verifiable domains, and access tier. His conclusion is that human review and comprehension are starting to fall behind. One caveat, stated plainly: this is opinion, not a measured study. It adds a sense of scale to the comprehension thread without testing it. Impressions formed on free-tier use will underrate what professional users already get, so test features against the paid and agentic tier before deciding a capability isn't there. The next data point is an amateur astronomer's TESS candidate, with falsifiable dimming times between October thirty-first and November twenty-sixth. It will show whether an AI-assisted finding survives checking.

Sakana AI and UC Berkeley have published MASS, a loop where one model acts as solver, workflow optimizer and judge. It then fine-tunes on the best-judged multi-agent runs and repeats. Starting from Qwen 3.6 27B, held-out win rate rose from about fifty-four percent to about seventy percent over two rounds. Score per output token improved by twenty to sixty percent on research-style suites, but fell about four percent on Terminal-Bench 2.0 and about six percent on SWE-bench Verified. The suite is twelve synthetic tasks with only three held out, the authors report sharp regressions after strong iterations, and every number is self-reported. Because the same model produces the score it optimizes, headline benchmark numbers are harder to trust, even with agreement against external judges rising from seventy-three to ninety-three percent. Nothing here justifies swapping a fine-tuned small model into a coding agent. A replication with error bars and an untouched held-out set would change that.

At the HumanX conference, PolyAI's chief technology officer Shawn Wen said full-duplex voice models exist, but reasoning is not yet fast enough for natural conversation, and that speech recognition missing keywords undermines context. Otter's Alex Gay added that transcription errors compound downstream, and said Otter will notify all participants in chat when a meeting is recorded. Both said voice tools should disclose recording or AI. This is conference commentary with no new numbers. Spend on reasoning speed and keyword accuracy before voice polish, and build disclosure into the default flow.

That's the briefing. Have a great day, and don't forget to subscribe.