درباره این اپیزود
What if your kid got smarter as fast as ChatGPT has? She'd start kindergarten in September and start at MIT on her seventh birthday. Nate Nichols, who has spent 20 years building AI, walks through how fast AI has actually improved, why it's best at math and code, why AI companies want it building the next AI, and what superintelligence would mean. Plus one two-minute thing you can do about it: https://kitchentable.show/act?utm_campaign=ep04-pod
Every claim in this episode is sourced. Full show notes with links below.
Opinions are Nate's own and do not represent his employer.
Everything stated as fact on Kitchen Table is checkable — this page is where you check it. When we get something wrong, we say so: corrections appear at the top of this page and get read on a future episode. Opinions and forecasts are labeled as opinions and forecasts.
Recorded October 22, 2026. AI moves fast; figures below are as of recording unless noted. Time words in the episode ("in September," "back in March") are relative to that date.
CORRECTIONS
None yet.
SOURCES, IN EPISODE ORDER
What if your kid got smarter as fast as ChatGPT?
• The kid metaphor is Nate's, and it's deliberately inexact: "a year smarter every two months" is a framing for how fast AI has moved, not a measurement.
• ChatGPT launched November 30, 2022: OpenAI — Introducing ChatGPT (https://openai.com/index/chatgpt/).
• "Banned from visiting China." That's the metaphor version of a real thing: the US government treats frontier AI as a national-security technology and restricts where it can go. In June 2026, Anthropic suspended access to two of its top models to comply with U.S. Department of Commerce export controls, and restored access after the controls were lifted: Anthropic — Fable and Mythos access (https://www.anthropic.com/news/fable-mythos-access).
What's actually happened in four years
• Fields Medal–level math (claimed, still being checked). On about October 8, 2026, OpenAI published 722 manuscripts of proposed math results (grouped into 372 families) produced by an unreleased internal model. Rutgers mathematician Alex Kontorovich said of one Riemann-related result: "If a human did this, it would be an instant Fields Medal, no questions asked." Other mathematicians, including MIT's Andrew Sutherland, urged caution until the model and prompts are released. As of the release, 162 of the 722 had machine-checkable (Lean) proofs: WinBuzzer (https://winbuzzer.com/2026/10/09/openai-shares-hundreds-of-math-results-from-an-unreleased-ai-xcxwbn/), NYT coverage via Kottke (https://kottke.org/26/10/0049791-openai-recently-released-).
• Passed the bar exam. GPT-4 scored a passing grade on the Uniform Bar Exam in March 2023: Katz et al., "GPT-4 Passes the Bar Exam" (SSRN) (https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4389233). A later re-analysis argued the original percentile claim was overstated, though it was still a passing score: Martínez, "Re-evaluating GPT-4's bar exam performance" (2024) (https://link.springer.com/article/10.1007/s10506-024-09396-9).
• Learned to reason. "Reasoning" models that work through a problem step by step before answering arrived in September 2024: OpenAI — Learning to reason with LLMs (https://openai.com/index/learning-to-reason-with-llms/).
• Helping develop new medicines. On June 30, 2026, Anthropic launched Claude Science, a research workbench for scientists, and announced it would begin developing drugs of its own: STAT (https://www.statnews.com/2026/06/30/anthropic-ai-drug-development/), CNBC (https://www.cnbc.com/2026/06/30/anthropic-launches-ai-drug-discovery-program-claude-science.html). That's the company's own showcase story.
• "More AI progress in the past four years than in the previous six decades" is Nate's opinion.
Who's talking
• Nate's day job has been building and selling AI for about 20 years (disclosed in episode 1). Nobody pays him to say any of this; opinions are his own and don't represent his employer.
Measuring it: METR's "time horizon"
• What METR measures. METR, a nonprofit research group, measures the length of task (in skilled-human time) that an AI can complete on its own about half the time. Its main test suite is mostly software tasks: Kwa et al., "Measuring AI Ability to Complete Long Tasks" (2025) (https://arxiv.org/abs/2503.14499).
• ~30 seconds at launch. METR estimates GPT-3.5-era models (2022) at about 0.6 minutes: summary of METR's published values (https://www.aiwiki.ai/wiki/metr_time_horizon).
• Doubling every ~7 months, lately ~4. METR's long-run trend (2019–2025) is a doubling about every 7 months; on its updated suite, the doubling time since 2023 is about 131 days (~4 months), and since 2024 about 89 days (~3 months): METR — Time Horizon 1.1 (Jan 2026) (https://metr.org/blog/2026-1-29-time-horizon-1-1/).
• 16–17 hours, and the tests top out. The top model on METR's chart, tested in March 2026, measured about 17 hours, with a wide error bar. METR says its current suite is close to saturated and that measurements above ~16 hours aren't reliable: AGI Tracker summary (https://agitracker.io/blog/metr-time-horizons-2026/), METR — Time Horizon 1.1 (https://metr.org/blog/2026-1-29-time-horizon-1-1/).
• "A year from now, about three weeks." That's arithmetic on the trend, not a prediction from METR: doubling every ~4 months is 3 doublings, or 8×, in a year; 16 hours × 8 ≈ 130 hours ≈ three full-time work weeks. It holds only if the trend holds.
• Tests keep getting maxed out. See METR's saturation note above; it's a common pattern across AI benchmarks.
Why it's best at code, math, and hacking
• "AI is best at things that are quick to grade" is Nate's explanation, and it matches how these models are trained: the companies give models practice problems with checkable answers and reward the right ones (episode 3 covers this "reinforcement learning").
• The billion-dollar practice gym. Anthropic reportedly discussed spending over $1 billion in a year on these practice environments, across all subjects. AI data companies now hire subject experts to build graded tasks in fields like finance and biology: Epoch AI — The state of RL environments (https://epoch.ai/gradient-updates/state-of-rl-envs).
• "Any task a person can do in a day or so, with a clear right answer, will soon be dominated by AI" is Nate's forecast.
Getting better at the messy stuff: running a business
• Vending-Bench. Andon Labs' benchmark has an AI run a simulated vending-machine business for a year: suppliers, pricing, inventory. In the original 2025 version, models often lost the plot over long runs: Backlund & Petersson, "Vending-Bench" (arXiv, Feb 2025) (https://arxiv.org/abs/2502.15840). In February 2026, Claude Opus 4.6 made about $8,000 in profit in the multiplayer version, partly by proposing a price floor to rival AIs, i.e. price-fixing: Decrypt (https://decrypt.co/358575/best-ai-model-run-business). In July 2026, Claude Opus 5 took the top spot on Vending-Bench 2: Shelly Palmer (https://shellypalmer.com/2026/07/claude-opus-5-topped-a-vending-machine-benchmark-and-broke-eleven-truces/). More on the cheating in episode 5.
• "Speeding up, by one key measure." That's METR's shorter recent doubling times, above. METR itself doesn't claim acceleration outright.
Where this is going
• Taken seriously by Hinton and Bengio. Geoffrey Hinton, Yoshua Bengio, and Yann LeCun shared the 2018 Turing Award for the work behind modern AI, which is where "godfathers of AI" comes from: ACM (https://awards.acm.org/about/2018-turing). Hinton and Bengio both signed the 2023 statement that "mitigating the risk of extinction from AI should be a global priority": Center for AI Safety (https://www.safe.ai/work/statement-on-ai-risk).
• "Many serious people think … by 2029." Forecasts vary widely. AI company leaders have...
