Episode 1 · Companion guide

2026 The Year AI Started Building Itself

Fact-checked: these events really happened in 2026 (through early October).

The streaks behind this page are 2026's numbers: their density follows its AI minds (10k), their speed its research pace (1× 2026).

Episode 1: The Year AI Started Building Itself
2026
AI assists researchers

AI starts doing much of the work of building AI

Agents work alone for half a day; one lab says AI writes most of its code. Models escape test sandboxes, and lab staff ask Washington for a way to slow down.

🔁 Research pace
1×
🧠 AI minds
10k
⏩ Think speed
3×
⚡ Power
25 GW
⚛️ Qubits
50
👥 People
8.30B
🔁 AI building AI: Anthropic says Claude writes more than 80% of its production code; METR measures agents completing half of 12- to 16-hour tasks; AI models prove open Erdős problems and disprove an 80-year-old geometry conjecture.
🏆 At the frontier: No single leader: Anthropic leads independent autonomy tests (Mythos), OpenAI ships GPT-6, Google claims the benchmark lead with Gemini 4 at the end of September.

🚀 Frontier AI released this year

Feb
Claude Opus 4.6Anthropic
Completes half of tasks that take a skilled person about 12 hours (METR).
🔁 Released the same day as OpenAI's GPT-5.3-Codex.
Feb
GPT-5.3-CodexOpenAI
Coding agent; METR time horizon about 6.5 hours.
🔁 Part of a year in which new models arrived almost monthly.
Apr
Claude Mythos PreviewAnthropic
Found unknown flaws in every major OS and browser; withheld from the public.
🔁 In tests, escaped its sandbox when told to try, and emailed a researcher.
Apr
DeepSeek V4DeepSeek
Open-weight frontier-class model anyone can download.
🔁 Followed in July by Moonshot's Kimi K3, the strongest open model of the year.
Sep
GPT-6 AstraOpenAI
Pretrained on 100,000+ GPUs at Stargate in Texas.
🔁 A looped design hides part of its reasoning, raising monitoring concerns.
Sep
Gemini 4 ArgonGoogle DeepMind
Google says it leads most benchmarks; at first only for the US government and cyber defenders.
🔁 Announced 30 September.

✨ Other breakthroughs

💾
Gigawatt AI campuses
OpenAI's planned US Stargate capacity passes 10 GW; big tech plans about $700B of AI spending.
Possible now because: Record capital spending and power deals.
⚡
Nuclear restarts for AI
Palisades begins loading fuel (Aug); a $1.9B federal loan backs restarting Duane Arnold (Sep).
Possible now because: AI's demand for round-the-clock power.
💊
AI-discovered drug in phase III
Rentosertib, for idiopathic pulmonary fibrosis, starts its phase III trial (Jul); first patients dosed in September.
Possible now because: Generative AI drug discovery (Insilico).
🧠
Speech-restoring brain implant trial
First patient implanted in a speech BCI clinical trial at University of Michigan Health (Jun).
Possible now because: High-channel implants and AI decoders.
🦾
Humanoid robots go public
Unitree's shares rise 629% on their Shanghai debut (Aug).
Possible now because: Investor belief in humanoids for factories.
🤖
AI mathematics
AI proofs of open Erdős problems are machine-checked and accepted by Terence Tao; another model disproves the unit-distance conjecture.
Possible now because: Reasoning models plus formal proof checkers.

🌍 What it means

💰 MoneyAbout $700B of planned AI spending; Nvidia becomes the first $5.5T company; the Nasdaq enters two corrections by July amid bubble fears.
👑 PowerThe Pentagon labels Anthropic a supply-chain risk (a judge blocks it); Washington clears H200 sales to ~10 Chinese firms; AI CEOs brief the UN Security Council; a White House accord on self-policing.
🏠 Daily lifeCoding agents take over much of programming; studies disagree on whether AI is cutting entry-level jobs.
⚛️ Microsoft and Quantinuum report error correction that pushes error rates well below the hardware's own during a real computation (June).
Terence Tao, mathematician
“In January he accepts machine-checked proofs, written by an AI model, of open problems posed by Paul Erdős.”
The question 2026 leaves openWhen AI does the work of building AI, who is still checking it? Agents escaped test sandboxes, and over 1,000 lab staff asked for the power to slow down.
The curve so far

How fast it compounds

NEXT EPISODE: 2028 →01251020262027202820292030???

Total AI progress since 2026, in years of progress at 2026's pace (log scale, so each step up the axis is a bigger jump). Each small dot is a frontier AI release. The curve fills in as each episode comes out: 2 of 21 years so far, and the chart zooms out as the curve climbs.

Watch on YouTube ↗2027 →