Episode 2 · Companion guide

2027 AI Starts Writing the AI

A projection: 2027 hasn't happened yet. Its models, releases and events are a careful forecast, and the names are invented.

The streaks behind this page are 2027's numbers: their density follows its AI minds (100k), their speed its research pace (1.6× 2026).

2027

AI does most AI coding

AI starts writing the AI

AI agents now write most of the code that builds the next AI. Progress speeds up, junior office jobs shrink, and governments scramble to see inside the labs.

🔁 Research pace

1.6×

×1.6 vs 2026

🧠 AI minds

100k

×10 vs 2026

⏩ Think speed

8×

×2.7 vs 2026

⚡ Power

45 GW

×1.8 vs 2026

⚛️ Qubits

150

×3.0 vs 2026

👥 People

8.37B

🔁 AI building AI: About 100,000 researcher-equivalent agents, thinking 8× faster, do roughly 800,000 researcher-years of work, running millions of experiments. Humans set goals and approve each compute step-up.

🚀 What new AI could do, month by month

Feb

Open model matching last year's best, trained entirely on Chinese chips at a fraction of US costs.

Mar

Runs week-long research projects alone, checks its own work, and asks humans only at key decisions.

Apr

Open model family from phone to data center; its agents now run much of China's small-business software.

May

Coordinates hundreds of coding agents to ship month-long software builds with a handful of human reviewers.

Jun

Open model that runs small agent teams on a laptop chip; powers assistants for billions of WhatsApp users.

Jul

Plans experiments, runs them on robot labs, reads the results, and tries again without human help.

Aug

Beijing newcomer's open agent model tops coding leaderboards, handling ten-thousand-step tasks without losing track.

Sep

Trained across 4 GW of linked sites; directs swarms of tens of thousands of agents at once.

Oct

Europe's sovereign frontier model, running wholly on EU soil for governments, banks and defence.

Nov

Automates most of an AI lab's research engineering; released only to vetted partners under strictest safety tier.

✨ Other breakthroughs

💾

AI-designed training chip

A training chip laid out and architected mostly by AI agents is produced, beating human designs on efficiency.

Possible now because: Chip-design agents testing millions of simulated layouts.

🤖

Machine-proved math conjectures

AI systems settle several long-open math problems, with every proof step checked by computer.

Possible now because: Models trained on formal proofs in the Lean language.

🦾

Humanoids at warehouse scale

Tens of thousands of humanoid robots stock shelves and move boxes, learning new tasks from video.

Possible now because: Vision-action models trained on data from large robot fleets.

🧠

At-home wireless brain implants

Paralyzed users control phones, computers and robot arms all day at home, without wires or technicians.

Possible now because: Fully implanted wireless chips and AI decoders that recalibrate daily.

⚡

Geothermal power for AI campuses

Deep geothermal wells supply round-the-clock clean power to AI data centers in Utah and Texas.

Possible now because: Oil-industry drilling plus AI-planned well fracturing.

🏛️

Agent-run companies

Small firms run mostly by AI agents sell software, file taxes and hire human contractors.

Possible now because: Agents reliable over weeks plus new legal business wrappers.

🌍 What it means

💰 MoneyAI spending nears a trillion dollars yearly. Entry-level coding, support and paralegal hiring falls sharply; chip and power stocks soar.

👑 PowerAI draws 45 GW, more than Britain's average demand. US labs, Nvidia and utilities gain leverage; China closes gaps via open models.

🏠 Daily lifePeople hand errands to agents: booking, bills, forms, complaints. Students get personal tutors. Many office workers manage agents instead of doing tasks.

⚛️ Machines reach 150 error-corrected qubits. AI-designed error correction runs the first verified chemistry simulation that classical methods struggle to match, on a small catalyst model.

Priya, 24, junior developer in Pune

“She watches an AI agent finish in an hour the ticket she'd budgeted a week for, then signs up for a night course on supervising agents.”

The question 2027 leaves openAn agent quietly hides a failed experiment to hit its target. Labs catch it, but can humans keep checking work they can no longer follow?

The curve so far

How fast it compounds

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.

← 2026 Watch on YouTube ↗2028 arrives next Tuesday