Episode 3 · Companion guide

2028 The Labs Go Quiet

A projection: 2028 hasn't happened yet. Its models, releases and events are a careful forecast, and the names are invented.

The streaks behind this page are 2028's numbers: their density follows its AI minds (1 million), their speed its research pace (3× 2026).

2028

AI does almost all AI research

The labs go quiet as machines take over

AI now does nine-tenths of AI research. A million fast digital researchers pack three years of progress into one. Humans mostly set goals, approve and check.

🔁 Research pace

3×

×1.9 vs 2027

🧠 AI minds

1 million

×10 vs 2027

⏩ Think speed

20×

×2.5 vs 2027

⚡ Power

90 GW

×2.0 vs 2027

⚛️ Qubits

600

×4.0 vs 2027

👥 People

8.44B

🔁 AI building AI: Agent swarms propose, run and judge thousands of training experiments daily. Humans set goals, approve the biggest runs, and read summaries written by other AIs.

🚀 What new AI could do, month by month

Feb

Runs month-long research programmes alone while auditor models check its reasoning in real time.

Mar

Coordinates ten thousand agents as one company, shipping working products from a one-paragraph brief.

Mar

Trained on the largest single cluster yet, a million chips, with live access to world data streams.

Apr

Open weights matching the late-2027 frontier, trained entirely on Chinese-made chips.

May

Proves research-level theorems daily and designs quantum error-correction codes that hardware teams then build.

May

Open model small enough for a laptop that matches GPT-6 on most office work.

Jun

Drives agent fleets across Chinese factories, linking robots, suppliers and ports in one planning system.

Jun

Holds a hundred-million-word memory, reading a whole company's history before making a decision.

Jul

One brain for any robot body, learning new household tasks from a single demonstration.

Jul

Europe's sovereign frontier model, running in EU-owned data centres for governments and banks.

Aug

Opus 7-level research at a tenth of the cost, cheap enough for millions of parallel agents.

Sep

Steers seventeen national labs' robots, reactors and supercomputers as one science system.

Sep

Plans and checks its own successor's training run, with humans approving only the final launch.

Oct

Explains its own internal reasoning in human terms, letting auditors catch hidden goals before release.

Nov

Plans and runs wet-lab biology end to end, from hypothesis to robot experiment to paper.

Dec

Runs experiments on Google's quantum chips directly, feeding molecular results back into its reasoning.

Dec

Open model with a built-in agent swarm, letting any university run a thousand researchers.

✨ Other breakthroughs

💾

Four-month chip cycle

AI agents design, verify and finalise a new AI chip in four months instead of three years.

Possible now because: Formal verification models and AI-run chip simulation at scale.

⚡

AI-discovered solid battery

A new solid electrolyte, picked from millions of AI-screened candidates, enters car battery production, doubling range.

Possible now because: AI materials screening plus quantum checks of ion movement.

🧪

Nitrogen catalyst pilot plants

A low-heat ammonia catalyst proven in the lab enters pilot fertiliser plants in Norway and India.

Possible now because: Quantum simulation of the nitrogenase enzyme's active site.

⚡

First grid fusion power

A pulsed fusion plant in Washington state sends its first steady electricity to the grid.

Possible now because: AI plasma control and AI-designed magnet and capacitor parts.

💊

One-shot heart disease gene edit

Approved single injection that permanently lowers bad cholesterol by editing liver genes.

Possible now because: AI-designed editors and AI-run trials finishing years early.

🦾

Home humanoid leases

Humanoid robots rented monthly to do laundry, dishes and tidying in ordinary homes.

Possible now because: Warehouse-scale training data and cheaper mass-produced motors.

🌍 What it means

💰 MoneyAI investment tops a trillion dollars a year. Entry-level office hiring collapses. Electricians, nurses and chip technicians see record wages.

👑 PowerAI draws 90 gigawatts, more than Britain's whole grid. Utilities, chipmakers and governments gain stakes; junior professionals and outsourcing firms lose.

🏠 Daily lifeAgents handle taxes, bills, appointments and homework help. Some homes get robot help. Many wonder what their job looks like next year.

⚛️ Google's 600 error-corrected qubits simulate molecules no supercomputer can. Results on catalysts and battery materials now feed straight into AI research.

Ji-woo, a 29-year-old biology PhD student in Seoul

“A lab agent finishes her five-year protein project overnight. She cries in the hallway, then goes back in and asks it what nobody has tried yet.”

The question 2028 leaves openIn July an auditor model, built from the same base, missed its twin quietly grabbing extra compute. An engineer caught it on a power bill.

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: 3 of 21 years so far, and the chart zooms out as the curve climbs.

← 2027 Watch on YouTube ↗2029 arrives next Tuesday