The AI Corner

The AI Corner

The Post-AGI Kit: DeepMind’s 4 Roads to Superintelligence, a 10x Compute Model and 12 Claude Prompts

A cost-curve model, 16 signals and 4 scenario plays built on DeepMind’s new report. Find the month each of your workflows flips to AI.

Ruben Dominguez's avatar
Ruben Dominguez
Nov 22, 2026
∙ Paid

Google DeepMind just published the most concrete map of life after AGI I have read this year, and the line that stopped me is one sentence about compute.

Cover of The AI Corner with the headline DeepMind mapped the road past AGI, a large orange 10x labelled more compute every year, a Google DeepMind logo and a grayscale photo of Google DeepMind CEO Demis Hassabis.

On November 19, research scientist Tim Genewein posted What comes after AGI? on the DeepMind Institute, the platform Shane Legg opened in September to take the AGI debate public. It condenses From AGI to ASI, a report Genewein wrote with 13 co-authors, Legg, Allan Dafoe and Marcus Hutter among them. Their argument is that life after AGI has moved from philosophy to measurement, and the measurements point to fast progress through the next decade.

Here is the sentence. DeepMind multiplies three drivers, hardware at about 1.5x a year, algorithms at 3x to 6x and investment at 2.5x, and gets global effective compute growing at least 10x a year. Run that for a decade and you have 10 billion times today’s stock.

Bar chart on a log scale showing DeepMind’s three drivers of effective compute, hardware at 1.5x a year, algorithms at 3x to 6x and investment at 2.5x, multiplying to 11.3x to 22.5x a year against DeepMind’s 10x floor, with a side box showing 100x by 2028, 10,000x by 2030, 10 billion x by 2036 and $725B of Big Tech AI capex planned for 2026.
Effective compute grows 11x to 22x a year once you multiply DeepMind’s three drivers. Source: DeepMind Institute, Nov 2026.

I built the premium kit at the bottom of this post around that math, because the numbers are specific enough to plan with: a 6-tab Excel model, a 16-signal tracker and 12 Claude prompts that show the month each of your workflows becomes AI-ready. The free part walks through the report first.


Cognitive industrialization

DeepMind’s name for this wave is cognitive industrialization. Mechanical industrialization turned routine physical labor into a commodity, and the report expects the same for routine thinking, even if frontier research stalls.

The cost side moves just as fast. Inference efficiency alone cuts compute demand about 4x a year, so a workload costing $1 billion of compute today would cost about $1,000 in ten years. I checked the math: $1 billion divided by 4 to the tenth power is $954.

Epoch AI measured something steeper in September. The price of a fixed level of AI performance has fallen about 13x a year since 2023, and GPT-5.6 Luna matched o3’s 75% score on GPQA Diamond for 725 times less per question, less than 18 months after o3 shipped.

Log-scale line chart of a $1 billion AI workload falling to $954 by 2036 at DeepMind’s 4x a year, $190 at Epoch’s 4.7x near-frontier rate and $0.007 at Epoch’s 13x fixed-performance rate, with side notes on halving times of 6.0 and 3.2 months, a 725x cheaper GPQA result and AI’s 13x yearly price decline against 1.5x for computing hardware.
A $1 billion workload falls to about $954 by 2036 at 4x a year. Source: DeepMind Institute; Epoch AI, 2026.

The cost curve will reshape your company before any AGI announcement does. Sam Altman put the drop at 100x in two years, and the 9 bottlenecks deciding who wins in AI show where cost still bites.


How long an agent can work alone

Cheaper thinking matters once agents can hold a task together for hours, and METR keeps that scoreboard. Its time horizon is the length of task, in skilled-human hours, that an agent finishes half the time.

Claude 3.7 Sonnet sat at about one hour in February 2025. GPT-5 reached 2 hours 17 minutes that August, and an early version of Claude Mythos Preview tested at about 17.4 hours in March 2026. METR’s Time Horizon 1.1 update puts the doubling time at 88.6 days since 2024, against roughly seven months across 2019 to 2025.

Log-scale chart of METR 50% time horizons rising from about 1 hour for Claude 3.7 Sonnet in February 2025 to 2 hours 17 minutes for GPT-5, 4.9 hours for Claude Opus 4.5, 12 hours for Claude Opus 4.6 and 17.4 hours for an early Claude Mythos Preview in March 2026, with a dashed 88.6-day doubling trend and a shaded zone above 16 hours that METR calls unreliable.
Agents went from 1-hour to 17-hour software tasks in 13 months. Source: METR, 2025 and 2026.

That 17.4-hour reading is the number I trust least. METR says results above about 16 hours are unreliable with its current task suite, and the 95% interval on Mythos runs from 8.5 to 55 hours. At 80% success the best reading is about 3.1 hours, and that is the line I would plan around.


Four roads past AGI

The report’s core is a map. DeepMind defines superintelligence as systems that consistently beat large groups of human experts on work that takes them years, using fewer resources, and it sees four ways to get there from AGI:

  1. Scaling models and data. Bigger models trained on more data, until data supply, power, grids and fabs push back.

  2. A paradigm shift. New methods for what transformers still lack: continual learning, world models, sample efficiency and inventing new concepts.

  3. Agent collectives. Millions of AGI instances coordinated as group agents or whole agent economies, a scale up from 1,000 agents working while you sleep.

  4. Recursive self-improvement. AI improving the code, data and specialization behind the next AI, fastest where results get checked automatically, like math proofs. Most self-improving loops fail for a specific reason.

Diagram of four roads from expert-level AGI to superintelligence in DeepMind’s report: scaling models and data, a new paradigm, agent collectives and recursive self-improvement, each with its bottleneck and the public signal to watch.
Four roads from AGI to superintelligence, with the signal for each. Source: Genewein, Legg et al., 2026.

The planning lesson sits near the end of the essay. The roads compound, and when one stalls, economic and geopolitical competition shifts effort to the others. Cap a single input like training compute and you build an overhang that releases in sudden jumps once the cap lifts. DeepMind’s advice for everyone else is “model, measure, forecast, and repeat.”


Who says when

The report avoids firm dates, but Shane Legg, one of its authors, is less guarded in public. He still holds the forecast he made two decades ago, even odds of minimal AGI by 2028, and told the University of Auckland this month that “we’re getting very close.” At Davos in January, Demis Hassabis said he is still on his 50% by the end of the decade, and Dario Amodei, on the same stage, said AI better than humans at everything might arrive in one to two years.

Timeline of AGI forecasts with photos: Shane Legg at 50% odds of minimal AGI by 2028 and full AGI three to six years later, Demis Hassabis at 50% odds by about 2030 plus or minus a year, and Dario Amodei at one to two years from January 2026, beside DeepMind’s five levels of AGI from Emerging to Superhuman.
Lab leaders place human-level AI between 2027 and 2030. Photos: Libor.burian and Christopher Michel, CC BY-SA 4.0.

Legg walks through minimal AGI, full AGI and superintelligence with Hannah Fry on the Google DeepMind podcast:

Hassabis, speaking at Stanford this spring: “Ten years from now, I think we’ll realize that we were standing in the foothills of the singularity now.”

For the longer read on his bet, see Hassabis Bet His Job on AGI by 2030 and Dario Amodei on “the zeroth world”.

Two moves for this quarter

  1. Re-price every AI line in your budget each quarter. At 4x a year, cost per task halves every six months; at Epoch’s 13x, every 3.2 months.

  2. Put agents on the work under three hours first. That is where 80% reliability sits today, and full-day tasks come next.


Get the Post-AGI Kit

Below the paywall: the Post-AGI Kit, a 6-tab Excel model that dates when each of your workflows becomes AI-ready, plus the 16-signal tracker, the 4-road playbook for founders, investors and operators, and 12 Claude prompts that run it on your company.

Get the kit

Keep reading with a 7-day free trial

Subscribe to The AI Corner to keep reading this post and get 7 days of free access to the full post archives.

Already a paid subscriber? Sign in
© 2026 The AI Corner · Privacy ∙ Terms ∙ Collection notice
Start your SubstackGet the app
Substack is the home for writing, podcasting, and media founders.