My Eight-Month AGI Bet

Put a date on the wall: 29 June 2027.

Writing on 29 October 2026, I predict that within eight months we will see a system that meets the working definition of AGI I set out below. I also suspect that the leading laboratories have achieved more powerful internal self-improvement loops than they have publicly demonstrated.

That second sentence is speculation. I have no verified evidence that Google DeepMind, Anthropic, or OpenAI has secretly solved unrestricted recursive self-improvement. I am not reporting inside information. This essay is a public forecast and an argument for preparing, with conditions under which I will call the forecast wrong.

What I am actually predicting

For this bet, AGI means a broadly capable digital system that can complete unfamiliar, economically useful knowledge-work projects across several domains with limited supervision, learn new tools from documentation, detect and recover from many of its own mistakes, and transfer useful methods between tasks.

It must do more than answer difficult questions. It must deliver work that independent evaluators can inspect. It need not be conscious, embodied, omniscient, or competent at every physical task. This is my operational definition, not a scientific consensus.

I would test it on fresh projects in software, research synthesis, quantitative analysis, and business operations, with explicit budgets and held-out environments. Domain experts should judge results without knowing which system produced them. Report failures and human interventions alongside successes.

These details matter because “AGI” can otherwise move every time the evidence becomes inconvenient.

Why I lean toward a short timeline

My argument is about several capabilities reinforcing each other. Better models can write better tools. Better tools can expand the tasks models complete. Better evaluators can support more automated experimentation. Cheaper inference can make larger search programs affordable.

The Darwin Gödel Machine provides experimental evidence that an agent can improve its own software through iterative modification and evaluation. AlphaEvolve provides evidence of AI-assisted improvements to computational algorithms and infrastructure. Neither demonstrates the complete AGI loop I predict. Together they make the feedback-loop hypothesis more plausible to me.

Anthropic's 2026 automated-alignment-research work is another suggestive piece: agents can help search for interventions against measurable failures, while the researchers must still detect cheating and evaluate generalization. The optimistic interpretation is that some research work is becoming automatable. The limiting interpretation is that measurable research environments are much easier than open-ended science.

My eight-month forecast depends on the optimistic interpretation transferring farther than public evidence currently establishes. That dependency should be visible, not hidden inside a confident tone.

Why I may be wrong

The most serious obstacle is evaluation. If a system improves against a flawed judge, its progress can be illusory. A research assistant may produce beautiful explanations while missing a basic experimental confound. Scaling that assistant creates more convincing errors, not necessarily more science.

Another obstacle is long-horizon reliability. A model that succeeds on most short tasks may still fail a project containing hundreds of dependent steps. Errors compound; context drifts; recovery consumes time. A benchmark jump can coexist with frustrating real work.

There are also physical bottlenecks. Chips, power, networking, instruments, and manufacturing do not appear because a model wrote a good plan. AI can help design around constraints, but verification and construction still take time.

Finally, laboratories may have promising internal systems that remain too expensive, fragile, or difficult to control for the role I have defined. “They probably have something better internally” is a reasonable possibility, but not evidence that they have crossed a particular threshold.

What would change my mind

If by 29 June 2027 the strongest accessible systems still require frequent expert rescue on fresh cross-domain projects, I will regard this forecast as a miss. If impressive demonstrations depend on undisclosed human labor, narrow task selection, or budgets unavailable to any realistic user, they will not satisfy the spirit of the bet.

I would become more confident before then if independent evaluations show sustained multi-day work, reliable recovery, strong transfer to unfamiliar tools, and genuine reductions in human supervisory effort. Model release frequency alone would not be enough.

I am deliberately leaving room to lose. A prediction that cannot lose is a slogan.

Preparation should survive a missed date

For Africa, the useful response is to build capabilities that matter under both fast and slower timelines: scientific laboratories, reliable energy, inference infrastructure, technical education, secure payment systems, and institutions that can evaluate automated work.

Do not put essential public services or household security at risk on the assumption that my deadline must be right. Build a staged strategy. Start with valuable present capabilities. Expand investment when demonstrated performance supports it. Keep the option to adopt stronger systems quickly.

I want AGI soon because the world contains too many solvable problems left waiting for expertise, coordination, and resources. I want Africa to arrive as a builder, investigator, and owner of the resulting institutions.

The date is my bet. The preparation is the work.

On 29 June 2027, we should return to the wall and examine the evidence together.

Editorial note: This essay expresses the founder’s ideas, developed with AI-assisted research and writing and reviewed against the linked sources on 29 October 2026. Convictions, proposals, and forecasts are identified in the text; scenes framed as imagined are illustrative.

Cover photograph / illustration: ESO/Bayerischer Rundfunk. CC BY 4.0. Original source. Illustrative image: ESO Stars in Space Night.jpg. Credit: ESO/Bayerischer Rundfunk; CC BY 4.0. This image does not document the research or proposed facility described.