AGI: Timelines & Definitions

This course is for anyone who wants to evaluate AGI claims — in a headline, an investor pitch, a policy debate, or a colleague's hot take — without either dismissing or accepting them on reputation alone. It's especially useful for founders and investors making multi-year bets, policymakers and analysts tracking frontier AI governance, and anyone planning a career or curriculum around an uncertain timeline.

No technical AI background is required. This is a course about how to read evidence and disagreement, not about how transformer models work internally.

There is no single, universally accepted definition of “artificial general intelligence.” Three influential framings currently compete, and each implies a different way of deciding whether AGI has “arrived.”

DeepMind's framework is worth understanding in more detail because it's the most explicit about grading, not just declaring. Its five performance levels are defined by percentile against skilled human adults: Emerging (equal to or slightly better than an unskilled human), Competent (outperforms 50% of skilled adults), Expert (90th percentile), Virtuoso (99th percentile), and Superhuman (outperforms 100% of humans). By this scheme, the paper's authors classify current large language models like ChatGPT as “Emerging AGI” — broad in the tasks they attempt, but not yet reliably better than an unskilled human across most of them, let alone at the “Competent” level. No system has reached a higher level as general-purpose AI as of the paper's analysis.

The practical takeaway: “has AGI arrived?” is not a single yes/no question. The more useful question is “AGI by which definition, at which performance level, in which domain?” — and most headlines that claim otherwise are quietly picking one definition without saying so.

Because static benchmarks get solved and replaced, researchers increasingly track trend lines instead of single scores. Two are widely cited:

Both organizations are explicit that these are descriptions of a historical trend, not a guaranteed law. Extrapolating an exponential curve years into the future is exactly the kind of forecast that has been wrong before in the history of technology — which is precisely why Module 3 treats the resulting date estimates as a wide range, not a countdown.

Three independent, verifiable data points show just how wide the real spread is — and none of them is obviously wrong.

None of these groups is obviously more credible than the others by default. Lab leaders have the closest technical view but also the strongest incentive (fundraising, recruiting, policy influence) to sound imminent. Surveyed researchers are less commercially exposed but were asked a single definitional question that may not match any one lab's roadmap. Superforecasters have the best calibration track record on short, resolvable questions, but that skill has never been tested on a question this far out. The honest conclusion is that this is a live, unresolved disagreement among serious people — not a case of some of them simply not knowing what they're talking about.

You don't need to resolve the expert disagreement to evaluate a specific claim well. Four questions catch most overstated AGI headlines:

1. Which definition are they using? Economic output (OpenAI), a performance level (DeepMind), or a capability list (Amodei-style)? Claims rarely state this, but the definition determines whether the claim is even coherent.
2. Is this a benchmark score, a real deployment, or an announcement? A strong eval result, a supervised pilot, and a press release are three very different levels of evidence — and headlines routinely collapse them into one.
3. What's the source's incentive? A funding round, a product launch, or a policy debate all reward sounding closer to (or further from) AGI than the evidence alone supports.
4. Is the underlying trend continuing, plateauing, or being reinterpreted? Check the actual metric — task horizon, compute, benchmark saturation — rather than the framing around it.

Use the template below the next time an AGI claim crosses your feed — paste it into your notes, or into an AI chat tool alongside the claim, to force yourself through all four questions before you update your beliefs or your plans.

I'm evaluating this AGI-related claim: [paste the headline, quote, or announcement here]

Walk me through it against this checklist:
1. Which definition of AGI (or "powerful AI") does this claim implicitly use — economic output, a specific performance level, or a capability list? Is that stated or assumed?
2. Is the underlying evidence a benchmark score, a real-world deployment, or a company announcement/funding event? Be specific about which.
3. Who is making the claim, and do they have a financial, political, or reputational incentive for it to sound more or less imminent than the evidence supports?
4. What concrete, trackable metric (if any) does this claim relate to — and is that metric's trend continuing, plateauing, or being reinterpreted based on recent data?

Give me a one-paragraph honest assessment of how much this specific claim should move my beliefs, separate from how it was framed.

Put the three data points from Module 3 side by side and the honest range for “AGI or powerful-AI-equivalent capability” runs from roughly 2026 (the most aggressive lab forecasts) to 2047 and beyond (the researcher-survey median), with serious, calibrated forecasters on record doubting the shorter end. That is not a failure of the field to agree — it is the actual state of the evidence.

The practical implication: career, investment, product, and policy decisions that only make sense if a specific date is correct are fragile bets, not informed ones. Decisions that hold up across most of that 2026–2047+ range — investing in judgment and accountability-heavy skills, building governance now rather than waiting for a clear signal, tracking the underlying metrics instead of the headlines — are the ones worth prioritizing today.

From here, The Coming Wave covers how governance and containment frameworks are trying to prepare for that range of outcomes, and The Future of Work applies the same “don't bet on one date” logic to your own career.