In July 2025, OpenAI and Google DeepMind independently announced that their models had reached gold-medal standard at the International Mathematical Olympiad. Neither system was purpose-built for contest maths. They were general reasoning models with some scaffolding bolted on for the competition. Within a fortnight, the result had been absorbed into the ambient noise of AI news and everyone was back to arguing about chatbots.
That cycle is worth understanding if you’re trying to follow AGI breakthroughs seriously. The field is producing genuine, verifiable results at a pace that would have looked ridiculous in 2020, and the discourse metabolises each one so fast that it feels unremarkable by the following Monday. Telling the real thing from the marketing means asking what was actually demonstrated, on what benchmark, and whether any of it transfers.
The target keeps moving
Part of the confusion is definitional. There’s no agreed finish line, so every announcement gets measured against a private, shifting standard. If you want a clear-eyed primer on what AGI actually means and how close we are, that’s a sensible starting point before you read another breathless thread about it.
What I’d argue instead is that a real breakthrough has three properties: it’s verifiable by outsiders, it generalises beyond the task it was trained on, and it would have been hard to fake. Most announcements satisfy one of those.
Science is where the results are hardest to argue with
Protein structure prediction remains the clearest case. AlphaFold 2’s 2020 CASP14 performance was followed by a database of more than 200 million predicted structures, now used by over two million researchers in 190 countries. The 2024 Nobel Prize in Chemistry went to Demis Hassabis and John Jumper for it, alongside David Baker for protein design. That’s a domain-specific system rather than a general one, but the scientific impact is not in dispute.
Then there’s AlphaEvolve, published in May 2025. It found a way to multiply two 4×4 complex-valued matrices using 48 scalar multiplications, beating the 49 that had stood since Strassen’s algorithm in 1969. The same system recovered roughly 0.7% of Google’s global compute through better scheduling heuristics, which sounds boring and is worth an enormous amount of money.
Maths has become the favourite proving ground, partly because answers can be checked mechanically. When OpenAI’s Astra worked through ten long-standing open problems, the interesting question wasn’t whether the model could produce plausible-looking proofs. It was whether referees could verify them and whether the techniques generalised.
Reasoning models bent the curve sideways
For years, progress meant bigger pre-training runs. The o1 generation in late 2024 introduced a different lever: spending more compute at inference time, letting a model search over its own reasoning before answering. The effect on hard benchmarks was steep.
The best public measure of this is METR’s task-horizon work, which estimates how long a task a model can complete with 50% reliability. Frontier models went from handling tasks that take a skilled human minutes, to roughly 50 minutes in early 2025, to more than two hours by late summer. The doubling time has hovered around seven months for a while now. Extrapolate that and you get some strange numbers within a few years, which is exactly why people argue about whether the trend can hold.
Where the claims start to wobble
Benchmark saturation is the obvious failure mode. MMLU and GSM8K were useful signals until they weren’t, and now they’re mostly decoration in launch posts. The ARC-AGI-2 suite was designed as a resistance test for exactly this reason, and frontier models still score in the low single digits to low teens on it while humans clear 60%. That gap is not a rounding error.
Cherry-picking is the other problem. Companies choose the evals, run them in-house, and rarely publish the failures. There’s a real methodological question buried in all this, and it’s the same one behind when we can credibly say an AI made a scientific discovery rather than generating a plausible-sounding hypothesis that a human then had to rescue.
Four constraints that haven’t broken yet
- Energy. Multi-gigawatt data centre builds are now a grid-planning problem, not a procurement one. Power is becoming the binding constraint on scaling before chip supply does.
- Data. High-quality human text is finite. Synthetic data helps and also risks model collapse and benchmark contamination in ways that are genuinely hard to audit.
- Reliability. A system that succeeds 95% of the time per step fails 64% of the time over 20 steps. Long-horizon autonomy lives or dies on that arithmetic.
- Memory. Models still largely relearn the world at each deployment. Continual learning without catastrophic forgetting remains unsolved in production systems.
Agents are the honest stress test
Chat benchmarks flatter models. Agents expose them. The moment a system has to hold state, recover from errors, and act on a real environment with real consequences, the failure surface explodes. The rollout of autonomous coding and browser agents in 2025 gave a good demonstration of what happens when capability and guardrails get out of sync, and the way AI agents created operational chaos is more instructive than any leaderboard.
This is where I’d look for the next meaningful leap. Not a higher score on a reasoning suite, but an agent that finishes a multi-day project unsupervised, notices when it has gone wrong, and asks for help at the right moment.
Risk arguments are getting narrower and better
The catastrophe discourse has improved mostly by shrinking. Instead of vague warnings about superintelligence, researchers now trace specific pathways and find specific bottlenecks. The case that AI is unlikely to hand a lone actor bioweapons rests on constraints in synthesis equipment and procurement, not on model refusals. That kind of concrete analysis is far more useful than either doom or dismissal.
Signals worth tracking over the next eighteen months
Ignore the demo videos. Watch for these instead:
- Task-horizon estimates crossing the eight-hour mark, independently replicated.
- Any frontier model that updates its weights during deployment without degrading on held-out evals.
- Cost per completed task falling by an order of magnitude rather than a few percent.
- Third-party evaluations on unseen benchmarks, with failures published alongside successes.
- AI-generated results that experts verify but admit they would not have found themselves.
The last one is the real test. Every other milestone is a proxy for it. And on the current evidence, the honest position sits somewhere between “closer than comfortable” and “a decade, maybe more” depending on whether the reliability curve bends. Anyone telling you they know which is selling something.