The slide goes up at 9:42 in the morning. A horizontal arrow, four dots, the last one stamped with a year somewhere between 2027 and 2030. The room lifts its phones. By lunchtime the photo is on six news sites with the headline “Lab says AGI by 2028”, and by dinner it has been quoted in a policy memo, three podcasts, and a very confident LinkedIn post from someone who wasn’t there.
That’s the standard life cycle of AGI conference announcements. The claim takes about four seconds to make and about four years to properly evaluate. And the gap between what got said on stage and what the lab actually committed to is usually wide enough to drive a research programme through.
Why a keynote is a different animal from a paper
Researchers hedge. Keynotes don’t. A paper can say “under certain assumptions” eleven times and nobody blinks; a keynote speaker who sounds uncertain gets cut from the highlight reel. That asymmetry is why the conference circuit is where the industry’s most quotable language gets manufactured, and why the most consequential announcements often appear as a blog post 48 hours before the talk, safely out of the Q&A window.
There’s a commercial logic too. A credible AGI timeline moves hiring, compute contracts, and enterprise procurement cycles within days. A single session at NeurIPS, ICML, or one of the dedicated AGI gatherings can change what a procurement team is willing to sign in Q3. The audience isn’t only listening. It’s buying.
The three shapes an AGI announcement usually takes
The roadmap slide
This is the one that travels furthest. Since DeepMind published its “Levels of AGI” framework in late 2023, everyone has a vocabulary for tiered progress, and tiers photograph well. The useful question isn’t which level a company claims to have reached. It’s whether the level definition comes with a published eval, or whether it exists only on the slide.
The benchmark claim
FrontierMath saturation, ARC-AGI records, Humanity’s Last Exam scores, agentic task benchmarks. These are the most falsifiable announcements, which makes them the most valuable and the most abusable. A number without the scaffold around it, meaning compute budget, attempt limits, tool access, and contamination checks, tells you close to nothing.
The safety framework drop
Capability thresholds, preparedness frameworks, evaluation gates before deployment. Increasingly these arrive paired with a capability announcement so that both land at once. Watch for whether the framework includes a defined pause condition and who has authority to call it. A threshold without a named decision-maker is a press release wearing a lab coat.
When a claim arrives with none of these, or with a demo that can’t be rerun independently, that’s the moment to apply the same tests you’d use to separate a real milestone from a marketing beat.
A genuine capability claim usually includes five things
- A name and a version. Not “our model” but a specific checkpoint, so results can be tied to something.
- The eval, not just the score. Task count, attempt budget, tool access, and whether the test set was held out.
- A cost curve. A result that needed $20,000 of inference per task is a research result, not a product roadmap.
- The failure cases. Speakers who volunteer where the system breaks are almost always the ones you can trust with the parts they don’t.
- A date you can check later. Vague timelines are unfalsifiable. “A public eval by June” is not.
Benchmarks move fast; the definition doesn’t
Worth remembering how quickly the ground shifts. Early in 2024 the best publicly reported score on ARC-AGI’s first version was around 5%. By December of that year a reasoning model had pushed past 87%, at a cost per task of thousands of dollars and with a lot of test-time compute thrown at it. Same benchmark, same year, an order of magnitude of apparent progress.
Both data points are true. Neither settles anything. That’s the trouble with milestone-shaped announcements: they measure distance along a track that someone else chose, and the choice of track is itself a claim about what general intelligence is. A lab that defines AGI as economic substitution will announce it at a very different moment from a lab that defines it as open-ended scientific discovery, and both will do so with total sincerity.
The academic circuit and the corporate circuit are not the same room
The dedicated AGI conference scene, the AGI-24 and AGI-25 gatherings in Seattle and beyond, runs on academic time: proceedings, peer review, arguments that continue into the corridor. The corporate circuit runs on announcement time. The two overlap in the media and almost nowhere else.
For practitioners, the academic track is usually the better early-warning system. You’ll hear about an architecture two years before a keynote mentions it, and you’ll hear the objections that a product launch quietly omits. If you want to know which of this year’s dates survive contact with reality, it helps to look at which AGI forecasts have actually held up rather than which got the most applause.
Announcements that aren’t about AGI at all
Half the news value at a modern AI conference is off-topic. Security researchers turn up with demos that get a fraction of the attention a timeline slide gets and three times the practical consequence. One that stuck with me: a walkthrough showing how a two-minute bypass defeated the EU’s new age-verification app. No stage, no countdown, no roadmap. Just a working exploit and a slide with a timestamp.
That’s the pattern. Deployment friction, regulatory exposure, and infrastructure cost tend to be announced quietly, in side rooms, by people who don’t have a marketing team. They also tend to determine whether the grand timeline is reachable at all.
How to read a roadmap slide while it’s still on screen
Four questions, roughly in this order. What exactly is being claimed, in a sentence you could falsify? Who is claiming it, a researcher with a track record or a comms lead with a deadline? What would have to be true for this to happen, in compute, data, and energy terms? And when can I check back?
Write the answers down with the date attached. Review them in six months. You will be surprised how often the most confident slide in the room turns out to have quietly changed its wording, and how often a throwaway line from a workshop turns out to be the thing everyone cites two years later.
Keep a running file for this. One line per announcement: date, claimant, exact claim, the falsifier, and a review date. Then actually review it. A team that does this for a year stops being surprised by AGI conference announcements, not because the announcements improve, but because the difference between a genuine capability jump and a well-lit slide becomes obvious at a glance.