Ask ten researchers when AGI arrives and you’ll get dates spread across four decades. Ask them what would have to happen first, and the answers converge with surprising speed. That gap is the useful part. Debating AGI milestones before 2030 gets you further than arguing about the finish line, because a milestone can be checked, dated, and proven wrong. “AGI by 2029” is a vibe with a year attached.
So here is a working list of the signposts that would count as real progress this decade, the ones that probably won’t show up in time, and how to separate a genuine step change from a launch event with good lighting.
What Actually Counts as an AGI Milestone?
Most public arguments squash capability, autonomy, and deployment into a single word. A system that scores 90% on a graduate-level exam is impressive and tells you almost nothing about whether it can run a three-week project without supervision.
Three different things people mean by “progress”
- Capability: the ceiling of what a model can do with a helpful human, good tools, and unlimited retries.
- Autonomy: what it manages when nobody is watching, over hours or days, on a fixed budget.
- Deployment: what survives contact with real users, messy data, and consequences.
Milestones worth tracking move all three at once. Benchmarks mostly move the first one. That’s why a model can jump 15 points on a leaderboard in a quarter and still fail to book a flight correctly.
The Milestones Most Likely to Land Before 2030
Task horizons stretching from minutes to days
METR’s task-horizon research is the cleanest measuring stick anyone has built. The finding that the length of task an AI can complete at 50% reliability roughly doubles every seven months has survived several model generations, and the 2025 revisions didn’t break the trend. A system that finishes a two-hour task today points toward one that finishes a full working day later this decade.
The headline number misses the important part. An agent that completes a ten-hour task half the time is a demo. One that completes it 95% of the time is a colleague. Watch for reliability curves, not peak scores.
AI that does credible AI research
Research automation is the milestone with the loudest hype and the muddiest definition. Systems already propose better matrix multiplication algorithms and generate workshop-grade papers that mostly get rejected. Neither is the threshold. The threshold is an AI producing a result that expert reviewers accept as novel and correct, then reproducing it independently without a human steering the experiments. Until replication happens, you’re watching a very fast literature review.
Continuous learning in production
Every deployed model today is frozen at training time. Retrieving from a memory file isn’t learning, and a long context window isn’t either. The milestone here is a system that picks up a new skill from a handful of examples, keeps it across sessions, and doesn’t degrade on everything else. Catastrophic forgetting is the tell. When a model can acquire a skill and hold it for a month without fine-tuning, that’s a real line crossed.
Agents you’d trust with a budget
Occupational evaluations like GDPval, where models produce actual deliverables from real job categories, are a better proxy than most academic benchmarks. The practical test is transactional: hand an agent $500 and a supplier list, ask for a correctly invoiced order, and see what comes back. Error recovery is where these systems fall over, not first-attempt quality.
The Date-Keeping Record Isn’t Great
Anyone tracking AGI milestones before 2030 should spend an afternoon with the historical forecasts. The track record of AGI predictions shows a pattern of confident dates sliding sideways whenever the technical bottleneck turned out to be different from the one people expected. Forecasters in 2015 were worried about commonsense reasoning. They got it, roughly, and immediately discovered that long-horizon planning was the harder problem.
Aggregate probability curves tell a similar story. When you plot AGI probability by year across major forecasters, the mass clusters between 2035 and 2050, with a stubborn minority betting on the late 2020s. The minority is worth taking seriously. It just isn’t the consensus, and headlines often imply otherwise.
What Probably Won’t Arrive Before 2030
A closed recursive self-improvement loop
For AI to accelerate its own development, you need a complete cycle: the model designs a better model, the lab trains it, an evaluator confirms the improvement is real, and the process repeats without a human unblocking each step. Compute and experiment latency are physical constraints, not software ones. Training runs take weeks. Even a perfect research agent can’t compress that queue.
Human-level sample efficiency
A teenager learns to drive safely in about 20 hours of instruction. Driving systems have absorbed billions of miles and still struggle with construction zones. Closing that gap is a different research problem from scaling pretraining, and nothing in the current paradigm obviously solves it. Expect incremental gains, not a step change, before 2030.
Genuine cross-domain transfer
Generality in the AGI sense means carrying a skill learned in one domain into a structurally different one. Models today transfer well within a format and poorly across them. A system that wins a chemistry problem and then applies the same reasoning structure to a supply-chain dispute would be a genuine surprise. The evidence for near-term AGI is thinner than the marketing suggests precisely because transfer is where it breaks down.
Reading the Announcements Without Getting Fooled
Every few months a lab ships something and the word “AGI” appears in a headline within 48 hours. The useful skill is knowing which claims to test. Telling a real AGI milestone from a marketing beat comes down to a few questions: was the evaluation run by an independent party, was the task chosen before the model was trained, and does the demo show the median case or the best of forty attempts?
Stagecraft matters too. Keynote demos are edited, latency is hidden, and the failure cases get cut. Reading the stagecraft of AGI conference announcements is a skill you can build in an afternoon and use for the rest of the decade.
What to Watch, Quarter by Quarter
Instead of waiting for a single announcement, track a small set of indicators that update regularly. Each one is boring on its own. Together they tell you whether the curve is bending.
- Task-horizon doubling time. Is it still around seven months, or has it stretched to twelve?
- Reliability at the 95th percentile. Not the best run. The worst acceptable one.
- Cost per completed task. Falling cost is the quiet signal that a capability is becoming deployable.
- Error recovery rate. Can the agent notice it went wrong and fix it without a human?
- Independent replication. Who else ran the evaluation, and did they get the same number?
None of these require trusting a press release. All of them are published, argued over, and revised in public.
The Signpost That Deserves Your Attention
If you only track one thing, track reliability on long tasks. A system that completes a 40-hour project with 95% success, at a cost below the human equivalent, changes more about the economy than any benchmark record. It’s also the milestone least likely to be announced with a countdown video, because it’s hard to make look dramatic.
The date AGI arrives will probably be argued about for years after the fact. The milestones on the way there are checkable now, and most of them are due before 2030 if the trend lines hold. Watch the boring numbers. They’re the ones that don’t get revised downward after the launch event.