AGI Timeline: What Founders Should Build Now

Every founder now gets some version of the same question from an investor, a candidate, or a co-founder at midnight: what happens to this company if AGI shows up in three years.

It is a fair question and a badly formed one. The answer depends much less on the date than on what the person asking means by AGI, and the people best placed to know do not mean the same thing. The useful move is not to pick a year. It is to work out which of your bets change if the year is right, and which change if it is wrong by a decade.

What an AGI timeline actually is

An AGI timeline is one person's estimate of when AI systems reach artificial general intelligence: competence at problems they were not specifically built for, at roughly human level, without being retrained for each new one.

The estimate travels badly on its own. A date is only as meaningful as the definition attached to it, and the definitions in circulation are not compatible with each other. Two of the most credible people in the field currently name the same year and mean substantially different things by it. That gap, not the year, is the part you can actually plan against.

It is worth separating AGI from what usually gets discussed next. Superintelligence, or ASI, is the step past it: systems that exceed the best humans across essentially every domain rather than matching a competent one. Almost every forecast in circulation, including the ones below, is about AGI. The gap between the two is contested and nobody has a credible date for the second, which is another reason to plan against capabilities you can observe rather than labels you cannot.

The AGI timeline the labs quote: around 2030

Around 2030, from the person running the lab most likely to get there first.

Demis Hassabis co-founded DeepMind in 2010 and leads Google DeepMind today. His lab's protein structure work, AlphaFold, won him a share of the 2024 Nobel Prize in Chemistry. In April 2026 he sat down with Garry Tan for a live episode of Y Combinator's How to Build the Future and gave the number as a planning input rather than a prophecy:

"Depending on what your AGI timeline is, you know, mine's like 2030 or something like this"

He has been consistent about it in public. At Google I/O in May 2025, Axios reported that Hassabis and Google co-founder Sergey Brin both put AGI at around 2030.

There is a faster camp, and it is the one most of the public argument is actually about. AI 2027, published in April 2025 by Daniel Kokotajlo, Eli Lifland, Thomas Larsen and Romeo Dean of the AI Futures Project, sketched a scenario in which AI systems automate AI research itself within a couple of years, and it shaped more conversations about timelines than any lab statement did.

The interesting part is what its authors did next, because they publish their revisions. By January 2026 they had pushed their own medians for transformative AI substantially later: Kokotajlo's moved from 2028 at publication to around December 2030, and Lifland's from 2031 to January 2035, on evidence including a slower than expected pretraining picture. Then in April 2026 both moved roughly 1.5 years back toward the present, on faster than expected progress in agentic coding.

Sit with that for a second, because it is the strongest argument in this entire article. The most engaged full-time forecasters in the field revised their own numbers by years, in both directions, inside twelve months. They were not being sloppy. They were doing the honest thing and updating on evidence. If that is what careful forecasting looks like up close, then a date is not something you build a company on. It is something you keep re-reading.

What gets quoted less is how much uncertainty sits underneath. Hassabis thinks the current ingredients survive into the final architecture. Large-scale pre-training, RLHF and chain of thought have come far enough that he cannot see the field waking up in two years to find they were a dead end. But he is explicit that the recipe may not be complete: "it could be that there's still one or two big ideas left" that need to be cracked, and he puts his own odds on that at about 50/50.

Read that as a founder and it says something a headline never will. The person with the best view inside the leading lab is describing a coin flip on whether a fundamental unsolved problem stands between here and AGI. A timeline with a coin flip inside it is not a schedule. It is a scenario, and scenarios get planned for differently.

Why two experts share an AGI timeline and disagree about everything else

They agree on roughly 2030 and on almost nothing else about what it means.

François Chollet created the Keras deep learning library, published the ARC-AGI benchmark in 2019, and launched the ARC Prize in 2024. He left Google in November 2024 to co-found Ndea, a lab pursuing AGI through program synthesis, with Zapier co-founder Mike Knoop. He is the field's best-known scaling skeptic by construction: ARC-AGI was built to test skill-acquisition efficiency rather than accumulated knowledge, which is exactly what scaling buys.

On Y Combinator's Lightcone he landed on almost the same year:

"I think we're probably looking at AGI 2030. Early 2030s, uh most likely."

Same decade, opposite reasoning. And when you look at what each man means, the agreement dissolves.

Chollet rejects the definition most of the industry uses. He notes that many people say AGI will be a system that can automate most economically valuable tasks, and says of that framing: "It's not about intelligence. It's not about general intelligence." His own test is about learning efficiency, not output: "a system that can approach any new problem, any new task, any new domain" and become competent at it "with the same degree of efficiency as a human could." Humans need very little data. A system that needs a corpus to learn each new thing has not passed.

Then there is a third definition, the one your customers will actually feel. At Sequoia's AI Ascent 2026 keynote, partner Pat Grady declined the technical argument outright, saying "we're venture capitalists not about to propose a technical definition for AGI," and offered a functional one instead: "if you can dispatch an agent to do a job and it can recover from failure and persist until that job is done."

So: three definitions, one date range.

• Hassabis points at a capability set: continual learning, long-term reasoning, memory, consistency.
• Chollet points at efficiency: how little data it takes to become competent at something new.
• Grady points at a job: work that finishes without a human rescuing it.

Grady's is the only one with a delivery date for your customers, and the only one that is already partly true. That is the practical read for the next few years. Nobody is going to announce that AGI arrived. What will happen instead is a sequence of specific jobs that quietly stop needing a person in the middle. Attach your planning to those jobs, not to the word.

What is still missing before the AGI timeline lands

Hassabis names the gaps precisely, and they are unusually concrete for a topic this abstract. Still unsolved: "continual learning, long-term reasoning" and some aspects of memory, plus getting systems to be consistent across the board. His verdict on all of them: "I think all of these are going to be required for AGI."

Continual learning is the one with the most immediate consequence for anything you ship this year. Today's models are stateless between sessions, and Hassabis is blunt that "not having continual learning currently is one of the things holding back agents" from completing whole tasks rather than pieces of them. They do not adapt to the specific context you put them in, which is exactly the gap between an agent that is useful for part of a job and one you can hand the job to.

He is equally blunt about the current workaround: "we're kind of using duct tape right now. So, like shove it all in the context window." And he makes a point that anyone who has watched a long agent session degrade will recognize. The problem is not storage. Even with a perfect ten million token memory, "there's still a cost to looking it up and finding the right thing" that matters for the decision in front of you. He gives the concrete case: an agent naively recording tokens from a live video feed burns a million tokens in about twenty minutes. That failure mode has a name and a set of fixes, and it is the single most useful thing to get good at while continual learning stays unsolved: see context rot for how to catch it, and AI memory for what a real memory layer does instead.

This is not one man's view. Engram, founded by Dan Biderman and Jessy Lin, is a company built on the premise that memory and continual learning are two sides of the same coin, and that the answer is baking knowledge into model weights rather than stuffing ever-larger prompts into the context window. When the head of the leading lab names a gap and an independent team raises money to close it, treat the gap as real and near-term rather than philosophical.

The planning rule Hassabis gives founders

Assume AGI arrives in the middle of your build, not at the end of it. That is the rule, and it arrives almost as an afterthought at the end of the interview.

The arithmetic is simple. A genuine deep tech company is, in his estimate, roughly a ten-year journey. His AGI timeline is about 2030. For anyone starting now, those two overlap by construction, so "you have to just consider AGI appearing in the middle of that journey."

His read on that is not alarm. "It's not bad necessarily, but you have to take that into account." And the instruction is specific: "imagine what that world would look like and then build something that would be useful if that comes in halfway through."

That reframes the midnight question entirely. The question is not whether AGI kills your company. It is whether your company gets better or worse if a much more capable general system arrives in year five. Those are different bets, they live in different rows of your roadmap, and you can sort them today without knowing the date.

Hassabis also gives a structural hint about what that world looks like. He does not expect one enormous model that does everything, because cramming every specialized capability into a general system degrades it: putting all of protein folding into Gemini "wouldn't make sense" and would cost it elsewhere. He expects general systems to use specialized ones as tools. If that is right, the unit of value is not the model. It is the tool, the workflow around it, and the judgment about when to call which.

Which bets survive the AGI timeline being wrong

Hassabis's own answer for where to build is consistent across the interview: "there's huge scope for combining where AI is going with some other deep technology area," ideally with a founding team expert in both. Those areas, in his words, are "pretty safe from just getting swamped by whatever the next update is to the foundation models," and "that's one of the more defensible areas I would say."

The counter-pressure is real, and Grady states it from the investor's side: "no lead is safe," because "there is a torrential downpour of new capabilities coming out of the foundation models."

Hold both and you get a sorting rule rather than a guarantee. Three properties tend to survive being wrong about the date in either direction:

1. A second discipline in the loop. If the hard part of your product is chemistry, biology, actuation or regulated physical process, a model update does not hand your problem to a competitor with a prompt. This is the bet behind AI drug discovery and physical AI, and it is the specific thing Hassabis is pointing at. The cost is that these teams are harder to hire for and slower to start, which is precisely why the advantage lasts.
2. A clean seam where the model is replaceable. If a better model makes your product better the week it ships, capability progress is a tailwind. If integrating it is a rewrite, every release is a fire drill. Design for the swap: see how to build for the next AI model.
3. Data and workflow you own. The corpus that accumulates from customers doing real work in your product is the one thing a lab cannot train on. That is the durable version of a moat, covered in data for AI and competitive moats in AI.

What does not survive is the bet nobody says out loud: a thin layer whose only advantage is that today's model cannot quite do the job unaided. That position is short the timeline, and it loses whether AGI arrives in 2030 or 2045. The only difference is how long you get to keep pretending.

How to tell the AGI timeline is moving: track markers, not dates

The most useful thing in the interview is not the year. It is that Hassabis hands out falsifiable markers, which is rare in this genre. A date cannot update. A marker can, and it moves your plan when it clears.

• The vibe-coded hit game. "We haven't seen a AAA game that tops the App Store charts that was sort of vibe coded yet." He says he can prototype in half an hour what took him six months at seventeen, and yet no kid has shipped a hit game, so something in the process or the tools is still missing. He expected the payoff he had not yet seen to land "in the next 6 to 12 months" from April 2026, which puts the window through roughly April 2027. He does not say the marker has to be a chart-topper specifically, so watch the whole class: a one or two person team shipping something commercially serious.
• Invent Go, not move 37. AlphaGo's famous move was creative inside a game. The higher bar is inventing the game: given a high-level brief, produce Go. On today's systems he is unambiguous, "clearly today's systems, I think, can't do that."
• The Einstein test. "Can you train a system with the knowledge of cutoff of 1901" and have it produce what Einstein produced in 1905. He suggests simply running the test periodically, which is better advice than it sounds: it converts a philosophical argument into an experiment with a date on it.
• A genuine scientific discovery. Every frontier lab is tinkering here, and his personal assessment is that he has not seen it yet. When a system produces a discovery rather than a solution to a posed problem, the definitional argument is largely over.
• The virtual cell. On a full working simulation of a cell, he says "we're about 10 years away probably from something like a virtual cell." That marker matters because it sits well past 2030, which tells you the timeline is uneven by domain. General capability arriving does not mean every hard domain falls at once.

Track three of these in whatever document you already read weekly. If the game marker clears next quarter, your assumptions about what a two-person team can produce should move that quarter, whether or not anybody has declared AGI.

The security decision an AGI timeline forces early

If you accept that a much more capable system may land in the middle of your build, the uncomfortable follow-on is that the data architecture you pick this year is the one it will run on. Three decisions get much cheaper to make now than to retrofit later.

Where inference happens. Hassabis makes the case for capable local models "for efficiency reasons, but also for privacy and security reasons, too," with the pattern being sensitive processing kept on the device and selective delegation to a frontier model in the cloud: "you process all of the audio-visual feed, let's say, locally, and that stays local." If your product touches personal, clinical or regulated data, that architecture is easier to adopt before you have a cloud-only pipeline in production. Self-hosted AI agents covers the tradeoffs.

What you retain, and on what consent. Data collected today under a vague permission becomes far more revealing when a more capable model reads it in 2029. Consent and retention are architecture decisions with a long tail, not legal boilerplate to sort out at Series A.

What the system may do without a human. Grady's functional definition is also a blast radius. An agent that can recover from failure and persist until a job is done is an agent that can persist through a mistake with great determination. Draw the autonomy boundary explicitly per workflow, and make the expansion of that boundary a decision someone signs rather than a default that drifts.

Hassabis's own framing on the risk is not marketing. Asked about how Promethean the science work feels, he agrees, then adds that "we have to also be careful with how we use that and what we use it for," including the misuse that can happen with the same tools. The founder version of that sentence is smaller and more actionable: the more capable the system you are planning for, the earlier the guardrails have to exist, because you will not get to add them under load.

What to do this week

1. Write your own definition of AGI in one sentence, and note which of the three above it resembles. If your investors mean Grady's and you mean Chollet's, you are having two different conversations about the same roadmap.
2. Sort your roadmap into two columns: bets that get better if a much more capable general system arrives in year five, and bets that get erased by it. Do not act on it yet. Just look at the ratio.
3. Find the seam. Identify every place a model swap would require more than a config change and a re-run of your evals, and put the worst one on next sprint.
4. Pick three markers from the list above and add them to whatever you already review weekly. Delete the date from your planning doc.
5. Make the two data decisions now: what you retain and under what consent, and which workloads never leave your infrastructure.
6. Name the one thing in your product that a foundation model update cannot ship next month. If you cannot name it in a sentence, that is this quarter's real problem, not the AGI timeline.

The date is not the decision. The date is an input to a decision you can make this week, which is whether your company compounds with capability or competes against it.

The pattern underneath all of this runs through the AI for startups pillar: the durable advantage is never the model, it is the thing you build around it. If you want the full system for running a startup this way, that is what I teach in AI Operating System for Startups.

Sources

• Demis Hassabis: Agents, AGI & The Next Big Scientific Breakthrough, Y Combinator's How to Build the Future with Garry Tan, April 2026. The interview this article distills: the 2030 estimate, the unsolved capability list, the deep tech planning rule, and the marker set.
• François Chollet: Why Scaling Alone Isn't Enough for AGI, Y Combinator's Lightcone, March 2026. The near-identical date reached by opposite reasoning, and the competing definition.
• This is AGI: Sequoia AI Ascent 2026 Keynote, Sequoia Capital, April 2026. Pat Grady's functional definition and the "no lead is safe" framing.
• Memory and Continual Learning: Engram's Dan Biderman and Jessy Lin, Sequoia Capital, June 2026. An independent team building against the same continual learning gap.
• Google leaders see AGI arriving around 2030, Axios, Ina Fried, May 2025. Corroborates the 2030 estimate from Hassabis and Sergey Brin at Google I/O.
• AI 2027, AI Futures Project, April 2025, and the authors' own revisions: Clarifying how our AI timelines forecasts have changed since AI 2027 (January 2026) and the Q1 2026 timelines update (April 2026). The faster camp, and the record of it moving in both directions.
• Background: Demis Hassabis (2024 Nobel Prize in Chemistry) and François Chollet (Keras, ARC-AGI, Ndea).