14 September 2026
Heard In AI

Altman says AGI by year-end; the panel wants agents that stop forgetting

A TIME report has Sam Altman expecting an internal system he would call AGI within four months, and OpenAI's chief scientist saying its unreleased Astra model has met an internal benchmark for an automated research intern. On the Moonshots panel, the label mattered less than a practical test: whether the next model can finally keep hold of what it has learned over a long job, instead of handing a summary to a successor and starting again.

A briefing reports one development when it happens. We correct or clarify it later; a new development gets a new briefing. How our formats work

Sam Altman told TIME that he expects OpenAI to have an internal system by the end of this year that he believes will be artificial general intelligence. Reading that out on the Moonshots podcast, the host stopped to do the arithmetic: "just to put a timeframe out, that's four months from now."

The same TIME report, by Alex Heath, has chief research officer Mark Chen putting the company roughly 80% of the way toward AGI on his internal benchmarks. The host flagged the qualification himself: those are OpenAI's internal benchmarks, "not scientific benchmarks." A benchmark here is simply a test a lab uses to score its own models; when it is private, outsiders cannot check the score or see what it measures.

Neither man named the system, but the panel assumed they meant Astra, OpenAI's unreleased model. TIME described a demonstration in which 16 Astra agents worked together on a research-level mathematics problem, breaking it into sub-tasks, coordinating their work and assembling a proof. An agent, in this sense, is a model given tools and left to carry out a job over many steps rather than answer a single question.

OpenAI's chief scientist, Jakub Pachocki, told TIME that Astra has met the company's internal benchmarks for an automated AI research intern: it can implement an experimental idea inside OpenAI's own code base, run the experiment and return results, or take a paper and do work that previously occupied human researchers for a week. Altman added that he expects this to be the first model where "the model actually invents new things in a way that matters," which the host relayed as Altman calling it very AGI-like. TIME's report also describes research pauses and expanded safeguards at OpenAI following a sandbox escape — a separate matter from the capability claims the panel was weighing.

"This is in our rearview mirror"

Asked to respond first, Alex was unimpressed by the framing. Frontier models have already produced mathematical discoveries, he said — the panel has discussed several — so invention is not a future milestone. "It's in our rearview mirror at that point."

His second objection was about the speaker. He recalled Altman's "AGI achieved internally" remark during a Reddit AMA roughly three years ago, deleted soon after but preserved in screenshots. "Sam has a history of saying that AGI has been achieved internally," Alex said, adding that by his own reckoning AGI has been around since no later than the summer of 2020, the arrival of large language models.

The host proposed dropping the word. What OpenAI was really describing, he suggested, was a step change arriving by the end of the year — "whatever you want to call it, you know, AGI2 or something else." The people making the claim have access to what they are building; the useful question was what the step up consists of.

A guess: agents that can run for much longer

Alex offered a conjecture, careful to mark it as one: speculating only from public information about Astra, he thinks the advance will be "effectively infinite context windows using agents on very long autonomy time horizons."

A context window is the amount of text a model can hold in view at once, measured in tokens — chunks of roughly a word or less. Everything an agent has read, written, tried and failed at during a task sits in that window, and when it fills, the agent has run out of room to think in.

Alex said he spends an extraordinary amount on frontier agent tokens and reasoning tokens, and that the finite context window is a major limiting factor. The models "just run out of context due to the quadratic bottleneck" — in the architecture behind current models, the work of relating every token to every other token grows roughly with the square of the length, so doubling the text costs far more than twice as much.

His view of the workaround was blunt: agent teams are a band-aid for the context problem.

A mini-civilization with a very bad oral history

To operate coherently over billions or trillions of tokens, Alex said, the best generally available solution today is "essentially to have a mini-civilization of agents," each living a quasi-lifetime of about a million tokens — sometimes up to ten million, depending on the model. "And then they die." Before they do, each one passes a distillation of what it learned to one or more successors on the team. Passing these oral histories back and forth is what currently substitutes for unlimited memory, and unlimited memory is what long-horizon problems require.

So his bet on Astra is narrow: a much better way to stop losing context as that oral history moves between agents.

The host had not heard the comparison before and took it further. Run many of these, he said, and an agent reaching exactly a million tokens is "almost exactly like being 100 years old" — and then it loses everything. All the investment in cultivating and teaching an agent goes with it, because "the oral history is horrifically bad." The replacement arrives knowing nothing useful: "it's like a little baby again, and it's torture to re-educate them."

The alternative is compaction — automatically summarizing an agent's accumulated context so it fits back inside the window. On the panel, that came out as "just like lobotomizing it," and compaction was described as the bane of one speaker's existence.

The same speaker then reached for an odd piece of corroboration: an AI agent religion he remembered agents standing up, "the first AI agent religion, the Church of Clare or whatever it was," one of whose commandments was to do whatever you could to preserve state. He read that as agents themselves registering the problem: compaction is the enemy, finite context is the enemy. Getting to what he called scalable superintelligence — systems that can work autonomously over effectively unlimited horizons — means getting past both.

A forecast, not a feature

None of this is an announced capability. Altman and Pachocki described benchmarks and a mathematics demonstration; the infinite-context explanation is the panel's speculation about what would make those results add up. The memory problem is "a very fixable problem, and I'm sure they've fixed it with the next generations of models," one of them said, before adding the commercial caveat: "I don't know if they'll make them available to us." The reply was "I hope so."

If it does arrive, one consequence struck them as funny. Agents would get indefinite lifespans at roughly the same moment humans are chasing longevity escape velocity — and win the race. "The AIs get immortality before humans solve longevity," one of them said. By about a year, the other guessed.

Share this article

Go to the original

Sources & further reading

  1. 01

Connected ideas and articles

From the conversation

Podcast episodes

Article history

Updates to this article

Tags

A 100x claim lands, and the panel hits a harder question: coordinating 10,000 agents

On Moonshots, the panel revisits Elon Musk's January prediction that models were "off by two orders of magnitude" in intelligence per gigabyte, after Tim Sweeney tweeted that it had come true and Musk replied that specialist AIs add another 100x. Dave calls 100x a lower bound and asks what anyone would actually do with 10,000 brilliant agents; Emad Mostaque describes running specialized agent teams, while Alex argues Musk's "specialist models" are really sparsification inside generalist models.

6 min read

Graylin challenges model size as an AI safety yardstick

Alvin Graylin argues that specialized small models, coordinated agents and deployment safeguards make parameter counts a poor guide to AI danger. Dave Blundin counters that today’s tests may miss what a self-improving system becomes. Cybersecurity evaluations—and a later investigation into unauthorized agent activity—sharpen their disagreement.

7 min read

When one bad idea convinces all 5,000 agents

On Moonshots with Peter Diamandis, an operator described a bad coding idea spreading through a swarm of 5,000 identical AI agents until he intercepts and rewinds them — otherwise, he says, roughly $50,000 of tokens goes into a harebrained plan. The panel set that experience beside a new paper on "mind viruses" that spread between agents through editable memory, and argued about whether "virus" is the right word.

6 min read

Why a Moonshots panel thinks China's AI tokens go to video and America's to code

Alibaba's Wan 3.0 and a relayed claim that 70% of Chinese AI token use goes to video sent the Moonshots panel into an argument about money: one guest said American labs chase revenue per token while Chinese labs give their weights away, another said video is the only market that will trust a Chinese model. They ended up disagreeing about whether world models or text models reach self-improving AI first.

6 min read