In 1901, two years before helping build the first heavier-than-air flyer, Wilbur Wright told his brother that powered flight was still fifty years away. Meanwhile, fusion power has famously been “twenty years away” for the past fifty years. Historically, scientists and engineers have seen far more success saying whether a technology is feasible and saying what the technology will be like, than predicting when a breakthrough will unlock that technology.
We consider the timeline to ASI (artificial superintelligence) uncertain, and given the difficulty of timing such developments, we expect it to remain similarly uncertain until ASI is imminent, or until it already exists. If policymakers are to intervene and prevent this technology from being built, then they will need to act before it is entirely clear how much time we have left.
That said, our best guess is that the timeline to ASI is probably not long. Recent progress has been sufficiently rapid that we at MIRI would not be surprised if ASI is developed in two or five years, and we would be surprised if it were still more than twenty years away.
Recent inventions like ChatGPT have helped bring AI into the public eye, but these inventions don’t represent a once-off breakthrough. Rather, tools like ChatGPT are the latest of a long series of advances.
Twenty years ago, researchers argued that ASI was far off because of the many things AI couldn’t do — basic image recognition and generation, translation, Winograd schemas, computer programming, games like Go and StarCraft, and so on. All of those challenges have fallen, with the result that the gap between us and ASI is made mostly of intangibles. While there are still many problems state-of-the-art AI fails on, it has become increasingly common to see people claim that AI will never solve some problem, only to learn that current AI already excels at that task.
Progress from current systems to smarter-than-human systems is likely to exhibit threshold effects. Chimpanzees and humans have broadly similar brains, and are only separated by a few million years of evolution; but the results of chimpanzee and human cognition are radically different, because humans happened to cross one or more thresholds that chimpanzees don’t reach.
While evolution occurs on geologic timescale, human technological progress occurs on a timescale of months and years; and AI may exhibit even sharper discontinuities than humans did, thanks to additional threshold effects such as self-improvement.
Over the last few years, AI has rapidly come to play an important role in software development. Once the ability of AIs to automate AI research tasks and improve on its general capabilities crosses a critical threshold, AI is likely to experience runaway growth, as additional improvements further accelerate the process of self-improvement. By the time this feedback loop hits sharply diminishing returns, it could easily be vastly beyond the human capability range, if self-improving AIs fell within that range in the first place.
These factors make it harder to rule out very aggressive timelines, while still leaving us in a highly uncertain situation.
No observations or theory rules out the possibility of ASI being built in the next year; nor does any observation or theory rule out the possibility that large language models could hit a wall and researchers could flounder as they look for new and better approaches. Full solutions in AI can come surprisingly early, as in the case of AlphaGo; but they can also come surprisingly late, as in the case of ironing out rare failure modes for self-driving cars.
Threshold effects and “the gap between us and ASI is made mostly of intangibles” also make it less likely that the situation will become clearer as new AI advances roll out. As new challenges fall, it’s likely to continue to be unclear how quickly the next challenges will fall, or what the gap is between the current state of the art and ASI.
Given this level of uncertainty, it seems reasonable to ask why MIRI researchers think five- or ten-year timelines are any more likely than, e.g., thirty-year timelines. The short answer, which seems to match how many other researchers are thinking about timing ASI, is: There isn’t any sophisticated formal method we can use in this context to calculate a year. What we can do is hazard a guess that some combination of algorithmic progress and hardware scaling is likely to unlock smarter-than-human general cognitive abilities relatively soon, based on the large qualitative improvements we’ve seen recent algorithmic advances produce in tasks ranging from Go to image generation to carrying on a conversation, and based on the qualitative leaps in capability GPTs have exhibited with increased scale.
Some researchers have argued that AI development is more predictable than this. A commonly cited paper in this context is Kaplan et al. (2020), which proposed “scaling laws” for performance improvement as model size, training data, and computation are increased. However, the kinds of measures tracked by scaling laws don’t give us practically important information about the capability levels of systems. Attempts to identify particular levels of computing power, data, etc. with “human-level intelligence” or “superintelligence” appear to be almost purely speculative, and equating empirical trendlines with these guesses can give the misleading impression that they’re on a similar footing.
We don’t know whether current architectures, typified by large language models, can scale all the way to superintelligence. But AI companies are searching hard for alternative architectures, and the most recent state-of-the-art models are likely already making use of architectural tweaks and refinement.
There have been past periods of AI hype followed by slowdowns, such as the AI winter of the late 1980s and 1990s. We may or may not see additional slowdowns, large or small. However, AI’s extremely rapid progress over the past decade has resulted in a situation very unlike the 1980s bubble.
Tens of billions of dollars and tens of thousands of researchers and engineers are pursuing what they believe to be a trail of gold. If simple scaling doesn’t work, they aren’t going to suddenly give up. Innovations the size of transformers — the key idea behind AIs like ChatGPT — don’t land every year, but they land occasionally. If scaling stops working on current architectures, the AI industry will try harder on other algorithms.
The field’s large increase in research effort and funding, then, suggests that progress is likely to speed up over the coming decade. This leaves us in the uncomfortable position of not knowing how many additional AI advances are survivable, but knowing that an enormous effort is underway to roll out technical breakthroughs. MIRI Senior Researcher Eliezer Yudkowsky’s best guess — albeit it’s only a guess — is that it’s likely to take anywhere from zero to two more innovations as significant as transformers in order for developers to achieve ASI.
We can also note, if only in passing, that the heads of top labs are forecasting fully general AI within a few years. In February 2024, Google DeepMind CEO Demis Hassabis predicted “AGI-like systems” by around 2030. Even Yann LeCun, Chief AI Scientist at Meta and often seen as a skeptic or par-human AI, said in October 2024:
I said that reaching Human-Level AI “will take several years if not a decade.”
[OpenAI CEO] Sam Altman says “several thousand days” which is at least 2000 days (6 years) or perhaps 3000 days (9 years).
So we’re not in disagreement.
But I think the distribution has a long tail: it could take much longer than that. In AI, it almost always takes longer.
In any case, it’s not going to be in the next year or two.
This particular line of evidence seems weaker from our perspective. On their own, we don’t put much weight on lab heads’ public predictions — labs can always hype their product. However, it seems worth noting that companies are speaking in terms of 5- or 10-year timelines, rather than 20- or 50-year ones; and it seems worth noting that even LeCun is in this boat.
Our own evaluation, while uncertain, largely agrees with these assessments. ASI may or may not be built with methods that resemble modern LLMs; but researchers are pursuing novel methods as well as efforts to scale LLMs.
As effort and funding scales up, it becomes likelier that someone will succeed, whether or not the world is equipped to survive ASI.
