But won’t we run out of data before AI goes all the way? Or electrical power? Or funding?
Probably not.
Humans use data much more efficiently than AIs do, so we know it’s possible in principle for intelligent minds to be much more data-efficient than modern AIs are. If AI labs “run out” of data when it comes to making LLMs better, that will slow them down for only as long as it takes to invent new methods that are more data-efficient.
Humans also use power much more efficiently than AIs. We run on 100 watts, which is proof that there’s no fundamental obstacle to general intelligences that run on about as much power as a large light bulb. Not only has leading AI hardware been getting 40 percent more energy-efficient each year, but algorithmic improvements mean that, by one 2024 estimate, over the years 2012 to 2023, “the compute required to reach a set performance threshold has halved approximately every 8 months.”
Remember that the field of AI has existed for far longer than the LLM architecture has, and it’s pretty good at coming up with new architectures that surmount obstacles. And more generally, when humanity has put its best minds and resources to something known to be possible, it has a pretty strong track record of success.
With skilled AI researchers now routinely commanding seven-figure salaries (or nine figures, for top leadership roles) and annual private investment in AI now measured in the hundreds of billions of dollars, it looks like the talent and resources will be there to overcome anticipated bottlenecks. See also our discussion of the field’s success at overcoming obstacles.
Don’t expect another “AI winter.”
People have been wrongly predicting an imminent “AI winter” for the last decade now. AI winters used to happen back in the 1970s through the 1990s, when AI funding was public and the public funders got sick of the lack of results. Because the AI of old did not, in fact, produce results.
Things have changed. ChatGPT was perhaps the fastest-adopted app in history, and it’s printing money hand over fist. It generated $3.7 billion in revenue in 2024, with projections to generate $12.7 billion in 2025. It’s spurred by private investment, and it’s making enough money to attract the top talent in the world without any public source that could cut them off.
It’s still possible that AI techniques will hit some sort of wall, and that humanity will have a respite of some sort before superintelligence hits. But the old pattern of “AI winters” — of public funding, no results, and a decline — has been shattered.
Notes
[1] large light bulb:As mentioned in a previous endnote: McMurray et al.’s paper gives an average basal metabolic rate (the minimum resting energy consumption) of about 0.863 kilocalories per hour per kilogram, which works out to about 1 watt per kg or about 60-80 watts for a human. That’s only 60-80% of total energy expenditure, which — including physical activity — is about 100 watts.