My One-Year-Old Predictions for What the World Will Look Like in 3 Years
I originally published these predictions on January 28, 2025 on my Ukrainian-language Telegram channel. I am re-posting them here in English, fourteen months later, mostly so they are registered somewhere visible and searchable, and so that I and others can track them properly over time. I have not edited the substance in any way; the translation is faithful to the original. The three-year horizon means these predictions target roughly January 2028. I would make slightly different predictions now, less because the trajectory has changed and more because my knowledge has improved.
My predictions for what the world will look like in 3 years (if we are still alive by then):
The major AGI labs are either nationalized or quasi-nationalized, meaning at the very least they operate under the protection and audit of government agencies.
Frontier models are no longer available to the general public, partly because a single query to such models costs tens or hundreds of thousands of dollars, and partly because they are now considered national security assets.
There is a full-scale trade and R&D confrontation between the US and China, with export controls on chips becoming a near-total ban on chip exports.
Frontier models are distinctly superhuman in all scientific and engineering domains where (1) physical experiments are not required, and (2) clear success criteria exist in principle. No mathematician or programmer, no matter how brilliant, can come even close to competing with AI given the same amount of time, and (less likely, but still fairly likely) given the same budget. In other theoretical fields (theoretical physics and chemistry, computational biology), the situation looks more like AI is better than any individual scientist, but sometimes it is hard to tell, opinions vary, and sometimes you need to run many experiments to verify, which is slow, complex, and expensive. The general trend is: the more a field relies on physical experiments, the less certainty there is that AI is superhuman.
Even in heavily experiment-dependent fields, many specialists consider AI to be smarter than themselves, though many do not. There are numerous examples of achievements where humans worked together with AI, but it is impossible to draw a clear line as to who contributed what. Either way, everyone seems to agree that without AI, these achievements would not have happened in this timeframe.
In biology, a large number (from dozens to thousands) of AI-generated drug candidates are going through clinical trials, and most experts generally expect them to pass, but this still requires many more years, and there are ongoing debates about how effective these drugs will be.
In macroeconomic statistics, the effect of AI is either very small or entirely unnoticeable. Most jobs that existed as of 2025 are performed better by AI agents than by humans, but mass displacement of workers has not occurred, and unemployment does not appear to be rising. Mass roboticization has not happened either, although robots that outperform humans in all or nearly all tasks involving interaction with the physical world have long since been developed.
However: many companies and startups that used to require thousands of employees now have only 10–50 staff. It is possible that the first unicorns with only a handful of employees are emerging. Young and successful companies frequently operate as “5–20 human managers + thousands of AI agents.” The capitalization and profits of some companies with microscopic teams look absolutely unprecedented.
More than a trillion dollars has been spent in total on building out server infrastructure. This massive buildout is somewhat slowing down, or at least looks like it should slow down, because energy scarcity is starting to play a role: servers already consume more energy than a large European country or an American state, and building nuclear power plants is not exactly a fast process.
Thanks to AI, a certain number of fantastic discoveries in the style of AlphaFold have occurred, meaning: AI solves some fundamental problem at a level that was considered impossible, but still imperfectly; then several edge cases are found where the problem remains unsolved; and so, although what happened are essentially absolutely fantastic discoveries, they are somewhat downplayed because of the unsolved edge cases. Of course, when it comes to discoveries in mathematics and computer science, things are “entirely unambiguous”: the discoveries exist, they are unexpected, and everyone acknowledges them.
Jailbreaks, at least in frontier models, stop working.
Either some relatively major AI incidents have occurred (less likely), or conversely, any external manifestations of misalignment that are periodically observed in models in 2025 have disappeared (more likely) or nearly disappeared.
All points from “Marcus’s bets” whose realization can in principle be verified have been realized. That is, excluding items 3, 7, and 8.
Nvidia is trading at a market cap of at least 7 trillion dollars, quite possibly more than 10 trillion.
Open-source models are not the leaders.
Nearly all (except, of course, Marcus) leading AI researchers believe that AGI has either almost certainly been created or has already been created, depending on the definition.
#1 seems to be confusingly worded. Is the banking industry in the US ‘quasi-nationalized’? The pharma industry? Construction? Restaurants? All of these do ‘operate under the protection and audit of government agencies.’ What level of government involvement are you envisioning?
Why do you expect that? The hardness of jailbreak creation did not grow as fast as capabilities, to the best of my knowledge.