AI predictions for 2025

Date: January 1, 2025

There aren’t many concrete milestones left on the road to AGI. A few years ago, I listed realistic video generation, math research, and solving MIT Mystery Hunt puzzles as the capabilities that would cause me to shorten my AI timelines if they were achieved anytime soon. Today, significant progress has been made on the first two of those, and solving the whole Mystery Hunt is probably equivalent to AGI. So I’m here to put forth some predictions for the next year as a test of my forecasting skills.

But first, let’s briefly go over the past year’s advances in AI. (If you’ve been keeping up with the field, you can skip to my predictions; for a more extensive review of LLM progress, check out Simon Willison’s post.)

State of AI

A lot happened in 2024. Unlike any other field, a year or two can be the difference between the almost unthinkable and the commercialized.

• Self driving cars that are safer than humans exist. They’re called Waymos and if you visit SF you’ll see them all over. If you’re in a city where they operate, you owe it to yourself to try one out and experience living in the future.
• In a coding test recently, I wrote essentially zero lines of code, instead relying on Cursor (powered by Claude). My programmer friends have had similar experiences.
• DeepMind’s AlphaProof system performed at the level of a silver medalist on the 2024 International Math Olympiad (IMO), though it was given significantly more time to think.
• AI art is far better than a year or two ago—misshapen hands are a thing of the past. Good AI art can be indistinguishable from human art.
• Video generation is decent now: compare Veo 2 to VideoPoet from one year earlier.
• Robot dogs have advanced tremendously since the days of Boston Dynamics. They’re now faster and more agile than the average human, even on rough terrain.

o3

OpenAI quietly announced this model two weeks ago, and it caught many off guard given that they had launched the o1 model just two weeks earlier. Outside of the AI community there’s been surprisingly little coverage of o3, but it represents a tremendous leap forward in LLM capabilities. By applying enormous amounts of compute to inference-time reasoning (as opposed to simply training larger models), OpenAI was able to attain results on science, math, and coding tests on par with the best humans. Access to the model is very limited for now, so we won’t know for a while what it’s like in the real world.

Predictions

Not everything here is original to me, but all of these are things I think will come to pass in 2025.

• AI jingles and art will proliferate in advertising, especially on the low-budget side. Some companies will experiment with AI-powered personalization to tailor ads to thousands of niche markets, or generating variants of ads to A/B test.
• Waymo will expand to a handful more cities with good weather but the rate of expansion will be hampered by high fixed costs (lidar, GPUs). Tesla FSD will improve to approximately Level 3—still behind Waymo but much more widespread.
• AI will be able to (in principle) apply, get accepted, and graduate from an online university with a STEM degree. This entails acing standardized tests and writing compelling college essays (which are already possible today), but also watching video lectures, completing coursework, and doing associated browser tasks.
• AI will get a gold on the IMO under the same limits as humans (i.e., the IMO Grand Challenge will be solved). Professional mathematicians will use reasoning LLMs (possibly attached to formal theorem provers like Lean) to provide inspiration in proofs and solve simple research problems on the level of a math PhD student.
• AI video will be as good as images are now. That is to say, the average output might not trick an attentive observer, but it’ll be hard to point to any flaws in the best generations.
• Very high-quality AI music will be possible, but its release will be hampered by copyright concerns.
• The entry-level call center job will be done better by AIs. There will be some notable mishaps amplified by the news media but it’ll generally work well.
• AI avatars will look and sound natural. You’ll be able to have video calls with AIs that can control and pick up on tone and body language. There will be much fretting about what this means for the future of human interaction/relationships, with anecdotes of lonely folks who spend hours videocalling their AI partners.
• Many people will consume AI-generated short-form content without knowing it (Reddit posts, tweets, TikToks). Some will bemoan the ascendancy of slop, but the quality will be good enough for cheap entertainment.
• Most lines of code at startups will be written by AI agents that are given high-level instructions by programmers. Rather than having to one-shot programs, they will iterate on designs autonomously. Even those with no programming experience will be able to create fairly sophisticated apps on the level of an unaided CS undergrad.
• The gap between the best closed- and open-weight models will remain under a year, though the age of running top models on a laptop (or even a top gaming GPU) is over.
• There will be some even cooler videos of robots doing household tasks or acrobatics, but you won’t see any in real life.
• Reasoning models with search capabilities will solve at least one 2025 MIT Mystery Hunt puzzle.

Conclusion

You can’t escape AI. You’re going to hear more and more about it every year. You will probably be shocked at new developments several times a year even if you are paying attention. To quote a sticker an Anthropic recruiter gave me, things will never be chill again.