Predictions for AI in 2025
Happy new year, everyone! Before we dive in, here are a few things that we thought were worth sharing from the holidays:
Simon Willison’s Things we learned about LLMs in 2024 was an awesome (and very comprehensive) review of last year. A lot of the themes we pointed at in our prediction review (cost, multimodality) and themes from our posts last year (evals, model commodification, test-time compute) are reflected there.
Joey shared some expectations for 2025 on the DeepLearning.AI blog. Unsurprisingly, the themes align pretty well with what we’ve been thinking about around here as well. The focus on Return on AI is going to be crucial!
As with last year, we’re kicking off 2025 by sharing our predictions for the year. We weren’t as accurate as we would have hoped in 2024, but our goal is to get a little better every year. We’re again going to share predictions grouped by categories, and each prediction is going to have a confidence level associated with it: 50%, 70%, or 90%. 50% predictions are things that we think might happen but are highly uncertain (roughly a coin flip in likelihood). 70% predictions are things that we think are likely to happen but not guaranteed (538 predicted the 2016 election with 70% confidence). 90% predictions are things that we think are very likely to happen, but of course nothing is a guarantee.
One interesting (and good!) thing that we noticed in making these predictions is that we’re dramatically more humble than we were last year. In our first pass, most of our predictions were in the 50% and 70% category — it felt wrong to put things in the 90% category, and we had to go back and add new things into the mix to spread out the distribution. The goal of doing is both to share what we’re thinking but also to help us better understand the world, so some humility is a good thing.
Let’s dive in!
Foundation Models
Test-time compute is here to stay even if the phrase doesn’t make much sense. Model reasoning skills will continue to accelerate, primarily by using extra inference compute which will means cost reductions are dramatically lower than the previous two years. Meanwhile, we will see diminishing returns from extended training runs. While OpenAI has a lead, others (primarily Anthropic and/or Google) will likely catch up.
We will see a 90%+ score on ARC AGI. 50%
An Anthropic model or a Google model will match o3’s ARC AGI score. 70%
Llama 4 is released with a suite of inference-time optimization compute techniques. 70%
While it’s not part of the prediction, we don’t think it’ll be beam search!
The per-token cost of the latest version of the o series model is no more than 0.8x the cost at the end of the year as compared to the beginning of the year. 70%
GPT-5 is not released. 50%
Claude 4 is released. 70%
Marketing has not been the strong suit for model providers, but our guess it that the next major Claude release will include advanced reasoning capabilities.
o3-preview is not available for public use before April 2025. 90%
The models are exciting, but there’s a long way to go!
AI Applications
Enterprises have to adopt AI… or else. We will see a significant number of AI application companies grow revenue, and companies that are behind in the AI arms race (ahem, Salesforce) are likely to be looking for acquisitions. That might lead to some irrational behavior.
There will be at least 5 new AI applications unicorns minted in 2025. 70%
A large tech company (> $100B market cap) will spend at least $500MM on a strategic acquisition of an AI application. 90%
A large tech company (> $100B market cap) will spend at least $1B on a strategic acquisition of an AI application. 70%
Venture investments in AI applications in 2025 is within ± 20% (i.e., roughly the same level) as it was in 2024. 90%
Venture investments in AI applications in 2025 is ≥ 20% higher in 2025 as in 2024. 50%
A large financial institution shares 10%+ cost savings from using AI applications in 2025. 70%
Whether or not it’s true, people are going to want to talk about it.
There are no layoffs announced from a > $30B market cap company as a result of AI adoption. 70%
We think we’re far from AI-driven job loss — AI products will need to be built for their human coworkers.
Miscellanea
There will at least $1B invested by the US federal government into AI adoption. 90%
Whatever your political opinions may be, it’s clear that the folks coming into the new administration are focused on AI adoption and anti-regulation.
Google’s leading model will have a higher Arena ELO than Anthropic’s at the end of 2025. 70%
Google finally seems to have figured foundation models out, and we think they’re worth keeping an eye on. Anthropic has been comfortably in second, but their focus has been UX.
Perplexity remains under 100M MAU in 2025. 90%
*Current numbers are fuzzy, but we think truly AI-powered search has a long way to go. In the meantime, Google has caught up with the basics of summarization, keeping with the theme of this post.*
Venture investment in AI infrastructure (not counting foundation model companies) is ≥ 20% lower in 2025 than in 2024. 70%
This is the natural consequence of our theory of the AI market.
At least 1 AI unicorn minted in 2025 will have < 100 employees. 90%
Bonus (Qualitative)
There’s a lot of things that we think will happen but don’t have good ways to measure yet.
The use of AI applications will qualitatively change the way modern companies operate. 90%
We feel strongly about this one. The companies that are on the forefront of adopting AI are starting to operate differently, and there’s tons of maturation that will happen this year.
The gullibility challenge for agents — trying to solve every problem even if it involves tying yourself in knots — will not be solved. 70%
It’s hard to know what an acceptable level of gullibility is, but on the day we’re writing this, we just had Devin go make a bunch of nonsense changes in response to a question on a PR comment. 🙂
A SOTA AI model company will release a non-chat based UX application that garners significant usage. 70%
We put this here because we don’t know what exactly non-chat based UX means yet or what significant usage will be based on when it’s release — but we’re confident that there’s going to be a significant change (improvement) in AI UX.
AI applications will generate ≥ 10x revenue in 2025 compared to 2024. 70%
We really wanted to make a quantified prediction about AI application revenue, but it’s hard to know what numbers are right now, so we’re making a wild guess.
That’s all we have for now. It’s going to be an exciting year. The opportunities for innovation at the application layer are incredible, and there’s going to be a fascinating level of improvement in foundation models. We’ll keep you posted on how our predictions do!