What happened with AI in 2024?

The idea this week: despite claims of stagnation, AI research still advanced rapidly in 2024.

Some people say AI research has plateaued. But a lot of evidence from the last year points in the opposite direction:

• New capabilities were developed and emerged
• Research indicates existing AI can accelerate science

And at the same time, important findings about AI safety and risk came out (see below).

AI advances might still stall. Some leaders in the field have warned that a lack of good data, for example, may impede further capability growth, though others disagree. Regardless, growth clearly hasn’t stopped yet.

Meanwhile, the aggregate forecast on Metaculus of when we’ll see the first “general” AI system — which would be highly capable across a wide range of tasks — is 2031.

All of this matters a lot, because AI poses potentially existential risks. We think making sure AI goes well is a top pressing world problem.

If AI advances fast, this work is not only important but urgent.

Here are some of the key developments in AI from the last year:

New AI models and capabilities

OpenAI announced in late December that its new model o3 achieved a large leap forward in capabilities. It builds on the o1 language model (also released in 2024), which has the ability to deliberate about its answers before responding. With this more advanced capability, o3 reportedly:

• Scored a breakthrough 87.5% on ARC-AGI, a test designed to be particularly hard for leading AI systems
• Pushed the frontier of AI software engineering, scoring 71.7% on a key benchmark using real tasks compared to 48.9% accuracy for o1
• Achieved a 25% score on a new (and extremely challenging) FrontierMath benchmark — while previous leading AI models couldn’t get above 2%

While not released publicly yet, it seems clear that o3 is the most capable language model we’ve seen. It still has many limitations and weaknesses, but it undermines claims that AI progress stalled in 2024.

It may be the most impressive advance in 2024, but the last year had many other major developments:

• AI video generation gained steam, as OpenAI released Sora for public use and Google DeepMind launched Veo.
• Google DeepMind released AlphaFold 3 — a successor to a Nobel Prize-winning AI system — which can predict how proteins interact with DNA, RNA, and other structures at the molecular level.
• Anthropic introduced the capability for its chatbot Claude to use your computer at your direction.
• AI systems are increasingly able to take audio and visual inputs, and larger amounts of text, while also engaging with users in voice mode.
• By combining the models AlphaProof and AlphaGeometry 2, Google DeepMind was able to use AI to achieve silver medal performance in the International Mathematical Olympiad.
• The Chinese company DeepSeek said that its newest model only cost $5.5 million to train — a dramatic decrease from the reported $100 million OpenAI spent training the comparably capable GPT-4.

And there’s a lot more that could be included here! We won’t be surprised if 2025 and 2026 see many more leaps forward in AI capabilities.

AI helping with science

Recent research indicates that AI can help speed up scientific progress, including AI research itself:

• AI appears to have improved top material science researchers’ output by around 80%, speeding up their ability to generate new ideas.
• AI agents could outperform human experts on ML research engineering tasks when given two hours, though the humans take the lead over longer periods of time.
• Researchers found that language models seem to be able to discover fundamental patterns in the neuroscientific literature, making them better than human experts at predicting the outcomes of new studies in the field.
• AI-discovered molecules appear more likely to pass through phase I clinical trials.

Some key developments in AI risk and safety research

Meanwhile, we’ve seen a mix of encouraging and worrying results in research on AI safety. Here are a few of the important publications this year:

• A recent paper from Anthropic found that AI models can “fake” being aligned with their developers’ intentions when they’re told they’re being trained, only to abandon these behaviours when deployed.
• Apollo Research found that under certain conditions, OpenAI’s o1 model has the capability to scheme to deceive its developers and even pretend to be less capable than it really is.
• In a paper on “Sleeper Agents,” researchers discovered that AI models can have deceptive and unwanted behaviours that standard safety training won’t eliminate.
• A new approach to preventing AI misuse and improving alignment involves using “circuit breakers” to directly target internal representations of harmful outputs, rather than simply trying to train models to refuse to produce harmful outputs.
• Anthropic reported that it successfully identified millions of “features” — patterns of neurons that can be linked to human-understandable concepts — within one of its frontier AI models. This points the way to better understand of how these systems actually work.

What you can do

These developments show the fast pace and potential risks of advancing AI. To help, you can:

• Learn about the the biggest risks AI poses
• Learn about how to carefully communicate about this topic
• Learn how to use AI in your own work
• Consider switching your career to work in AI governance or AI safety
• Consider other potential jobs that can help

We also recommend checking out recent posts from our founder Benjamin Todd on:

• Funding opportunities in AI safety
• How ordinary people can prepare for the possibility of transformative AI in the near future

We plan to continue to cover this topic in the coming year, and we wouldn’t be surprised to see many additional changes and major AI developments. Continue following along with us, and consider sharing this information with your friends by forwarding this email if you find it helpful.

Learn more:

• Can AI Scaling Continue Through 2030? from Epoch AI
• What just happened: A transformative month rewrites the capabilities of AI by Ethan Mollick
• An update from Jack Clark’s “Import AI” newsletter: “OpenAI’s O3 means AI progress in 2025 will be faster than in 2024.”
• Synthetic data is more useful than you think by Lynnette Bye
• Is AI progress slowing down? Making sense of recent technology trends and claims by Arvind Narayanan and Sayash Kapoor
• Things we learned about LLMs in 2024 by Simon Willison
• How A.I. Could Change Science Forever — a video by Cool Worlds