OpenAI and Google win gold at global math competition

Quick Note

FYI President Trump is scheduled to unveil his AI Action Plan later today. This plan will effectively replace former President Biden’s AI executive order, so there will be some interesting policy changes in play. Check back tomorrow for more details.

OpenAI and Google’s AI models win gold medals at global math competition

AI models from OpenAI and Google DeepMind achieved gold-medal scores at the 2025 International Math Olympiad (IMO), showcasing how quickly the technology is advancing despite failing to beat five human participants. 

Key background: The IMO is one of the world’s oldest and most challenging high school-level math competitions. Participants attempt to solve six problems within nine hours over the course of two days. 

The rundown:

• Both companies entered reasoning models that solved five out of the six problems within the time limit, earning 35 out of 42 possible points and a gold medal score. 

• That puts both models on par with the top 10% of human contestants who also won gold medals, although five participants received perfect scores. 

• This is the first time that any AI system has scored gold at the IMO. Tech commentators say this signals that OpenAI and Google are evenly matched in the AI race. 

Why it matters: Google has entered this competition before. Last year the company earned a silver medal, but only by using a system that required humans to translate each problem into a machine-readable format. This time around, OpenAI and Google both entered models that were able to digest questions and generate answers entirely on their own. Although their scores show that AI still hasn’t surpassed the world’s smartest highschoolers, this is a big step up from last year.

How do AI models solve problems they’ve never seen before?

Remember when I told you that AI doesn’t really think? You’re probably wondering how that can be true if OpenAI and Google were able to compete in the IMO, which uses brand-new, original problems every year. 

It all comes back to pattern recognition. AI models are trained to replicate patterns from their training data, which likely includes hundreds of thousands of math problems. When an advanced AI model sees a new problem, it can usually break it down into familiar patterns and go from there. 

What’s interesting is that this is kind of similar to how students solve new problems. A high schooler might see a tough problem on a test and follow the exact same process, breaking it down into familiar steps to eventually solve it. This makes advanced math one of the best measures of AI’s “intelligence.”

Mixus believes that email is the key to usable AI agents

San Francisco-based startup Mixus is attempting to make AI agents easier to access and monitor by integrating them into existing communication platforms, rather than forcing users to adopt another third-party application.

Key background: AI agents are designed to complete tasks independently, unlike standard models that specialize in language processing and require more guidance from users.

The details:

• Mixus’ founders say that because almost the entire workforce uses email, it presents an opportunity to democratize access to agents. 

• The company is building an agent that users can access by simply emailing instructions to [email protected]. It will then follow up with its work directly within the same email thread.

• Technology news outlet Techcrunch (which tends to be brutally honest about these things) placed Mixus’ tech at the higher end of the agent spectrum after a private demo.

The bigger picture: Tech companies promised big things from agents in 2025. They largely haven’t delivered, in part because they’re still hard to access. Even building a simple agent like an email assistant requires a lot of coding, both to create the agent itself and then to integrate it into your email provider. Mixus’ product aims to offer a much simpler alternative.

More trending news

Amazon acquires AI wearables startup that records everything users say

Apple relaunches AI-generated news summaries months after complaints by the BBC

ChatGPT users send 2.5 billion prompts according to Axios

Thanks for reading! How was today’s newsletter? My goal is to make these emails as easy to read as possible — if there’s anything I can do for you to achieve that goal, let me know!

See you tomorrow,

Chain of Thought