DeepMind's AI math genius

PLUS: Altman speaks on GPT-5, Stargate, and open-source

Author

Rowan Cheung
February 10, 2025

Good morning, AI enthusiasts. While chatbots previously stumbled over basic arithmetic, DeepMind just announced that its AlphaGeometry2 outperforms math olympiads.

With AI about to reach superhuman math capabilities, are the next wave of models about to unlock a whole new realm of scientific discovery?

In today’s AI rundown:

  • DeepMind AI surpasses math olympiads

  • Sam Altman speaks on GPT-5

  • Run CLI commands from your prompts

  • Apple's lamp robot makes expressiveness key to home AI

  • 4 new AI tools & 4 job opportunities

LATEST DEVELOPMENTS

GOOGLE DEEPMIND

Image source: Ideogram / The Rundown

The Rundown: Google DeepMind just introduced AlphaGeometry2, a new version of its math-focused AI model that solved 84% of International Mathematical Olympiad geometry problems from the past 25 years — surpassing the average gold medalist performance.

The details:

  • The system combines a Gemini model with a symbolic engine to tackle complex geometry problems requiring rigorous proofs and deductive reasoning.

  • AlphaGeometry2 solved 42 out of 50 problems to surpass the average gold medalist score of 40.9, a massive improvement from its predecessor's 54% solve rate.

  • The model generated over 300M synthetic theorems and proofs of increasing difficulty for training, featuring a larger and more diverse set than AG1.

Why it matters: Math has typically been one of the areas that language models seem to struggle with (sometimes in simple and comical fashions). Still, DeepMind is quickly cracking the code to unlock systems tackling super-complex problems. This can also play a key role in accelerating other math-heavy scientific areas like physics.

TOGETHER WITH VANTA

The Rundown: Join John Hammond — cybersecurity researcher, practitioner, and content creator — and Matt Cooper, Vanta’s Director of GRC, for a fireside chat on AI, security maturity, and the top security risks in 2025.

What you’ll gain:

  • A deep dive into 2025’s top cyber risks, including the impact of AI

  • Actionable insights to refine your security priorities

  • Strategies tailored to your organization’s security maturity level

  • A live Q&A at the end

Save your spot and get ready to future-proof your approach to cybersecurity.

OPENAI

Image source: UTokyo Center for Global Education

The Rundown: OpenAI CEO Sam Altman and CPO Kevin Weil spoke at the University of Tokyo Center for Global Education, handling questions on the future of AI, upcoming model advances like GPT-5, the company’s new Stargate Project, and more.

The details:

  • The CEO said the $500B Stargate Project will allow the next class of models to start developing the first ‘new bits of scientific knowledge.’

  • He also revealed that OpenAI’s internal model ranks as the 50th-best programmer in the world and could reach the No. 1 spot by the end of the year.

  • Altman expects OpenAI ‘to move in the direction of open source,’ saying that ‘society is willing to take the trade-offs’ that open models bring.

  • Altman also compared AI’s capability increases to ‘trying to outrun the calculator,’ saying the tech will be better in every general domain than humans.

Why it matters: Nobody in the business is better than Altman at making mind-bending statements about the future of AI — and this talk had no shortage of them. But if his confidence regarding the coming AI acceleration is any indication, we could see some wild paradigm shifts, like new scientific discoveries, coming sooner rather than later.

AI TRAINING PRESENTED BY PROJECT IDX

The Rundown: Project IDX is your AI-enabled development environment in the cloud with code assistance from Gemini.

Step-by-step:

  1. Log into Project IDX with your Google account.

  2. Create a project from the templates dashboard.

  3. Click the Gemini icon at the bottom of the workspace or press Cmd+Shift+Space (Ctrl+Shift+Space on ChromeOS, Windows, or Linux).

  4. Select “Interactive Chat with Gemini” and pass it a prompt to update your project configuration files and code.

Pro tip: Project IDX offers a streamlined development experience by allowing you to execute terminal commands directly from the user interface.

APPLE

Image source: Apple

The Rundown: Apple researchers introduced ELEGNT, a framework aiming to showcase how natural movements can make home robots more engaging and relatable — using an expressive robotic lamp inspired by Pixar's Luxo Jr for demonstration.

The details:

  • Apple's prototype combines basic functionality with movements that convey emotions and intentions, like "looking" out a window when discussing weather.

  • The robot integrates Siri's voice capabilities while using its movable head and arm to create more natural interactions through gestures and positioning.

  • Testing revealed that expressive movements, like nodding or showing curiosity, significantly improve comfort and engagement compared to static responses.

Why it matters: Tech companies are racing to bring robots into our homes, and while many have featured the typical humanoid builds, Apple's research suggests that success may depend on both advanced capabilities and creating devices that can interact in ways that feel more natural and emotionally resonant to users.

QUICK HITS

  • 🇫🇷 Le Chat - Mistral’s revamped AI assistant platform with 10x response speed and new iOS and Android apps

  • 🤖 GitHub Copilot - Microsoft’s coding assistant with new agentic features

  • 🎥 Pikadditions - New video-to-video feature that enables users to integrate any subject or object into existing footage

  • 👅 Tough Tongue - Multimodal AI agent for navigating difficult conversations

Ilya Sutskever’s SSI is reportedly in talks for new fundraising at a $20B valuation, a 4x increase from September’s round despite no public product or revenue.

OpenAI is establishing a new office in Munich, citing the country’s leading position in European AI adoption with the highest amount of ChatGPT users, paying subscribers, and API developers outside of North America.

OpenAI co-founder John Schulmann is reportedly joining former OpenAI CTO Mira Murati’s new startup after leaving Anthropic after just five months.

Perplexity announced ‘The Million Dollar Question,’ incentivizing users to use the platform and ask questions during the Super Bowl for a chance at a $1M prize.

Over 2,000 artists signed an open letter calling for the cancellation of ‘Augmented Intelligence,’ an upcoming AI art auction at Christie’s — arguing the models use copyrighted work in training.

Krea officially launched its previously teased Chat tool in open beta, allowing users to generate and edit images via a natural language chat interface.

COMMUNITY

Join our next workshop this Thursday at 4 PM EST to learn how to build your own AI-powered personal assistant to streamline your daily tasks with Dr. Alvaro Cintas, The Rundown’s AI professor.

RSVP here. Not a member yet? Join The Rundown University on a 14-day free trial.

We’ll always keep this newsletter 100% free. To support our work, consider sharing The Rundown with your friends, and we’ll send you more free goodies.

That's it for today!

Before you go we’d love to know what you thought of today's newsletter to help us improve The Rundown experience for you.

Login or Subscribe to participate in polls.

See you soon,

Rowan, Joey, Zach, Alvaro, and Jason—The Rundown’s editorial team

Reply

or to participate.