THE NEWS

🤖 Google’s AlphaGeometry is outperforming Mathematics Olympiad medalists

Mathematics is on the brink of a new era, where AI might soon outpace human capability in solving complex problems.

Leading this charge is Google's AlphaGeometry, a new AI that is capable of solving really hard geometry problems from the Mathematics Olympiad, almost at the level of gold medalists.

Here's the scoop: Google's team has created an ingenious two-part system. The first part is what they call the 'engine': think of it as the problem solver. When it sees an equation it knows what to do.

The second part is the creative genius, a language model created to come up with different ways to solve geometry puzzles.

This language model comes up with diverse solutions for the engine to test. If a solution doesn't hit the mark, they regroup and try another approach. It's a dynamic duo of creative thinking and methodical problem-solving, delivering some truly impressive results. In the standard 4-hour Olympiad timeframe, this system almost matches the problem-solving success rate of the gold medalists.

Small but feisty: the language model is surprisingly small-scale. It's got just 151 million parameters – far less than typical language models, which often exceed 100 billion parameters. Plus, it's been trained entirely on synthetic data, which is basically computer-made stuff. This is pretty groundbreaking and shakes up the whole conversation in the AI world about whether you can really rely on computer-made data to train these models.

It does not go for the looks: Unlike human solutions that often feature symmetry, AlphaGeometry's solutions don’t have a bias for elegance and are purely functional – they may not be pretty, but they get the job done.

The achievements of AlphaGeometry are even more remarkable considering it operates on V100 GPUs, which aren't the latest or most powerful GPUs available today. This suggests that we haven’t seen the full potential of AI yet.

In a generous move, Google plans to share AlphaGeometry with the world. They're open-sourcing the code and model, and aim to integrate it into Google’s Gemini platform, alongside AlphaCode 2, within a year.

This development is not just a win for Google but a promising advancement for the field of mathematics and AI.

ONE MORE THING

Mark Zuckerberg doubles down on AI, wants to build AGI

He announced that Meta is building out huge compute capabilities (they spent $10.5billion on NVIDIA GPU’s), that they want an open-source AGI and are already training their next model, LLama 3.

This is great news! Meta is one of the biggest driving force behind open-source AI development. The AI community desperately needs more resources to balance the scales, preventing just a handful of companies from holding all the power in this field.

⌚ If You Have One More Minute

• 🐇 Rabbit Inc and Perplexity partner up to introduce real-time, precise answers to Rabbit R1

• 🇺🇸 OpenAI’s Altman says U.S. and AI will be ‘fine’ no matter who wins White House after Trump’s Iowa landslide

• ⚖️ A New Nonprofit Is Seeking to Solve the AI Copyright Problem

AI Art of the Day 🎨

Krea AI just rolled out three cool new features: text-to-image, background removal, and an eraser tool. And guess what?

They're all accessible for free account users. Feel free to experiment with them here.

It's amazing how effortlessly we can now create any image we imagine.

🏄🌊🏄🌊🏄🌊🏄🌊🏄🌊🏄🌊🏄🌊🏄🌊🌊🏄🌊🏄🌊🏄

That's all for today, folks!

If you enjoyed this, please consider sharing this hand-crafted newsletter with a friend.