Welcome back!

We may be watching the Kasparov vs Deep Blue moment of mathematics. This week, AI models from Google and OpenAI cracked Math Olympiad gold, using only natural language and no tools. LLMs can now handle abstract reasoning for hours, and start to scale into domains once thought untouchable. Let’s unpack…

Today’s Summary:

• 🏅Google and OpenAI win gold at Math Olympiad

• 📹 Decart streams live AI video

• 🏥 Google unveils open-source healthcare AI

• 📈 Alibaba’s Qwen3 top open source model

• 🖼️ Gemini adds image segmentation

• 🎮 New ARC-AGI-3 benchmark

• 🛠️ 2 new tools

TOP STORY

Google and OpenAI score math gold medal with new AI models

The Summary: AI models from Google DeepMind and OpenAI have both reached gold-medal performance at the International Mathematical Olympiad. For the first time, AI solved five out of six official IMO problems under real exam conditions using only natural language. Google Gemini was formally graded by IMO, while OpenAI’s model was evaluated by former medalists.

Key details:

• DeepMind’s Gemini Deep Think scored 35/42 under 4.5-hour IMO conditions, the first AI model formally certified with a gold score

• OpenAI’s unnamed model matched the score using a general-purpose reasoning engine, evaluated by three former IMO gold medalists

• Gemini solved the problems end-to-end in natural language without any formal languages, tools, or internet access

• More than scores, which are difficult to compare, the deeper milestone is that models were able to sustain several hours of abstract reasoning without added tools

Why it matters: This might be a Kasparov vs Deep Blue moment for math. Solving Math Olympiad problems requires hours-long, structured thinking under constraint. Now, two different AI systems have done it end-to-end, in natural language, without tools, within time limits. While we’re still far from automated discovery or provable rigor in high-stakes science, the feat shows how LLMs are starting to scale into domains previously thought untouchable.

DeepMind solution to the IMO 2025 problems

DECART AI

Decart AI launches realtime streaming video model

The Summary: Decart released MirageLSD, the first AI model that transforms live video feeds in real time. The web-based Mirage platform is now live. Quality is maintained indefinitely at 24 FPS and 768×432 resolution. Several innovations make its real-time performance possible.

Key details:

• Produces infinite video streams, breaking past the 20–30 second limit of earlier AI video models

• History augmentation trains on corrupted frames so the model anticipates and corrects visual drift

• Andrej Karpathy calls MirageLSD “real-time magic”, and says diffusion models can now style live feeds frame by frame

• Roadmap includes improved facial consistency, voice prompts, and precise object controls in mobile apps

Why it matters: Decart live-stream AI lets users create dynamic video effects for games, events, or creative projects. The visuals stay sharp for hours, with live prompts and instant editing. This could power AR filters, virtual production sets, and real-time visual effects on any live camera feed.

Try it here

FROM OUR PARTNERS

The Missing Tool in Your AI Stack

Create How-to Videos in Seconds with AI

Stop wasting time on repetitive explanations. Guidde’s AI creates stunning video guides in seconds—11x faster.

• Turn boring docs into visual masterpieces

• Save hours with AI-powered automation

• Share or embed your guide anywhere

How it works: Click capture on the browser extension, and Guidde auto-generates step-by-step video guides with visuals, voiceover, and a call to action.

Best part? It’s 100% free.

👉 Download the extension now

GOOGLE

Google MedGemma makes medical AI customizable and local

The Summary: Google has released MedGemma and MedSigLIP, two open-source AI models designed for real-world healthcare applications. These models are multimodal, efficient, and can run locally to interpret medical text, images, and electronic records. They can also be fine-tuned without relying on the cloud. MedGemma outperforms larger models in clinical tasks and reduces data retrieval errors in EHR systems.

Key details:

• MedGemma 27B is multimodal and scores 87.7% on MedQA at a tenth the inference cost of top-tier models

• MedSigLIP is a 400M encoder for medical image classification and search. Trained on over 33M images, including histopathology and dermatology

• Tuning the models may lead to major gains, like tripling accuracy in tissue analysis and reducing errors in patient record search

Why it matters: AI in healthcare has long been stuck between black-box cloud giants and fragile local tools. MedGemma stands out as open, local, and adaptable, letting researchers, clinicians and hospitals build and control their own AI tools. It can run without internet, and be tailored to local needs or specific rules.

Technical report

QUICK NEWS

Quick news

• Alibaba’s Qwen3-235B becomes top open-source model

• Gemini 2.5 now supports conversational image segmentation

• New ARC-AGI-3 benchmark tests how well AI learns new tasks

TOOLS

🥇 New tools

• Ash - AI support for anxiety, stress and growth

• Suno v4.5 - Make any song you can imagine