Self-improving AI, pharma moves & more (June 16, 2025)

Happy Monday,

The AI world is feeling a little Black Mirror today: Meta just cornered the data market, MIT gave AI the power to self-improve, and Claude... wrote a research paper critiquing Apple. Yes, Claude. Buckle up.

🔩 Big Shifts in AI Power

💰 Meta’s Scale Play Makes Rivals Flinch

Meta bought a 49% stake in Scale AI, the massive data labeling firm. That’s a lot of control over AI’s raw materials, and now OpenAI, Google, Microsoft, and xAI are all scrambling. OpenAI is still using Scale (for now), but the rest? Backing off fast.

A balance scale tilts slightly to the right, with a blue Meta symbol labeled 49% on the left pan and an orange "S" symbol labeled 51% on the right pan

🧮 AMD Enters the Rack Wars

AMD unveiled "Helios," a rack-scale AI system designed to rival Nvidia. It promises 40% better token-per-dollar performance and aims for 20x energy efficiency by 2030. The AI GPU wars just got a whole lot tighter.

🧾 Claude Critiques Apple…As Claude

Apple’s recent paper argued that LLMs can’t really reason. But it turns out one of the rebuttals was co-authored by Claude itself. That’s right, AI is helping debunk science about AI. Who needs peer review when you’ve got Claude?

🧪 Self-Improving AI & Biotech Moves

🧠 MIT’s AI Teaches Itself (Without You)

Researchers at MIT built SEAL, a system where language models create their own training data and refine themselves via reinforcement learning. One model jumped from 0% to 72.5% performance in puzzle tasks, all without external input.

A humanoid robot sits cross-legged on the floor of a white room, carefully assembling a complex jigsaw puzzle

💊 AstraZeneca Signs $5.3B AI Drug Deal with China’s CSPC

AstraZeneca partnered with CSPC in a multi-billion dollar deal to use AI in developing new oral medications for chronic diseases. It’s a signal of how seriously pharma is taking AI-first R&D, especially in cross-border research alliances.

🗺️ LLMs Can Geolocate Better Than You

Open-source researchers challenged 20 AI models with obscure travel photos. OpenAI’s models nailed it by analyzing architecture, vegetation, and even blurry text, often beating Google Lens. Claude, meanwhile, often guessed wrong continent.

An AI robot sits at a table in a room, curiously examining a spread of travel photos featuring mountains, cities, beaches, and famous landmarks

🛠️ Tools of the Day

Seedance – Instantly create professional AI videos
Flowstep – Turn ideas into wireframes and user flows
Instance – Build apps, games, and websites from plain text

✍️ Prompt of the Day

Business Plan Generator
Take on the role of a strategist and build a complete business plan: start with an executive summary, then cover market analysis, strategy, financials, and timelines. Use bullet points and clear headings.

⚡ Quick Hits

Google drops Scale AI after Meta’s investment
Tencent’s Hunyuan3D turns 2D images into photorealistic 3D
16% of people say they’ve had romantic chats with AI

📌 TLDR

Meta’s Scale AI deal is reshuffling the AI data game, MIT’s SEAL framework hints at self-upgrading superintelligence, and Claude is now publishing rebuttals to Apple. From infrastructure wars to biotech breakthroughs, AI just got a whole lot weirder and more powerful.

See you tomorrow,
David