DeepMind's Great Reset, Jeff Dean's Auto-Science Spinout, and Meta Joins the Coding-Agent Race

Google DeepMind's Leadership Reset

The headline of the day: Demis Hassabis is stepping back from running Google DeepMind day-to-day to become its Chair and Chief Scientist of Alphabet, refocusing on long-term AGI and AI-for-science. Koray Kavukcuoglu takes operational control as SVP, overseeing Gemini, frontier research and product teams. The industry read is that this is both a governance reset and a push to sharpen how quickly Google ships Gemini.

The Discovery Loop Spinout

On the very same day, four of Google's most influential engineers — Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Quoc Le — announced Discovery Loop, a public benefit corporation aimed at automating machine learning, science and engineering. Backed by Radical Ventures, Khosla Ventures and even Alphabet, it's a strong signal that automated scientific discovery is becoming a primary frontier rather than a side quest.

Meta Enters the Coding-Agent Race

Meta shipped Muse Spark 1.2, a coding-focused model, alongside Muse Code, its first serious terminal agent — deliberately co-trained together for better first-attempt tool use and less reprompting. With roughly 82.9% on Terminal-Bench 2.1, an Intelligence Index score around 54, and competitive pricing near $1.25/$4.25 per 1M tokens, Meta is now firmly in the coding-agent conversation alongside the established players.

Open Harnesses Take Centre Stage

Prime Intellect released Prime Agent, an open-source harness built around a single persistent Python REPL, where the agent writes its own tools and spawns sub-agents programmatically instead of picking from a fixed menu. A related research finding drove the theme home: swapping only the harness, with the same underlying model, moved accuracy by more than 15 points.

Tiny Models, Big Reach

Liquid AI's LFM2.5-2.6B runs at roughly 30 tok/s on a phone, with a 128K context window and tool calling — pushing capable, private, offline AI onto everyday edge devices, even if testers note it still trails slightly larger models on hard reasoning.

Efficiency Wins for Local AI

A proposed llama.cpp change caches frequently used "hot" mixture-of-experts on the GPU, lifting throughput from about 33 to 56 tok/s on a modest 8GB card, though results vary by model. Separately, Qwen3-TTS voice cloning landed in mainline llama.cpp via llama-tts, clocking around 8x real-time on a high-end GPU from just a few seconds of reference audio.

AI Reviewing AI

A controlled study found that a stronger model reviewing a weaker one's code lifted the pass rate from 71.6% to 89.7% — but flipping the roles, with a weaker reviewer checking a stronger model, actually dragged the score down. It's a useful cautionary tale for anyone chaining models together: who reviews whom genuinely matters.