Google DeepMind dismantles the AlphaFold team; three core researchers leave for Anthropic

Google DeepMind has broken up the team behind AlphaFold — its Nobel Prize-winning protein-structure system — reassigning most of the original paper authors over the past year, with close to a quarter having left the company outright. Among the departures: Nobel laureate John Jumper and researchers Jonas Adler and Alexander Pritzel, all three now at Anthropic. Remaining staff have moved onto Gemini-related work, enzyme design, fusion, genomics, and Isomorphic Labs; DeepMind frames the change as a strategic shift away from single grand-challenge science toward Gemini-powered tools for scientists. The loss of AlphaFold's core talent to a direct competitor is the sharper signal underneath the official framing.

This gist is based on secondary reporting (The Decoder, citing the original Financial Times story), which is paywalled and could not be fetched directly this session.

Document-borne AI worms can self-propagate through Copilot for Word

Security researcher Håkon Måløy demonstrated a prompt-injection technique that turns a Word document into a self-replicating carrier: hidden instructions embedded in a source document are read by Copilot as if they were user requests, and Copilot then copies those same instructions into the document it produces. When the next person runs Copilot over that output, the payload fires again and spreads — no further attacker involvement needed. It is a responsibly disclosed research demonstration, not an observed in-the-wild attack; Microsoft was notified 144 days ago, and per the writeup there is still no mitigation covering the full class of attack.

Moonshot adds K3-256k, a half-quota 256K-context variant of Kimi K3

Moonshot's Kimi Code docs now list K3-256k, a 256,000-token-context version of Kimi K3 that consumes about half the quota of full K3 (which offers up to 1M-token context for higher tiers). It is aimed at everyday coding — Q&A, completion, and single- or small-file edits — rather than large-context work, and is available to Moderato-tier subscribers and above through the Kimi Code CLI, the VS Code extension, and the API via OpenAI- or Anthropic-compatible protocols. No dollar pricing or launch date is stated on the page; the model takes image input but not video, with low/high/max reasoning-effort settings.

TurboFieldfare runs Gemma 4 26B in ~2 GB RAM by streaming experts from SSD

TurboFieldfare is an open-source (Apache 2.0) Swift/Metal inference engine that runs the 4-bit-quantized Gemma 4 26B-A4B-IT mixture-of-experts model in about 2 GB of RAM on Apple Silicon. It keeps only a small shared core (~1.35 GB) plus the KV cache resident and streams each token's required experts from SSD on demand, using a 16-slot least-frequently-used cache and bounded parallel disk reads to hide latency. Measured throughput is 5.1-6.3 tokens/sec on an 8 GB M2 MacBook Air and 31-35 tokens/sec on a 24 GB M5 Pro. It is limited to this one text-only model, needs macOS 26 with Metal 4, and can run only one instance at a time; weights (~14.3 GB on disk) stream separately from Hugging Face under their own license.

Cryptographer Matthew Green: Anthropic's crypto results don't threaten deployed systems

Following yesterday's news that a Claude model produced two cryptanalysis results — a key-recovery attack on the HAWK post-quantum signature scheme and an improved attack on 7-round AES — cryptographer Matthew Green published an independent read. His conclusion: neither result threatens real-world systems. HAWK was never deployed (it was a post-quantum standard candidate), the attack only halves its security bits while remaining exponential-time and was run against a deliberately weakened instance, and the AES result (~2^89 operations, ~2^105 chosen plaintexts) is a modest, impractical refinement of a 2013 finding using known techniques. Green sees value in the timing, since cryptography is mid-transition to post-quantum standards, but warns that AI-generated cryptanalysis often looks correct while hiding subtle errors and still needs expert verification.

Google launches Lyria 3.5 in Flow Music

Google DeepMind released Lyria 3.5, an update to its music-generation model, inside Google Flow Music on 2026-07-29. Google describes richer and more natural melodic structure, higher-quality lyrics with better prompt adherence, more realistic and emotionally nuanced vocals with improved pronunciation, and more direct control over tempo and output duration. The announcement is qualitative only: it names no benchmark scores, no pricing or subscription tier, and no regional rollout details, so how it compares to Lyria 3 on any measurable basis and who can access it are not established by the source.

Self-hosting a frontier-class open-weight coding model still needs an 8-GPU node

An imec AI-stack post weighs self-hosting large open-weight coding models against renting GPUs or using commercial APIs, measured on an internal 64-task coding benchmark. A model small enough to fit on a single GPU solves roughly a third of the tasks; matching frontier performance (around 40 of 64) requires the biggest open-weight model running on an 8×B200 node, and even then only a couple of parallel sessions at a time. The post's cost takeaway is that an owned box, priced by the hours it actually spends working, lands in the same ballpark as renting equivalent hardware.

The circulating headline figure — "20% more hardware cost, 20% better task resolution" — could not be verified: the article text as fetched does not contain those numbers, and the readable cost claim is the "same ballpark as renting" line above. Treat the 20/20 framing as unconfirmed.