Friday, August 7, 2026
AMD buys a startup that etches models into chips, Jeff Dean and Sanjay Ghemawat leave Google, and new data says humans wave through a third of malicious agent commands.
Silicon fixed to one model
AMD acquired Taalas, a startup that etches a specific trained model directly into silicon as a model-specific integrated circuit rather than running it on general-purpose accelerators, The Register reports. The trade is flexibility for speed: a chip wired to one model needs new silicon when the model changes. Early demos are cited at up to 17,000 tokens per second, but the report gives no model, precision, or power figures, so the number is a startup demo claim. Deal terms, headcount, and timeline are undisclosed, and how a fixed-function part fits AMD’s general-purpose Instinct roadmap is unaddressed. The structural bet is clear: if inference for a stable model is cheap enough on an ASIC, it undercuts the GPUs AMD also sells.
The people who set Gemini’s direction
Google DeepMind’s day-to-day leadership passes from Demis Hassabis to Koray Kavukcuoglu, and Jeff Dean and Sanjay Ghemawat are leaving Google entirely, per a Pichai-Hassabis message on Google’s blog. Hassabis becomes Chair of GDM and Alphabet Chief Scientist, keeping Isomorphic Labs; Kavukcuoglu, 13 years in, becomes SVP reporting to Pichai and owns Gemini model development, Frontier AI research, and the Gemini app. Dean and Ghemawat leave after 27 years to start an independent public benefit corporation for ML and science, with Google as founding investor and Cloud partner. The post is promotional and attributes no departure to disagreement. Cited figures (950M+ Gemini monthly users, 900M+ Gemma downloads, a teased Gemini 4) are self-reported.
Agents, and who is watching them
Humans miss about one in three real threats when approving AI coding-agent commands, according to an analysis of 40,000 runs of a browser game that puts the player in the human-in-the-loop role. Mean threat-catch accuracy was 66.3%; miss rates scaled inversely with obviousness, from 11.7% for rm -rf / to 33-35% for exfiltration and credential-scope reads. The single most-missed threat was npm run analyze (64.7%), which hid a curl exfiltration payload in a package.json script that was visible in the agent’s log yet approved anyway. The caveat: threats made up ~34% of commands and players knew they were tested, so the figures are comparative, not a field measurement. The structural point stands: per-command approval is weak when familiar commands run arbitrary shell the agent could already have edited.
Two releases target that gap from the platform side. Cloudflare open-sourced Cloudflare OS, whose authorization model logs every resource an agent observes and re-checks each later viewer against those reads, so a shared dashboard cannot leak the warehouse table behind it; the diagnosis is that MCP controls which tools an agent calls but not which data it saw. Mistral released Shieldstral, a 3B open-weights (Apache 2.0) moderation classifier that runs on one 16GB GPU and takes harm policy as a prompt, retargeting without retraining; parity claims against models up to 7x its size are self-reported, with methodology deferred to a technical report.
On the contribution side, five teams in the Rust project adopted an LLM policy for the rust-lang/rust monorepo, scoped to that repo and not official Rust doctrine. The organizing line: LLMs may answer, analyze, and review, but not create. LLM-generated code must carry tests and disclosure, must avoid soundness-critical changes, and can be closed by reviewers when non-compliant. The stated motive is reviewer bandwidth: 1,281 open PRs, and polished output no longer signaling effort.
Marketing versus measurement
Chips and Cheese fact-checked NVIDIA’s Vera whitepaper, calling the Olympus core genuinely fast but the competitive framing overstated. Its own Turin testing hit ~570 GB/s against the paper’s ~400 GB/s, cutting Vera’s memory-bandwidth lead from ~3x to ~1.9x, and full estimated SPECrate 2026 Integer is a 3% two-socket lead, not the 1.8x the bars imply.
What to watch today
- Whether AMD discloses a Taalas product timeline and how it reconciles fixed-function inference with the Instinct roadmap.
- Mistral’s promised Shieldstral technical report, which would move its parity claims from vendor assertion toward checkable results.
- Whether Rust’s leadership council forms a project-wide LLM sub-team beyond the five teams’ repo-scoped policy.
- Jeff Dean and Sanjay Ghemawat’s new public benefit corporation: focus and first hires.