AMD acquires AI inference chip startup Taalas, which hardwires models directly into silicon
AMD signed a definitive agreement on 6 August to acquire Taalas, a startup whose chips etch a specific AI model's weights directly into custom silicon rather than running general-purpose inference hardware, cutting compute and memory bottlenecks for that one model. Financial terms were not disclosed. AMD plans to fold the technology into its accelerator roadmap and combine it with its Instinct GPU line as part of a full-stack AI inference push.
Why it matters: Locking a model into hardware is a real tradeoff (you lose flexibility to switch models) for a real gain (much faster, cheaper inference on high-volume workloads) — AMD betting on that tradeoff signals where it thinks the next inference cost war gets won.
Legal AI startup Harvey in talks to raise $500M at a $15.5B valuation, more than 5x its value 18 months ago
Harvey is in talks to raise at least $500 million at a $15.5 billion valuation, a 40% jump from its $11 billion valuation five months ago, with Lightspeed Venture Partners reportedly interested in leading. The company says it's now generating more than $350 million in annualized revenue, up from roughly $190 million ARR at the end of 2025. Existing backers include Sequoia, Andreessen Horowitz, GIC, Coatue and Kleiner Perkins.
Why it matters: Going from a $3 billion valuation in February 2025 to a proposed $15.5 billion now is one of the fastest re-ratings in this AI cycle, and it's happening in a vertical (legal) that was supposed to be one of the harder ones for AI to actually monetize.
Anthropic loosens Fable 5's biology safeguards, cutting unnecessary blocked queries by about 85%
Anthropic rewrote the "constitution" governing what counts as safeguarded biology content in Claude Fable 5 and retrained the classifier on new data after consulting internal and external experts, cutting biology-related false-positive blocks by roughly 85% in testing. Everyday questions like interpreting lab results or learning biology should now go through cleanly; virology, toxicology and molecular design stay blocked, and Anthropic is developing a separate "trusted access pathway" for vetted researchers.
Why it matters: This is a safety team publicly admitting its own classifier was overblocking legitimate use for months and fixing it with a formal methodology, which is a more useful signal about how Anthropic actually tunes safety tradeoffs than most safety announcements.
xAI ships Grok 4.6, betting on post-training gains over raw scale to compete with Kimi K3 and Claude Opus 4.8
Grok 4.6 launched 7 August as a 1.5-trillion-parameter model built on the same foundation as Grok 4.5, with xAI improving supervised fine-tuning and reinforcement learning rather than scaling up parameters. It shipped as closed-weights, available via the xAI API, the Grok app, grok.com, and X's Premium+/SuperGrok tiers. A larger 2.1-trillion-parameter Grok 4.7 is expected within weeks.
Why it matters: xAI explicitly choosing post-training over scale to compete with much larger models (Kimi K3 at ~2.8T params) is a bet that the scaling-law era of "bigger wins" is genuinely over for the current generation.
Meta ran AI-generated child sexual abuse ads for nine months after its own review process approved them
An investigation by Wired and the Tech Transparency Project found more than 50 paid ads containing AI-generated child sexual abuse material ran across Facebook, Instagram, Messenger and Threads between November 2025 and August 2026, some directing users to AI "nudify" apps. TTP director Katie Paul said these weren't posts that slipped past moderation but ads Meta's own team reviewed, approved and ran. Meta has since removed them and says most had minimal reach.
Why it matters: The failure here isn't detection at scale, it's that Meta's paid-ad review process — a smaller, theoretically more scrutinized queue than open posting — actively cleared AI-generated CSAM, which is a different and worse failure mode than content slipping past an automated filter.
AI Impact Hub