AMD acquires Taalas to etch AI models into silicon, 48x faster
Why it matters: model-specific silicon is where cheap inference heads next. Also: Cursor's router beats Opus 4.8 for 41% less, a free model, agents in Slack.
Copy markdown
Silicon that memorizes one model
Taalas burns a model's weights straight into the chip instead of streaming them from memory — its HC1 test part ran Llama 3.1 8B at ~17,000 tokens/sec, 48x an Nvidia GPU and 8.5x a Cerebras. The catch: one model per chip, though a re-spin touches only two metal layers. AMD folds the tech into its Helios racks, with the deal closing Q4.
Cursor's router now beats Opus 4.8 for 41% less
Cursor's Auto Balance mode now outperforms Opus 4.8 at 41% lower cost, while Auto Intelligence clears above-Fable satisfaction for 68% less — both tuned on live production traffic, with Opus 5 newly in the routing mix. If you're on Cursor, switching to Auto is a real bill cut with no quality tradeoff.
A gateway model that's free through mid-August
Vercel's AI Gateway added Ling 3.0 Tiny from inclusionAI: bring-your-own-key, no markup on provider pricing, high rate limits, and automatic failover — free until mid-August. A zero-cost way to trial a new model behind a single gateway endpoint before you commit.
Drop any agent into Slack and Teams
CopilotKit open-sourced the Channels SDK (MIT): run any AG-UI agent — LangGraph, CrewAI, Mastra, Pydantic AI, Google ADK — inside Slack, Teams, Discord, or Telegram with native interactive UI, streaming, tool calls, and human approval. `npm i @copilotkit/channels` gives your weekend agent a front door your team already lives in.
Elsewhere: durable human-approval in one call
Vercel's Chat SDK added Durable Approvals — a single requestApproval call renders an approval card and pauses the run, and the suspension survives deploys and restarts. It's the missing primitive for agents that need a human yes before they act.