AMD Buys Taalas: The Chip That Bakes Weights Into Silicon
AMD acquired a startup that hardwires LLM weights into transistors. At 17,000 tokens/sec, the real question: was the GPU always just a phase?
Explore our latest articles on AI tools, agents, product development, tutorials, and industry news.
AMD acquired a startup that hardwires LLM weights into transistors. At 17,000 tokens/sec, the real question: was the GPU always just a phase?
Jeff Dean, Ghemawat, Vinyals, and Le leave Google DeepMind to found Discovery Loop. What the exodus reveals about AI's talent war.
Moonshot dropped 2.8T params on HuggingFace. The VRAM math says almost nobody can self-host. Delta Attention is the real story.
A GrapheneOS duress PIN wiped a phone during a CBP search. The feds called it obstruction. What it means for your rights.
Vendor model routing is a cost AND data control point. Why open-source routers are the contested infrastructure play.
Moonshine and transcribe.cpp shrink speech AI to sub-megabyte, but the real shift is architectural guarantees over policy promises.
A prompt-guided GPT-5.6 attack proved an optimal lower bound in convex optimization while coverage of the same model decayed into pricing tips.
Mozilla's data shows open-weight models flipped from 5% to majority token share in two years. What the cost curve means for your inference stack.
Inkling and Kimi K3 shipped within 24 hours. Prediction markets repriced China, not Anthropic.
3 seconds of audio breaks bank voice authentication. Combined with prompt injection exfiltration, AI's attack surface is expanding faster than defenses.
MIT, Tencent, and Huawei independently published continual learning papers in 2026. Their convergence reveals AI's real bottleneck.
Tachikawa's 6-month physics problem fell to Fable in one night. GPT-5.6 claims a 50-year math proof. Polymarket is repricing it all.
Get a weekly digest of the best AI infra writing — Claude Code, agent frameworks, deployment patterns. No fluff.
WEEKLY. UNSUBSCRIBE ANYTIME.