AI digest: Hardware wars heat up
Taalas ditches GPUs for hardwired AI chips, Pentagon summons Anthropic CEO, and ByteDance cracks the reasoning code.
This week’s big theme is specialization. Everyone’s realizing that general-purpose everything might not be the answer.
Taalas wants to kill the GPU with 17,000 token/sec chips
Toronto startup Taalas is ditching programmable GPUs for hardwired AI chips that hit 17,000 tokens per second. Their bet is that flexibility is actually holding AI back, and custom silicon beats general-purpose every time. Bold move, but if they’re right, this could reshape the entire inference game.
Pentagon calls Anthropic CEO to the principal’s office
Defense Secretary Pete Hegseth summoned Anthropic CEO Dario Amodei for a tense chat about Claude’s military use. Hegseth is threatening to label Anthropic a “supply chain risk.” This feels like the opening shot in a much bigger battle over AI companies working with the military.
ByteDance maps AI reasoning like molecular bonds
ByteDance researchers figured out why long chain-of-thought reasoning falls apart and how to fix it. They’re mapping reasoning patterns like molecular bonds instead of just copying keywords. Smart approach that could finally make multi-step reasoning reliable.
Google says think harder, not longer
New Google research shows that longer chain-of-thought isn’t better, deeper thinking is. Their “deep-thinking ratio” improves accuracy while cutting inference costs in half. Makes sense - quality over quantity has always been true for human thinking too.