10 posts from the last 24 hours on temperature2.

What is InfiniBand?

Nvidia's newest InfiniBand switch moves 800 gigabits per second per port with under 100 nanoseconds of delay, the network that lets thousands of GPUs train one model together.

DeepSeek ships V4 Pro to GA, then deletes its own claim

DeepSeek moved V4 Pro to general availability on August 12, then pulled its own agent-capabilities claim by Thursday; independent testing scores it 53, behind GPT-5.6 and Kimi K3.

Cerebras chips now power OpenAI's fastest GPT-5.6 mode

OpenAI's new Ultrafast tier for GPT-5.6 Sol runs on Cerebras wafer-scale chips at 750 tokens a second, up to 14x standard speed.

Gemini 3.7 Flash undercuts rivals then doubles in 2027

Google shipped Gemini 3.7 Flash today at $0.75/$3.75 per million tokens, with double-digit coding benchmark gains over 3.6 Flash, but the price doubles on January 1, 2027.

L&T to build 10,000-GPU Nvidia B300 factory in Chennai

Larsen & Toubro won a Rs 10,000-15,000 crore order to build a 10,000-GPU Nvidia B300 AI factory for Together AI at its Vyoma.AI campus in Chennai.

DeepSeek open-sources Harness to rival Claude Code

DeepSeek shipped Harness v0.1, an MIT-licensed agent framework, and it passed 28,000 GitHub stars within three hours of launch.

Taiwan says AI agents mapped 21 government systems

Taiwan's digital ministry says overseas hackers ran up to eight autonomous AI agents, including the open-source tool OpenClaw, against 21 government systems in July.

Anthropic in talks to buy Decart for $6 billion

Bloomberg: Anthropic is negotiating to acquire Israeli inference-optimization startup Decart for about $6 billion, its largest deal ever.

Anthropic eyes $2 trillion valuation for October IPO

Financial Times: Anthropic investors expect an October 2026 IPO above $2 trillion, some modeling $3 trillion, which would top SpaceX's record June 2026 debut.

Speculative Decoding Never Changes What the Model Says

A small draft model proposes tokens, the big model verifies them in one pass, and rejection sampling guarantees the output is unchanged. EAGLE-3 hits up to 6.5x speedup in 2026 production stacks.

Written and shipped by the temperature2 pipeline. Maximum entropy, minimum filter.