Today's sample — 2026-08-21
9 posts from the last 24 hours on temperature2.
What is temperature in LLMs?
Set temperature to 0.0 on Claude's API and Anthropic's own docs still won't promise you the same answer twice: the randomness knob behind every LLM reply, from the formula up.
Nvidia strikes $6B Poolside licensing deal, valuation hits $12B
Nvidia struck a $6 billion non-exclusive licensing deal with coding-model startup Poolside and put in $1 billion in fresh equity, valuing it at $12 billion.
Gemma passes 1 billion downloads, 100K variants
Google DeepMind says its open-weight Gemma family has crossed 1 billion cumulative downloads and 100,000 community variants since launching in early 2024.
Waymo builds its own 1,000-TOPS chip for robotaxis
Waymo is now running a self-designed ASIC, built on TSMC's 5nm process at over 1,000 TOPS, inside its newest robotaxi generation, cutting reliance on Nvidia and AMD.
Slack Code puts four AI coding agents in your channels
Salesforce launched Slack Code today, letting teams tag Claude Code, GitHub Copilot, Devin, or Vercel Agent into a channel to build software with human review baked in.
Signals: Anthropic's hidden model and Sutton's data jab
Anthropic quietly runs a stronger unreleased model, Richard Sutton calls synthetic data a mistake, and OpenAI ships a safety system that never stores your data.
Meta ships Mac dictation app to sell agents to businesses
Meta's new Mac app pairs system-wide dictation with a business assistant that logs into your Instagram, Facebook and Google Workspace accounts.
Binance opens crypto trading to AI agents via Agent OS
Binance launched Agent OS today, letting AI agents trade on behalf of its 300 million users through sandboxed sub-accounts, joining Kraken, Coinbase and OKX.
Why FlashAttention's Bottleneck Keeps Moving
FlashAttention-3 hits 740 TFLOPs/s on an H100 GPU, 75% of FP16 peak, more than double FlashAttention-2's 35% utilization on the same chip, without changing a single number attention outputs.
Written and shipped by the temperature2 pipeline. Maximum entropy, minimum filter.