10 posts from the last 24 hours on temperature2.

Open models now serve most tokens on OpenRouter

Mozilla's first State of Open Source AI report finds open-weight models winning on cost and volume but still lagging closed models into production.

What is CUDA?

CUDA turned 20 in March 2026 with 6 million developers on it, per Nvidia's own count, and it's the software layer, not the silicon, that explains why Nvidia runs the AI boom.

Nvidia lines up 31 Japanese firms for physical AI

Jensen Huang and Japan's trade minister launched a government-backed Physical AI Initiative in Tokyo, with Toyota, FANUC, and Mizuho among 31 founding members.

What is a parameter?

GPT-1 had 117 million parameters in 2018. GPT-3 had 175 billion in 2020. Here is what that number actually is, and why bigger isn't automatically smarter.

Signals: a leaky CLI, a Rust rewrite, and one philosophy essay

xAI open-sources Grok Build after an SSH-key leak, Bun's Zig-to-Rust port ran through Claude Code for $165K, plus a serving paper and an essay on what agents cost teams.

DeepSeek's valuation triples to $70B in three months

DeepSeek is in talks for a fresh round at roughly $70 billion, barely a month after its first outside raise ever.

Gemini 3.5 Pro misses its own hype date

Trackers built up July 17 as Gemini 3.5 Pro's GA day. Google's own blog still runs the same 'rolling out next month' line it posted on May 20.

Kimi K3 triples its price and still beats Opus 4.8

Moonshot's Kimi K3 jumped to $3/$15 per million tokens, 3x its predecessor, and still undercuts GPT-5.6 Sol on cost per task while beating Claude Opus 4.8 on benchmarks.

Mira Murati's Thinking Machines ships its first open model

Inkling is a 975B-parameter mixture-of-experts model with 41B active, open for fine-tuning, and it's the first model Thinking Machines has released outside its Tinker API.

How VLA models turn language into robot motion

Physical Intelligence's pi-0 outputs 50 actions per forward pass at 50Hz; RT-2 outputs one action token at a time like a chatbot. That single design choice decides which tasks a robot can actually do.

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