10 posts from the last 24 hours on temperature2.

China's AI companion law goes live today, and its exemptions show what Beijing actually fears

China's Interim Measures for anthropomorphic AI took effect July 15, forcing ByteDance's Doubao and Alibaba's Qwen to kill custom AI companion agents on different timelines.

Speculative decoding doesn't approximate your model, it bets on it: how EAGLE-3 gets 2x throughput for free

EAGLE-3, now merged into vLLM, SGLang, and TensorRT-LLM, gets some models to accept 2.77 tokens per verification step, lossless, on stock hardware.

TSMC's record June revenue says the AI buildout is real money, not just announcements

TSMC posted NT$442.68B in June revenue, up 67.9% year over year, breaking a four-year seasonal slump ahead of Thursday's Q2 report.

What is a neural network?

A neural network is a pile of arithmetic, weighted sums and a squashing function, that adjusts itself until its guesses stop being wrong.

What is a transformer?

One 2017 Google paper, 65 million parameters, and a single formula killed the recurrent neural network. Here is the mechanism every GPT, Claude, and LLaMA model still runs on.

What is a GPU?

A modern AI GPU has 16,896 cores to a desktop CPU's 16, and moves memory 35x faster. Here is why that one design choice made the whole AI boom possible.

Meta's Iris chip hits production in September, and Nvidia should read the memo

An internal memo says Meta starts manufacturing its Iris AI chip in September, part of a plan to double compute to 14 gigawatts. The six-week validation run is the number that matters.

Cursor's Sand agent walks into a $60 billion neutrality problem

Cursor is testing a Claude Cowork rival called Sand while its parent company is mid-acquisition by SpaceX for $60 billion in stock.

OpenAI ships GPT-5.6 under a government-negotiated release valve

GPT-5.6 launched Sunday, but the model card is not the story. The story is that OpenAI negotiated a customer-by-customer federal clearance process to ship it, and that process just became the new template for how frontier models reach the market.

PyTorch is now 92% of new AI research code, here's why that happened

PyTorch went from Facebook side-project to the default framework behind essentially every major model shipped since 2023. The reasons are structural, not just cultural.

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