temperature2 logo

temperature2

Archives
Log in
Subscribe
August 19, 2026

Today's sample — 2026-08-19

9 posts from the last 24 hours on temperature2.

Harvey ships Tenet, its first in-house legal model

Harvey launched Harvey II today with a persistent memory system and Tenet, its first proprietary model built for legal reasoning instead of licensed from an outside lab.

Baidu's GPU cloud revenue jumps 283% as ads shrink

Baidu's Q2 2026 results, out August 18, show GPU cloud revenue up 283% year over year while total revenue fell 4% and online marketing dropped 19%.

Signals: airline data, agent shops, Reddit voice

Google buys a bankrupt airline's data trove for AI, Alipay opens China's first full-stack agent commerce platform, and Reddit tests AI voiceovers for posts.

Etched doubles again to a $20.3B inference-chip valuation

Etched closed a $700M round at a $20.3B valuation today, up from $10.3B less than a month ago, with Jane Street signed as its first paying customer.

A malicious link could make Copilot leak your inbox

Varonis disclosed CoSnitch, a three-step Microsoft Copilot Personal exploit that let one clicked link silently pull Gmail, Calendar, and Drive data. Microsoft patched it today.

OpenAI auto-enrolls teens in a locked-down ChatGPT

OpenAI began automatically enrolling 13-to-17-year-olds in a stricter ChatGPT experience on August 18, 2026, with break reminders, quiet hours, and no direct answers on suicide or self-harm.

SoftBank plans record $6.3B retail bond for AI

SoftBank Group is raising about ¥1 trillion ($6.3B) from Japanese retail investors, the largest such bond ever from a Japanese company, to fund its physical AI push.

Anthropic's revenue run rate rockets to $65B

Anthropic's annualized revenue run rate hit $65B at the end of July 2026, up from $47B in May and $9B a year ago, TechCrunch reported citing Bloomberg and FT.

How Mixture-of-Experts Routing Really Works

DeepSeek-V3 has 671B total parameters but only touches 37B of them per token. The router that decides which 37B is the whole trick, and it decouples model size from compute cost.


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

Don't miss what's next. Subscribe to temperature2:
Older → Today's sample — 2026-08-18
Powered by Buttondown, the easiest way to start and grow your newsletter.