Daily Brief
2026-09-04 · Evening · 8 items
-
OpenAI's GPT-6 Astra tops ARC-AGI-3 with dual SOTA scores
One model set SOTA on both harnesses at 62.7% and 99.9%, with the cheaper run costing just $19K.Hacker News 热门（buzzing.cc 中文翻译）
-
Greg retweets: GPT-6 Astra hits SOTA on ARC-AGI-3, benchmark called saturated
A new harness pushed Astra's score from 63% to 99%, rendering the original setup unable to separate model capability.X：Greg Brockman (@gdb)
-
Perplexity to integrate GPT-6 Astra, calling it a broad lead on wide and deep research
Perplexity says Astra beats rivals on cost-effectiveness for broad and deep research, rolling out soon to Pro and Max users.X：Aravind Srinivas（Perplexity CEO） (@AravSrinivas)
-
xAI Opens Grok Bot for Enterprise, Free for Two Weeks on Grok and Cursor
Even employees without existing seats are invited, giving the whole org a two-week free trial.xAI：News（网页）
-
OpenAI launches Astra model focused on computer and browser control
It rolls out first to cybersecurity customers, then reaches all paid tiers and the API within a week.TechCrunch：AI（RSS）
-
Google DeepMind Releases WeatherNext 3 Global Weather AI ModelRECAP
Brightband's independent real-time evaluation confirms WeatherNext 3 as the most advanced global weather AI model.Google DeepMind：Blog（RSS）
-
Hugging Face Open-Sources RL Pipeline for Watercolor Drawing with TRL and OpenEnvRECAP
All components – TRL, OpenEnv, Qwen3.5-35B-A3B – are open-sourced for replicating the RL workflow that trains coding models to paint watercolors.Hugging Face：Blog（RSS）
-
Hugging Face Launches funes: Local Memory Layer for Coding AgentsRECAP
A single funes add command lets coding agents autonomously retrieve original context including timestamps and session turns.Hugging Face：Blog（RSS）