Fetching from the wire…
01
02
03
04
05
06
07
08
Policy2026-09-15 · source-backed
Socher detailed Recursive on Latent Space September 14: eight co-founders including Jeff Clune, Tim Rocktäschel and Alexey Dosovitskiy, valued near $5B (Latent Space). Its agent reached 0.937 bits per byte on NanoChat/NanoGPT in under two days, beating all human researchers and prior agents, topped SOL-ExecBench GPU kernel leaderboards without deep CUDA expertise, and found 30 bugs in the evaluation harnesses along the way. Socher's builder-facing claim is the one I'd keep: harness optimization and reward engineering beat retraining, and simple reward specs get hacked.
Each link below shares sources, entities, or timing with this story.
Astra can take an experimental idea, implement it in OpenAI's codebase, run the experiment and report results, and can follow up on a paper in work that used to take a human researcher about a week. Altman told TIME the company is "not quite yet" at AGI but will declare it int...
Within days: Claude Code capped concurrent subagents at 20 and disabled nested spawning by default, Cursor exposed a subagentStart hook, and Claude Code fixed --max-budget-usd not applying to background subagents (changelog). Three independent responses to the same recognition...
The Frontier AEO Tracker ran 6 prompt variations across 7 models over 161 product categories with every prompt-answer pair publicly inspectable. Models heavily favor their own ecosystem. 28 of 161 categories return a unanimous first choice across all models. Source consumption...
The claim covers selected attention and MoE blocks, alongside a claim that model-assisted kernel work brought three previously unplanned open-weight models to high performance on the new chip in about two months (Latent Space). Treat the numbers as first-party until the full H...
Ryan Lopopolo from OpenAI Frontier went on the Latent Space podcast and described something I've been circling around for months. His team of three engineers built Symphony, OpenAI's internal orchestration layer, as a million-line Elixir codebase. Not one line was written by a...
For about a year, "run your agent locally" meant accepting a model that couldn't reliably call a tool twice in a row. That excuse is gone. Meta Superintelligence Labs published Muse Glimmer today: a 29.6B dense causal transformer, 52 layers, 6,656 hidden dim, with a ~1.8B ViT-...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.