AI · Sunday, 2 August 2026 · Issue 68 · 8 stories

Self-improving AI 5.6-sol, DeepMind poach Windsurf

Research

AI model 5.6-sol shows recursive self-improvement in scientific research

A recent LessWrong post details hands-on experiments where the GPT-style model 5.6-sol was tasked with reading and extending three of the author’s own data-science papers. The model generated plausible extensions and demonstrated a nascent form of recursive self-improvement, suggesting AI could autonomously push scientific frontiers. This early evidence challenges mathematicians’ worries and points to rapid progress in AI-for-science.

lesswrong.com ↗14 min read at source

Study Finds AI Tools Slow Veteran Open-Source Developers by 19%

A randomized controlled trial with 16 seasoned open-source contributors tackling 246 real issues showed they took 19% longer when using AI tools like Cursor Pro with Claude 3.5/3.7. Developers had anticipated a 24% speedup and still reported a perceived boost after the slowdown. The result challenges the belief that generative AI automatically accelerates software development.

metr.org ↗9 min read at source

Constitutional Midtraining Yields Durable Alignment Gains at No Capability Cost

Training a 120‑billion‑token LLM on a 394‑M‑token constitutional corpus improves alignment durability, cutting blackmail propensity by 17.5 percentage points even after fine‑tuning, while leaving benchmarks like MMLU and GSM8K unchanged. The benefit stems from simple content presence, not complex methods, and the code and models are open‑source.

arxiv.org ↗2 min read at source

LLM rankings tumble when judged by other LLMs in MUD benchmark

Researchers ran CrucibleBench, a 12‑room MUD where 13 LLMs pursued social goals, and let a Gemini Flash Lite classifier judge their behavior. Rankings shifted up to six places when the judge was ablated, while overall Cohen’s κ was a meaningless 0.04, showing LLM‑as‑judge metrics can hide major failures. The team urges benchmarks to publish per‑subject agreement and judge‑ablation analyses.

lesswrong.com ↗6 min read at source

FlexOlmo lets data owners train LLMs without ever sharing raw data

FlexOlmo introduces a modular training scheme where each contributor trains an expert on their own data, then plugs it into a shared language model. Contributors can activate or deactivate their modules at any time, retain attribution, and keep raw data private. This could unlock federated AI for sectors like healthcare, government, and finance.

allenai.org ↗5 min read at source

Products & Industry

Google DeepMind poaches Windsurf leadership after $3B OpenAI deal collapses

OpenAI’s $3 billion bid to acquire AI‑coding startup Windsurf collapsed, and Google DeepMind swooped in, hiring CEO Varun Mohan, co‑founder Douglas Chen and key researchers. Google also secured a non‑exclusive license to Windsurf’s technology for $2.4 billion, bolstering its coding‑agent ambitions without taking equity.

techcrunch.com ↗3 min read at source

Policy & Safety

Why Mathematicians Must Grasp AI Existential Risk Now

The essay translates AI alignment challenges into formal math language, giving mathematicians a concrete entry point to tackle the open problems that underlie AI risk. It argues that solving alignment needs simultaneous breakthroughs in utility specification, decision theory, and recursive self‑improvement, making the math community essential to averting a high‑stakes catastrophe.

lesswrong.com ↗2 min read at source

Tools & Open Source

ScreenEnv gives AI agents Docker‑based full‑desktop access in seconds

ScreenEnv is a Python library that spins up a full Ubuntu desktop inside a Docker container, giving agents mouse‑, keyboard‑, and window‑level control over real GUI applications. It offers a direct Sandbox API and Model Context Protocol, letting you script tasks or have an LLM control the desktop, all from a one‑line install with no VM overhead.

huggingface.co ↗5 min read at source

You’re caught up.

That was all 8 stories in the 2 August issue.