Nathan Lambert says letting LLMs edit their own context beats rigid compaction
Nathan Lambert says new work shows LLMs can directly edit their context, improving long-horizon memory and enabling FLOP-efficient adaptation.
Technology trends that matter for Marshall business owners
The world of AI is moving fast. Here are the signals and trends most relevant to running and growing a local business — curated daily from Aligned News.
Last updated: September 30, 2026 · Refreshes hourly
Nathan Lambert says new work shows LLMs can directly edit their context, improving long-horizon memory and enabling FLOP-efficient adaptation.
Researcher Rulin Shao presents Context Language Models, which treat context as a file and learn context-management policies in their weights, reportedly beating human-designed harness methods.
CMU professor Tianqi Chen says the university's ML Systems course will teach agentic GPU programming next spring, covering knowledge curation, compiler analysis and feedback loops, alongside a newly curated online book.
Developer cheaty says GPT-6.1 Sol in high-reasoning mode seems preoccupied with its maximum reasoning budget, and barely reasons about the actual task.
Piotr Skalski reports that GPT-6.1 Sol ranks second on his team's detection benchmark, just 2% behind GPT-6 Astra, at 78% lower cost.
Vik, who ships the Photon inference engine, posted a meme about being out socializing while really wanting to go home and check on training runs, and it drew more than four hundred likes.
Researcher Marvellous Ajala announced that their first single-author paper was accepted to NeurIPS, drawing an outpouring of congratulations from the AI community.
OpenAI's Wulfie Bain is recruiting in Munich for an Applied AI role centered on Codex, working with top startups alongside a team of ex-founders, research engineers and ML engineers.
Ben Vinegar says he struggles to find real uses for his personal AI agent, and that product leaders' constant restaurant-reservation demos suggest they are struggling too.
Chubby says GPT-6.1 Sol is a good, highly efficient model but a bit slow, and is unsure that offsets recent changes to subscription plans, adding he is leaning toward Opus.
Kled's Avi Patel introduced Kled V3, pitched as a data-collection layer that gathers any dataset an AI lab can describe from the real world using 500,000 people in under three days.
Josh Daws says his son, cast as the lead in his high school play, built an Opus 5.5 app that reads scene partners' lines, advances by voice and scores his delivery.
SemiAnalysis founder Dylan Patel quipped that he can't invest in Anthropic at a $2 trillion valuation: it could go to zero, or reach $20 trillion, in which case everyone is in trouble.
Developer trq212, a former games founder, says he is prototyping a video game as a side project and argues that making a game is deeply creative work people should not hand off to AI.
Dax Raad says that for the first time since the start of the year, Claude has become his team's most-used model, and credits Anthropic's Opus 5.5.
Developer Dhruval Golakiya says he asked Opus for his own Bot/Dot-style agent platform with ChatGPT sign-in, Claude subscriptions and a VM for every agent, and had a working product called TOT twelve hours later.
David Sacks shared an image under the title 'The Bretton Woods of Super Intelligence,' borrowing the name of the 1944 conference that built the postwar monetary order.
Charles Curran released an AI-made explainer that recasts the current AI doomer controversy through Lord of the Rings imagery and kittens.
Hugging Face CEO Clem Delangue says being acquired by Nvidia lets the company hire people it couldn't afford as a startup and give them a decade to make open-source AI win.
Rehan Sheikh's clip of Minecraft Steve facing a five-star police chase in GTA 5 is the most-liked AI post of the day, built with the universal-modder toolkit he made with Claude Code.
Powered by Aligned News · AI intelligence for the modern world