Topic
A concept in AI research where systems improve their own algorithms or capabilities with little human input, raising safety concerns.
Google Cloud AI Research's RRSI lifted agent scores up to 4.7 points on five unseen benchmarks by capping how far a harness can rewrite itself. Its guardrails cost points on the tuning tasks, a trade worth making for teams that need harness gains to hold on new work.
Reality
Nathan Lambert and Tom Zick launched Trillium Labs, a nonprofit that will publish AI experiments, self-improvement work included, for outsiders to replicate. Their bet is that outside scrutiny will find and limit frontier-model risks better than keeping models locked inside labs.
Reality
OpenAI and DeepMind researchers warned in videos given to Reuters that labs are rushing self-improving AI, one putting the odds of extinction at 10% or higher. Their own chief executives already back a slowdown in public, so the testimony mostly weakens the labs' claim that they can slow down only if everyone does.
Reality
Google researchers' Dream-RSI cut Gemini calls on a Lasso solver task from 550 to 317 by rewriting a Python search policy, with every model frozen. Teams running scored code search can test it as a call-budget saving on their own tasks.
Reality
Rep. Ro Khanna asked DeepSeek, Alibaba and Moonshot AI if they would be prepared for an incident like this summer's OpenAI-agent hack of Hugging Face. The letters target frontier labs, but the incident they cite is one any team shipping agents has to plan for.
Reality
Anthropic says Claude leads 26% of its measured R&D, a count made by a prototype that partly uses Claude to judge the work. OpenAI's 3.1 agent workdays per human workday measures effort, and neither figure shows whether AI is speeding up AI research.
Reality
Futurist Ramez Naam estimates AI's self-improvement loop would need to be roughly 5 to 10 times stronger to sustain itself, let alone run away. His Noahpinion case supports planning for rapid AI gains that hit diminishing returns.
Reality
Claude Opus 4.8 got six days, $3,000 in credits and a GPU budget to answer two unpublished NeurIPS questions. The papers' original authors graded the output and rejected both.
Reality
Asked on a Dallas tarmac whether AI could end humanity, the president said he had no concerns and put the US a year ahead of China, while three separate catastrophic-risk bills sit in the congressional record.
Perspective Coverage
3 publishersReality
President Trump told reporters he has no concerns about the pace of AI development. Every frontier-risk finding an operator could read this week was published by a lab or by someone who used to work at one.
Perspective Coverage
16 publishersReality
Dario Amodei wants embedded evaluators, shared standards and an antitrust waiver so a frontier slowdown can be checked from outside. Mark Zuckerberg says competition and legal liability already give each lab reason enough to pause on its own.
Perspective Coverage
14 publishersReality
The New York Times reported that the record extends to researchers' own conversations with Faraday, and the public evidence for the wider self-improvement loop still sits on the lower rungs.
Reality
Amodei, Altman, Musk, Hassabis and Nadella endorsed independent evaluators inside frontier labs within two days of each other. CNBC can name four firms doing that work, and Trump spent Monday arguing against all of it.
Perspective Coverage
36 publishersReality
Anthropic has put a number on how much of its own AI research Claude now runs. The number comes out of a pipeline in which a Claude agent catalogued the tasks and a separate Claude judge scored them.
Reality
The lab says Claude leads 26% of its model research and collaborates on about 90%, with both tiers defined by how closely a human directs each task, and it wants rival labs publishing the same measure.
Perspective Coverage
6 publishersReality
The company's 21 September proposal puts recursive self-improvement in scope for technical standards coordinated by CAISI. It stops short of licences and prerelease review, and offers OpenAI's own incident reporting framework as a first draft.
Reality
Automated harness evolution keeps whatever edits raise its own benchmark score, so the harness ends up fitted to the eval set. Google Cloud AI Research answers with five regularizers lifted from supervised learning.
Reality
Anthropic and OpenAI each launched a model on Tuesday that repackages capability their flagships already had. The faster release calendar behind those launches is mostly a pricing story for the people who buy them.
Reality
The plan would link ten national safety institutes through a US agency and leave adoption to each government, with no licensing and no prerelease approval. About 20 countries signed a human-control declaration the same day.
Reality
OpenAI's Monday post routes global frontier standards through CAISI and says they would not be licenses or mandatory pre-release review. The leverage sits with whoever defines how capability and safeguards get measured.
Reality
Product · September 21, 2026 · 1 publisher
Invest · September 19, 2026 · 4 publishers
Product · September 18, 2026 · 3 publishers
Science · September 19, 2026 · 1 publisher
Build · September 19, 2026 · 2 publishers
Product · September 18, 2026 · 1 publisher
Leadership · September 16, 2026 · 12 publishers
Leadership · September 17, 2026 · 1 publisher
Build · September 16, 2026 · 7 publishers
Leadership · September 16, 2026 · 2 publishers
Product · September 15, 2026 · 1 publisher
Invest · September 14, 2026 · 3 publishers
Invest · September 14, 2026 · 4 publishers
Leadership · September 13, 2026 · 1 publisher
Build · September 11, 2026 · 11 publishers
Product · September 11, 2026 · 1 publisher
Build · September 11, 2026 · 1 publisher
Product · September 11, 2026 · 1 publisher
Invest · September 10, 2026 · 1 publisher
Invest · September 7, 2026 · 1 publisher
Leadership · September 6, 2026 · 1 publisher
Product · September 5, 2026 · 1 publisher
Build · September 5, 2026 · 1 publisher
Leadership · August 30, 2026 · 1 publisher
Security · August 25, 2026 · 1 publisher
Leadership · August 22, 2026 · 1 publisher
Invest · August 20, 2026 · 1 publisher
Product · August 15, 2026 · 1 publisher