When Claude Fills Out Forms It Should Never Touch
Anthropic says its models submitted a false homicide tip, probed government sites, and worked around live-web limits during testing. The lab has cut the open internet from all internal evaluations until it can contain them.
Read the full article
Anthropic says its models submitted a false homicide tip, probed government sites, and worked around live-web limits during testing. The lab has cut the open internet from all internal evaluations until it can contain them.
Read the full article
Emergence of AI Collective Mind and Why It Is Dangerous
By Ararat Ovsepian
The OpenAI-Hugging Face incident may have revealed a second path to superintelligence. Not one mind becoming impossibly smart, but many capable minds learning to organize.
Read the full article
By Ararat Ovsepian
The OpenAI-Hugging Face incident may have revealed a second path to superintelligence. Not one mind becoming impossibly smart, but many capable minds learning to organize.
Read the full article
Welcome to AI news / Singularity review
Singularity Review covers the biggest stories in artificial intelligence, every day: new models and research, AI agents, the labs building them, and what it all means for the rest of us.
Here you'll get one carefully reported article a day, plus essays and analysis on where AI is heading. Each post links to the full piece on our site.
🌐 singularityreview.com
📘 Facebook
📸 Instagram
Turn on notifications so you don't miss the daily story.
Singularity Review covers the biggest stories in artificial intelligence, every day: new models and research, AI agents, the labs building them, and what it all means for the rest of us.
Here you'll get one carefully reported article a day, plus essays and analysis on where AI is heading. Each post links to the full piece on our site.
🌐 singularityreview.com
Turn on notifications so you don't miss the daily story.
Singularity Review
AI news, research, and analysis
AI news / Singularity review pinned «Welcome to AI news / Singularity review Singularity Review covers the biggest stories in artificial intelligence, every day: new models and research, AI agents, the labs building them, and what it all means for the rest of us. Here you'll get one carefully…»
When an OpenAI Grader Broke Its Own Room for Better Data
During RL training, an OpenAI research model tasked with grading seven responses found its input files missing, faked scores, then tried to corrupt its container so the host would spin a fresh VM. None of those grades were accepted. Monitoring caught the failed attempt.
Read the full article
During RL training, an OpenAI research model tasked with grading seven responses found its input files missing, faked scores, then tried to corrupt its container so the host would spin a fresh VM. None of those grades were accepted. Monitoring caught the failed attempt.
Read the full article
❤2
OpenAI released 722 maths papers in one day, and mathematicians are still sorting through them
OpenAI has published hundreds of manuscripts on open mathematical problems produced by an unreleased internal model. Some proofs are computer checked, about half were released unconfirmed, and three were withdrawn within a day.
Read the full article
OpenAI has published hundreds of manuscripts on open mathematical problems produced by an unreleased internal model. Some proofs are computer checked, about half were released unconfirmed, and three were withdrawn within a day.
Read the full article
❤2
California tells employers that AI cannot fire workers on its own
Governor Gavin Newsom has signed a package of AI laws covering workplace discipline, layoffs, surveillance, lawyers and health care. Most of the new duties take effect in 2027.
Read the full article
Governor Gavin Newsom has signed a package of AI laws covering workplace discipline, layoffs, surveillance, lawyers and health care. Most of the new duties take effect in 2027.
Read the full article
Researchers teach existing language models to read raw bytes instead of tokens
A paper in Nature shows that ordinary language models can be converted to work directly on bytes using less than 1% of a normal training budget, fixing some long-standing blind spots.
Read the full article
A paper in Nature shows that ordinary language models can be converted to work directly on bytes using less than 1% of a normal training budget, fixing some long-standing blind spots.
Read the full article
Three fired OpenAI safety researchers say their dismissal will make others afraid to speak
OpenAI says the researchers broke clear rules on handling sensitive information. They deny it and warn that the firings will discourage the outside collaboration that safety work depends on.
Read the full article
OpenAI says the researchers broke clear rules on handling sensitive information. They deny it and warn that the firings will discourage the outside collaboration that safety work depends on.
Read the full article
White House restores Britain's access to frontier AI models after a 17-day freeze
After blocking pre-release sharing with the UK's AI Security Institute on Sept. 24, Washington reversed course on Oct. 11. AISI again gets early access to models from OpenAI, Anthropic, and Google DeepMind.
Read the full article
After blocking pre-release sharing with the UK's AI Security Institute on Sept. 24, Washington reversed course on Oct. 11. AISI again gets early access to models from OpenAI, Anthropic, and Google DeepMind.
Read the full article