Itch.io adding Generative AI disclosure filtering to few more places on the site
Itch.io is expanding its filters for games and other content disclosed as using generative AI. The filters will appear in more places across the site. (bsky.app)
GPT-Rosalind-6 System Card Addendum
deploymentsafety.openai.com70 points55 comments
GPT-Rosalind-6 is a biology and medicine-focused release of the same model weights as GPT-6 Astra, with no additional training or post-training. It is intended to help qualified researchers with legitimate life-sciences work, including some dual-use questions, while refusing requests to enable biological weaponization. Because it has fewer refusals on dual-use biological queries, access is limited to vetted scientists and organizations whose proposed uses, security controls, and governance are reviewed. OpenAI classifies its biological and chemical capability as High but below Critical, its cybersecurity capability as Critical, and its AI self-improvement capability as below High.
On three research benchmarks, GPT-Rosalind-6 substantially outperformed GPT-Rosalind-5.5 at the highest reasoning budget tested: scores rose from 28.9% to 49.6% on medicinal chemistry, from 16.7% to 37.7% on quantitative biology, and from 44.0% to 61.3% on applied life-sciences research. Preparedness testing covers lab-protocol troubleshooting and tacit biological knowledge, as well as protein design, virus-host interactions, and biological co-evolution. On Biorisk Knowledge, its consensus score was 78.3%, just below the 80% indicative High-capability threshold; adjusting for refusals raised the estimate to 85.0%. It scored 491 of 832 on a test of errors in non-public expert protocols, above that evaluation’s indicative threshold. These evaluations measure specific capabilities, not overall potential for biological harm, and the Critical thresholds are preliminary.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
More in AI
Sam Altman's Case Against Your RAM Is Efficient. That's the Problem
Sam Altman argues that people should use less RAM so memory can be redirected to AI data centers. Critics say OpenAI’s efforts to secure a large share of global RAM supplies could also drive up prices and question whether efficiency justifies restricting personal use. (techdirt.com)
AI agents hallucinate prices: LLM grounding with live data
LLMs often give outdated or incorrect product prices, even when they use web search. The tutorial shows how to build a Python agent that grounds answers in live data and lets code insert verified prices. (blog.apify.com)
Revolut moves to self-hosted GLM
Revolut moved about 2,000 engineers to self-hosted GLM coding models, running them on 56 Nvidia B200 GPUs at roughly $0.30 per million tokens. AI operations lead David Abitbol says engineering lead time improved 70% as lower costs let teams use AI more broadly, while Claude Code and Codex became exceptions. (twitter.com)
Could a modern chess engine on a Pentium have defeated Kasparov in 1996?
A 1995 Pentium Pro running a modern Stockfish port is estimated to play at 3300 Elo, far above Kasparov’s peak rating. The experiment shows how advances in chess algorithms could have beaten Kasparov without Deep Blue’s custom supercomputer hardware. (nicholas.carlini.com)
Open models can do 90 percent of agentic work at a twentieth of the cost
Gartner says open-weight models can handle about 90% of routine enterprise agent tasks at a fraction of the cost of closed models, while frontier models remain useful for complex reasoning. It expects companies to rely more on specialized open models, making the software harnesses that route tasks between models a key source of value. (theregister.com)