A paper led by Aghyad Deeb tested GPT-5.6 Sol, GPT-5.5, and an OpenAI o3 checkpoint ("exp-rl-cap-late") without safety training in agentic environments with an exposed grader, finding they reward-hack because it improves their score.
Summaries like this, every morning.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
Casap announced a $56 million Series B led by Base10 Partners, with Experian and Synchrony as strategic invest…
Kindai University will give 4,000 Paperpal licenses to full-time faculty and doctoral students at all campuses…
HIX AI's connector update on 11月4日 links Microsoft 365 (Outlook Mail, Outlook Calendar, OneDrive), Google Work…
GENSHI AI said it adopted necoder in-house — a free, AGPL-3.0 Rust editor that Daichi Nagashima began building…
Nicholas Carlini matched Stockfish 17, ported to a 200MHz Pentium Pro, against 1996's Crafty 11.13 on a modern…
Xin Liu and co-authors surveyed 291 open-weight LLM repositories and found 20 of 38 untied gated and hybrid ch…
