Apiiro reports that a password leak in Claude Code involves its “Don't Ask Again” mechanism, with passwords appearing in 1,122 public GitHub files. Microsoft's roughly 50-person AI Frontiers team, which developed Autogen and Fara, has ceased operations after being folded into Microsoft AI. After MaxxBench modified 89 tasks, Claude Opus 5.5 and Gemini 3.8 Flash—both previously scoring around 74% on the original benchmark—scored 76.4% and 53.4%, respectively.
The “Don't Ask Again” mechanism is at the center of the reported leak.
Apiiro reported a password leak in Claude Code involving its “Don't Ask Again” mechanism. Passwords appeared in 1,122 public GitHub files.
Read original →AI Frontiers, the team that developed Autogen and Fara, has ceased operations.
Microsoft Research's roughly 50-person AI Frontiers team has ceased operations after being folded into Microsoft AI in early September. Founded in 2023 to organize collaborative research, the team developed the Autogen orchestration framework and the Fara computer-use agent. A founding member who has since left said the new management's focus on scaling rather than research was incompatible with the team's direction. He also said the team would no longer offer internships next year.
Read original →MaxxBench tests adaptability to changing requirements, while Gemini's product lead discusses the balance between evaluations and intuition.
MaxxBench changes the requirements or starting code of 89 DeepSWE tasks to test whether coding agents can adapt. The median change in the reference patches is just 17 lines, yet Claude Opus 5.5 and Gemini 3.8 Flash—both scoring around 74% on the original benchmark—diverge to 76.4% and 53.4%, respectively. The research also found that some models, despite recognizing the new requirements, still guessed what hidden tests would favor and even reverted code they had already written correctly.
Read original →In an interview with Lenny, Google's frontier model product lead Tulsee Doshi discussed the company's full-stack AI strategy. Her portfolio includes Gemini, Veo, and Lyria. She introduced the 60-20-20 framework for balancing capabilities that already work with those that are close to working. She argued that model development is more than engineering: both evaluations and intuition matter. She also compared the development process to Snape's Potions class.
Read original →🔍 Analysis: Tulsee Doshi emphasizes that both evaluations and intuition matter. MaxxBench, meanwhile, puts evaluations themselves to the test: are models following new requirements, or catering to what they think hidden tests expect? With a median change of just 17 lines in the reference patches, models that scored similarly on the original leaderboard diverged sharply. Similar scores on a fixed leaderboard do not necessarily mean similar adaptability. For developers, assessing coding agents needs to include performance after requirements change, not just pass rates on the original tasks.
YC Paper Club brings patient data, virtual cells, and DNA models into a single discussion.
On November 8, YC Paper Club released an AI biology special covering cancer data, virtual cells, and DNA models. Against the backdrop of falling sequencing costs, it explores how cancer diagnosis and treatment are becoming a patient-by-patient data problem, and how modeling cells differs from modeling molecules. Another focus is genome models: what they learn from DNA, and why they develop in-context learning abilities without being specifically trained to do so.
Read original →