9:30 PM
Today in brief.
Karpathy: the capability gap is widening
Andrej Karpathy's post (widely shared via HN) argues the shared public understanding of what LLMs can actually do is diverging faster as models improve.
🔧 Tech
Ask the agent why it did not do it first
Venturetwins reports a 6× speed and 8× cost win on Astra by asking one follow-up after a feature ship: "why didn't you just do it this way the first time?"
2 sources
🔬 Science
Labs want 100B models on phones by 2028
Qualcomm CEO Cristiano Amon says frontier AI companies are asking for phones that can run at least a 100-billion-parameter model continuously by 2028, and some are already building their own devices.
10 minutes with AI erodes persistence
A peer-reviewed COLM paper (Berkeley CHAI + CMU/MIT/Oxford/UCLA) finds that about 10 minutes with ChatGPT on fractions or SAT reading cuts accuracy and grit the moment the tool is taken away.
🔬 Science
Sakana MASS shows a working RSI loop
Sakana AI and UC Berkeley open MASS: one model plays executor, workflow optimizer, and evaluator, then fine-tunes on its own multi-agent trajectories for a bounded recursive self-improvement loop.
💼 Business
AI is minting a human premium for art
BBC's Ros Atkins maps the emerging "human premium": Deezer says 44% of uploads are AI-generated and 97% of listeners cannot tell, yet more than half still care that a human made it.
🤖 AI
500B tokens later: agents decompile an FPS
Maurice Heumann's team spent ~600–700B tokens orchestrating autonomous agents to decompile a commercial FPS to C++; the unlock was a byte-match oracle, not smarter models.
🔬 Science
OpenMed 3.0: local clinical AI, no cloud fallback
OpenMed 3.0 ships Apache-2.0 local-first clinical extraction and a Patient Journey timeline that refuses to call the cloud if a local model is missing.
No stories in this category for this edition.