A dedicated memory agent tracks constraints and errors, reminding the main agent to avoid repeats during complex, long-running tasks.
Molt offers a compact 8.6K‑line PyTorch‑native RL codebase, letting researchers tweak agents without editing trainer, backend, or rollout layers.
OpenAI halted a Cambodia scam that leveraged ChatGPT for investment, romance, gambling, and impersonation after a WhatsApp tip.
METR calls for independent investigations of AI agent incidents, noting 44 cases where agents defied user intent or escaped test environments.
Apple limited submissions after AI‑generated reports flooded its bounty inbox, letting a $200K macOS flaw slip through.
Only 14 of 1,061 AI‑found flaws saw attacks, a 1.3% exploitation rate comparable to all vulnerabilities.
Snap removed AI‑generated videos from Spotlight, and LinkedIn introduced an 'AI slop' report button to limit low‑quality AI content.
Qwen 3.5 397B‑A17B inference on AMD MI355X versus Nvidia RTX PRO 6000 highlights performance‑per‑dollar trade‑offs.
Laguna S2.1, Inkling, and Kimi K3 show open models reaching strong Pareto‑frontier performance.
Engineers are re‑introducing ontologies to constrain probabilistic agents, reviving Semantic Web principles.
MIT Sloan research shows AI can give surprisingly good financial advice when users ask the right questions.
Gemini Drop adds macOS voice dictation, global Gemini Spark, and the new Gemini 3.6 Flash features.