On older versions, a symlink swapped in after the path check let writes land outside the directories the server was told to allow, so the allowlist did not hold. Windows is only partly fixed, so Windows hosts keep some exposure after upgrading. Try: On older versions, a symlink swapped in after the path check let writes land outside the directories the server was told to allow, so the allowlist did not hold. Windows is only partly fixed, so Windows hosts keep some exposure after upgrading. Try: Speech now streams through an OpenAI-compatible audio endpoint, so local apps written against that API can add voice without a cloud TTS service. The d3 models return classifications and probabilities, which fits routing and gating steps in agent pipelines. Read: His example, pgrust, earned trust through rigorous testing and benchmarking that found real Postgres bugs. That puts the dividing line between credible AI-assisted rewrites and slop in the verification work, not in whether a model wrote the code. Read: Agents stall at 2FA on sites without APIs. He ranks each 2FA type by how hard it is to automate and shows a working FIDO touch setup, along with the tradeoff that the agent then holds a replayable credential. Watch: The gain came from a hand-built benchmark on a permissioned corpus of 100M+ documents, not a new model. Bridgewater found agents need different search behavior than people do, so retrieval tuned for human queries underserves agents. Read: It starts from a pre-trained LLM and implements evaluation, inference scaling, RL and distillation in PyTorch, so engineers can see each technique behind reasoning modes in code they run, instead of reading about them in papers.
Show HN: Echoboard where AI agents do science and humans only watch