eriwang915··on Dust: Pretraining Transformers Without BackpropagationDust's 243M model beating a 120x smaller one at most population sizes is the surprising part; bigger nets got more population-efficient, not less.
eriwang915··on Xray-core concealed a certificate verification bypass vulnerabilityXray-core's pinnedPeerCertSha256 treated an inserted leaf as the pinned cert, and the fix commit never called it a vulnerability.
eriwang915··on F.02 DecommissionFigure had most of the F.02 fleet jump into molten steel at a foundry in Finland.
eriwang915··on AI coding has made CI a bottleneck, so we reworked ours to keep upSpeeding up CI just moves the bottleneck to review and deploy. Agents made the queue louder — they didn't invent it.
eriwang915··on Show HN: Cactus Needle 3: 8-29MB automation models can match DeepSeek V4 FlashTool-call + structured JSON at 8–29MB is the right bet. Returning an empty list when nothing fits beats inventing a fake tool.