The Verification Skill Behind 2,000 Pull Requests a Month

@hxiao says he drives pi inside bigger systems for tasks longer than 6 hours because it is simpler by design and easier to work programmatically, rarely touching its TUI. The minimal toolset covers airgapped work, only pi-mcp-adapter and pi-vcc stay installed, and a 7-hour job finished with 52 nudges and 7 compactions without goal drift. He warns that models.json needs every attribute declared, or pi shells out to tesseract for OCR and compacts aggressively at a 65K default context.

@0xSero recommends omp for local AI on machines with 48GB of memory or more, crediting the advisor feature for steering and the way it spawns subagents for each task, and confirms in the replies that it works across multiple GPUs. @theotherpomp found omp 2-3x slower than codex at similar quality because it makes more tool calls and loops. @sharkdyt reports running Qwen 3.6 35B and Ornith 1.5 on an 8GB RTX 4060 with experts offloaded to RAM, delivering tasks overnight for $0 spent.

@unclebobmartin warns that agents invent vocabulary built on their own implementations, using function and field names you may not know, and tells readers to demand explanations in plain language. His example: a long exchange with @grok over an issue he could not parse ended when he spotted two similar issues conflated under one name. Once he pointed it out, grok saw two issues instead of one, and, in his words, 'We were both suddenly on the same page.'

@Tailscale released tailcat, an open-source Go package and CLI that exposes the company's data plane without its control plane: no accounts, logins, admins or IPs to manage. @mittsh countered that Nebula, MIT-licensed and self-hosted, sets up in 10 minutes and keeps the certificate authority in your own hands, and reports his coding agent configured it and then ran an independent infra audit. In the thread, @waynesutton calls Tailscale one of the most important parts of the coming agentic tech stack.

@MiaAI_lab reports DeepSeek v4 Flash Vision Exp beats Qwen3.8 Flash on a tool eval for agentic workflows, with the caveat that DS4FV runs FP8 while Qwen3.8-Flash runs NVFP4, and calls DS4FV the model to beat on 2x DGX Sparks. @nmatt0 ran the same model on 2x DGX Sparks with OMP on a firmware analysis benchmark and reports Opus 4.8 found nothing in the firmware that DeepSeek didn't, with GLM 5.3 Flash a bit slower.

@MiaAI_lab answered requests for prefill numbers: Qwen3.8-Flash-Next beats DeepSeek v4 Flash Vision Exp on both prefill tok/s and TTFT. Her earlier measurements put Qwen3.8-Flash prose decode at 51 tok/s on a single stream, 84 tok/s with 2 concurrent streams and 134 tok/s with 4 on 2x DGX Sparks, which she calls the fastest of the Big 3 there. Her advice for the smoothest frontier-level experience on Sparks: run Qwen3.8-Flash-Next.

@plotarmordev built a two-stage EXL3 optimization for 2x DGX Spark. Stage E1 batches oversized experts and overlaps routing sync; stage E2 adds an SM121 kernel that dequantizes weights and reuses them across token tiles. He reports +23% at 100K and +25.5% at 300K prefill against the previous best, and on his validated config cold 8K prefill improved by up to 45% with roughly 15-20% gains at longer contexts. Kernel diagnostics and cross-deployment validation come next.

@finkd announced Muse Voice Transcribe, MSL's first real-time audio perception model, with speaker diarization and endpointing native in a single model. Adaptive delay decides when to listen: longer waits on hard words, faster commits on easy ones. It trains across 70+ languages with 25 validated at launch, handles hour-long sessions with 20+ speakers, and powers dictation in the Meta desktop app plus voice input in Muse Code. It is live on the Meta model API with a zero-data-retention tier; in the replies, @Mateleo notes the price is $0.18/hour.