A 27B Model Now Runs on the Card Already in Your PC

pstack 0.15.9 ships /correct: when you keep correcting an agent for the same mistake, it finds the pattern and fixes it with architecture, types and checks. @poteto also upgraded /architect for agent-friendly design and added a /benchmark-checklist skill built on @brendangregg's checklist. A poteto-mode prompt sends Grok Bot to rearchitect a repo and report back with a plan.

@CardilloSamuel rebuilt his personal assistant around the hermes agent. Two DGX Sparks run GLM 5.3 Flash for calendar, news and email duty, tied to omnivoice for speech and whisper for listening on an RTX 5090. His lab gear, he adds, runs GLM 5.3 Flash uncensored on two RTX 6000 Pro Blackwell cards under a compute dashboard.

@0xSero's EXL3 build took a month to finish. It reports 2,000 tok/s prefill, 29 tok/s prose decode, 41 tok/s code, a 2 million KV pool, 262k context, 8-way concurrency, vision, 92% top token agreement and 0.07 KLD. @bertholomusai's TensorFold build v0.2 raised prefill from 1,357 to 1,935 tok/s and code from 69.9 to 80.4. Four streams rose from 80.5 to about 92, and start time fell from 65 to 35 seconds, bit for bit.

@christopherowen cabled four DGX Sparks into a ring with no switch and taught NCCL which way is clockwise. His all-reduce finishes before NCCL picks an algorithm: 17 microseconds for a 10 KiB message against NCCL's 91. It runs as one kernel, replays in a CUDA graph and stays bit-identical on every rank; the repo is public under Apache-2.0. Next up: letting the GPU ring the NIC doorbell itself; the first attempt rebooted a node.

@ItsmeAjayKV generated a 60-second motion graphic on KV cache in one DeepSeek V4.1 Flash API call. The API stats: 5 minutes 25 seconds of “Max” thinking, 85K total tokens, 275 t/s average decode and $0.09228 total. All CSS and HTML sat in one file. @superalesha had Claude Opus 5.5 direct a pop video about the US China chip war with no video model, and published it as a skill.

@cryptopunk7213 sorts out the Cerebras saga: OpenAI uses Cerebras chips for Ultrafast inference and ran short because Jane Street bought up supply. @FPuklowski relays SemiAnalysis numbers implying Ultrafast sells at about $200M per megawatt a year against a Cerebras cost of roughly $26M, and notes people keep re-running the maths hoping for a typo. @sama confirmed the partnership without the numbers.

President Trump announced a “Super Intelligence Force” to ensure America leads the world in super intelligence, @KobeissiLetter reports. @_The_Prophet__ reads the move as the state organizing around a capability before it fully exists. Compute, energy, chips and talent, he writes, get wrapped inside sovereign strategy. He calls it “how an intelligence industrial complex begins”.

@chamath argues today's AI sellers charge cost per thousand tokens plus markup, and tomorrow's will sell the task itself: “Accomplish the task, participate in the ROI”. He expects the transition to debunk current gamesmanship around revenues, gross margins and run-rates, and points to Web1 and Mobile for the value destruction that followed.