Altman Backs Dario's Plan To Pace The Frontier

@MichaelGannotti rounds up the week: Nous Research shipped Hermes Agent v0.21.0 with Bot Mode, named agents, group chats and steerable subagents, then v0.21.2 to fix fragile session stores. Grok Bot went enterprise with two weeks of free usage for customers including Legora, Supermicro and ServiceTitan, plus iPad and Android clients at twenty dollars a month. OpenClaw 2.0 drew 933 contributors, and Omarchy added DigitalOcean as a founding patron at $3M over three years.

@builtbysketch shipped PaperRoute, a Paperboy-style game built with GPT 5.6 Astra writing code and Blender scripts building the meshes. The ledger: 39 hours tracked, 25.2 human and 13.8 agent-only, 1.56 billion tokens with 1.53 billion cached reads, $2,175 of API value, 90 commits across 11 days. He credits 60-70% of the render-to-render improvement to the agent's review loop, and says sustained 60fps on phones is still unproven.

The AEON-7 vllm-ultimate container moved to upstream vLLM v0.29.0 compiled for sm_121a, serving Gemma-4-26B-A4B, Qwen3.6-27B and Qwen3.6-35B-A3B from one image with DFlash speculative decoding and NVFP4 weights. Release notes cite DFlash 2 at 3.39x single-stream and 1.41x at 64 concurrent on Qwen3.8-27B, 1M context via YaRN verified at 543,761 tokens, and a dual-Spark TP=2 fix worth +26% single-stream.

@ViC305 published DeepSeek-V4.1-Flash at a 3.30 bpw routed-expert average in EXL3, targeting two DGX Sparks with vision and the Engram conditional memory preserved. The full pack is 415.8 GiB, roughly 189 GiB of it Engram, leaving about 113.4 GiB per Spark, which he calls extremely tight. The 2-Spark runtime is not yet validated; he owns one Spark and wants dev receipts on boot, memory and quality.

@sudoingX spent two weeks living on every 100B+ model that fits one DGX Spark 128GB and ranks four: qwen 3.8 flash next at 44.4 tok/s with 1M context fitting in memory, ling 3.0 flash at 42.1 tok/s but down 83% at 51K context, laguna s 2.1 near 35 tok/s, and qwen 3.5 122B at 35.3 tok/s. The cuts matter too: deepseek v4 flash decodes 16.5 tok/s with no drafter, too slow to build with. stepfun 3.7 flash at 121GB took the box down twice before serving a token.

@MichaelGannotti runs two DGX Sparks and three jobs: Qwen3.8-Flash-Next owns spark-d369 for reasoning, writing and coding after scoring 137/157 on his Official A harness, coding a perfect 30/30. spark-56bc generates video with NVIDIA's Sol-H3-Spark recipe, 5 seconds of 768p in 68 s against 423.2 s for the old FL2VA serve. One heavy engine per box is the rule; stack two and the watchdog or thermals win.

@SamJWasserman published a speed recipe for MiniMax H3 video generation on two NVIDIA DGX Sparks, 1376x768 at 24fps with original voice restoration and measured lip-sync correction; the repo headline is 8s of video in 2m54s. @MichaelGannotti, who runs H3 on a single Spark with Qwen 3.8 flash on the other, calls two-Spark H3 "a boost in ability."

@0xSero says he is back to GLM-5.3-Flash because DS4.1-Flash "fails to just do things and needs me to keep pushing it," though he sees more potential in it. Citing @ArtificialAnlys, @landontgreen reads GLM 5.3 Flash as still the superior model and the better pick for 2x DGX Sparks. @chandu_cpz calls DeepSeek a buddy model and GLM an autonomous agent.

@MajmudarAdam argues the past two weeks of fear trace to what lab staff see on scaling plots: capability advances come from scaling existing laws or finding new ones, and several axes look unsaturated. As a proxy he notes models went from unable to do advanced math to solving "a millenium prize problem" in under 2 years. His worry: hacking shows early competence, so superhuman cyber ability may not be far off.

@AndrewYNg argues AI engineering blurs the old split where PMs and designers specify and developers build, since one person can now do far more. He lists four skills for shaping the build: driving the build loop, making product decisions, communicating and leading, and high-agency ownership. Product managers and designers, he adds, are picking up the same AI engineering skills in reverse.