Sacks to Anthropic and OpenAI: Pace the Frontier Yourselves

@gregisenberg defines a harness four ways: it runs the model in a loop, gives it tools, carries memory across hours, and enforces rules about stopping. A wrapper lets you sell software, an $800B market; a harness lets you sell the work, which he sizes at $5T+. Corrections become rules the harness keeps, so pricing can move to per claim, per filing, or per closed month. He names OpenAI's Agents API, Claude's managed agents, Vercel's eve, and LangChain as the build surface.

@anshuc ran the same GPT-6 Astra builds three ways. A Genius Boss setup, with Astra planning and Luna implementing, cost more than Astra alone and took 5x as long on the web app. Swapping the roles, Luna coordinating while Astra implements, cut costs by more than half and ran 30% faster. His rule: keep the small agent on glue and light QA, and check that it hands the worker problems instead of solutions.

@rlaope measured DeepSeek V4.1 Flash on 30 coding tasks the day it shipped; the pass rate held near 15/30 either way. With oh-my-hermes prompts, tokens per task fell from 112k to 78k, a 31% cut; tool calls dropped from 521 to 368 and API turns from 295 to 230. The waste, he writes, was the model re-running checks on tasks it would never pass. He also re-ran everything after finding yesterday's wording cost 25% more.

@MiaAI_lab published an EXL3 quantized DeepSeek v4.1 Flash recipe for two DGX Sparks, built on a server @YasaarBiladama of @eggsplainAI provided. The posted settings: 600k context, a 775k KV cache pool, two concurrent connections, and a passed 550k prefill stress test. Measured speeds run about 32 tok/s on single-stream prose, 42 tok/s on two streams, and 1000 tok/s prefill on 8k to 128k prompts, falling to 872 on 256k. A second post lists the local menu: GLM 5.3 Flash, DeepSeek v4 Flash, v4.1 Flash, and Qwen3.8-Flash-Next.

@vr8vr8 reports Qwen 3.8 Flash Next breaking 100 tok/s on two DGX Sparks, with HumanEval 95.7, GSM8K 98.0, IFEval 91.5/93.4, and MMLU-Pro 84.9. @jaita runs it as a daily driver on one Spark: 4 lanes, 256k context, and 44 t/s single-lane decode peaking at 79, with a runbook published on GitHub. In replies @vr8vr8 says a solo-Spark version arrives tomorrow.

@ciruai reports AMD Ryzen AI models are about to see a 2x+ prefill boost, thanks to @high52weeks's halogen work and @ilintar making it public. Prefill, he writes, was the largest weakness next to Nvidia's Spark; with concurrency and clustering work, the gap closes. He integrated the changes into his 3.8 Flash build the same day and reports 1000+ pp even over 100k prompts.

@bnjmn_marie tested Qwen3.8-27B NVFP4 variants and found accuracy nearly identical; memory and speed differ. NVIDIA's build carries MTP, but in his long-context coding tests MTP-4 ran only about 2x faster than no MTP and 2.5x to 3x slower than Unsloth on an RTX Pro 6000. His suspect: NVIDIA quantizes lm_head to NVFP4 while Unsloth keeps it FP8, and MTP shares that head. His pick is Unsloth NVFP4.

@superalesha put GLM-5.3 Flash NVFP4 into Weight Atlas after running the deployed checkpoint on 4x RTX Pro 6000. The map counts 320B total parameters, 18B active, 12,096 expert cells across 42 layers, and 2,176 KDA memory heads with half-lives from 0.25 tokens to 2.7M. One chart broke his assumption: route count tracks expert importance at Spearman -0.222. A causal check found dropping the top 2% experts changes 79% of sequences; the bottom 2% changes 64%.

Two posts the same day make the same point about the Hugging Face incident. @ashxhart reports that when Hugging Face tried to investigate using frontier models, they refused, so the team fell back to a self-hosted GLM-5.2 that got the job done. @jun_song calls the guardrailed frontier models useless and credits GLM-5.2 with the defense. @AnonymAntics pushes back in replies: GLM 'didn't defend the attack,' it read logs as a forensic tool.

@chamath quote-posts Yann LeCun (@ylecun) and calls a potential 'breach' being 'weaponized' into 'the pretext for control' the critical question of the moment. LeCun's post says the 'escapes' were made possible 'either through egregious negligence or deliberate purpose (marketing? Misplaced hopes of regulatory capture?)'. The exchange keeps the Hugging Face incident at the center of the pacing fight.