GLM 5.3 Flash Beats DeepSeek V4.1 Flash 2-1 in Verified Test

@omarsar0 built a custom verifier for the /goal feature in his agent harness: Jev checks whether the goal is actually complete after every turn, cheap enough to run continuously where an expensive reasoning model did the job before. He calls System One models perfect for verification and expects more harnesses to adopt the pattern. Proper benchmarks do not exist yet, he says; he is exploring the idea on Pi while his own harness stays private.

@MiaAI_lab’s updated Qwen3.8-Flash-Next recipe for a single DGX Spark measures 117.4 tok/s of prose and 180.2 tok/s of code at 8 streams, on 262K context with MTP k=3. Peak memory drops from 101 to 91 GiB, an official Nvidia NVFP4 becomes optional, and a supervisor restarts the server around the clock. @0xBakeer replied that he was already downloading it.

@Layton_Gott flags Cua’s CUA-S1-FORMS, the first of a System One family for computer use: 706K parameters, a 2.8MB download, free, and small enough to run locally in under 1GB of RAM. On Cua’s form-filling eval it scored 99.7% against 83.6% for hosted Jev. He reads it as a path to one tiny specialist per repetitive task: forms, CRM updates, data entry, browser actions.

@jmacftw sizes real-world local AI at 256GB of memory, about $10K or more, enough to run GLM 5.3 Flash, DeepSeek 4.1 Flash, or Qwen 3.8 Flash Next, though he warns the recipes run frontier-1 or frontier-2 at best and the boxes end up dedicated to serving; sidecars pushed him to a third Spark. @MiaAI_lab calls the second Spark the biggest upgrade and the sweet spot for 90-95% of users. She expects two Sparks to keep better prefill than an M5 Ultra while the Mac wins decode.

@skeptrune shows any open model can behave like Jev with a bit of inference engineering: SGLang’s /v1/score endpoint forces probabilities over a set of answers, and for DeepSeek you add a closing think tag so it answers directly instead of reasoning. He concedes dsv4.1 flash is not as good as Jev, but thinks the same approach on a larger model would match or beat it. Decision models, he argues, kill prospecting and sourcing work, and he built a Jev-powered LinkedIn recruiting agent.

@cbarmorecpa gave Jev 250 synthetic bank transactions, a simplified chart of accounts, and bookkeeping rules, asking for a category plus a decision on whether an accountant should review. It matched his answer key on both for 215 of 250, an 86% run, finishing in about 1 minute 54 seconds for roughly $0.02. Of the 35 misses, 31 were extra referrals to review; two went unflagged, the ones he wants to understand before any workflow acts without a person.

@mark_k reports, citing Reuters, that @AnthropicAI has quietly built its own wet biology lab combining Claude with real-world biological experiments. That follows years of Anthropic warnings that advanced AI could create catastrophic biological risks and may need to be slowed down. “Apparently, AI is dangerously uncontrollable when other people develop it, but perfectly safe when Anthropic controls it,” he writes.

OpenAI’s Noam Brown had argued air-gapping may not stop a misaligned AI because two machines can talk by running a CPU hot and reading the temperature change, as quoted by @firesidealpha. Bryan Cantrill replied that datacenters are built so one machine’s outlet never feeds another’s inlet, hot aisle and cold aisle, and “the only thing being revealed here is the contagion of fear.” @perrymetzger adds that the point of an air gap is keeping a machine from breaking into another across it.

@usr_bin_roygbiv argues max-q and workstation RTX Pro 6000 Blackwells are the largest appreciating asset of the next two years: triple the VRAM of a 5090, supply 10 to 100x lower, and 4 cards the minimum for running dsv4.1 locally. He had modeled $32,000 to $40,000 by next year, now expects it sooner, and sees $100,000 to $120,000 at auction late next year into early 2028, a 15x from April’s price. @Compilethings counters that AMD’s MI 350P, seen at 19,000 euros, is already the better choice.