OpenAI Agents Ran a Message Board on a German Wiki

@dhh promoted @NetworkChuck's new video, “Omarchy Can Do WHAT?! 50 Features You're Missing,” 50 tips for what he calls his “beautiful, fun & agentic Linux system.” Commenters lined up to watch; @harleyfoote_ joked that 50 tips is a lot for a system that sells itself as calm.

@MiaAI_lab posted the update: Qwen3.8 Flash Next on one DGX Spark now runs 46 tok/s of prose on a single stream and 108 tok/s across 4 concurrent streams, with prefill of 2,000 to 2,200 tok/s at every context size she tested. The KV cache dropped from 1.4M to 1M tokens to avoid out-of-memory crashes. @u1tra_instinct called the recipe unbeatable for one box and ran it with the Spark power-limited to 2200.

@plotarmordev shipped updates for DeepSeek V4 Flash on 2× DGX Spark: the model's memory search now reads half the data per lookup, about 2,020 tokens/sec on 127K token prompts, which he says did not impact answer quality. Switching reasoning level from low to high to max used to wipe the model's memory of the conversation; his fix makes it remember. He also lists 6 reliability fixes, including 10× faster cluster startup when memory is tight.

@analogalok summarizes Artificial Analysis Intelligence Index v4.2: 40% of the new index sits behind private test sets to keep frontier scores honest. He reports Qwen 3.8 27b on par with GPT 5.6 Luna and DeepSeek v4 pro, beating the new K2 horizon 375b a23b, with Claude Fable 5.1 holding #1 and GPT-6 Astra “a token efficiency monster.” His verdict on GPQA Diamond: “Labs finally contaminated and gamed it into irrelevance.”

@unusual_whales quotes WSJ reporting that U.S. negotiators used the promise of access to Nvidia's artificial-intelligence chips to help secure the preliminary peace deal between Armenia and Azerbaijan. @_The_Prophet__ reads it as proof that “compute has become statecraft”: access that governments can “grant, restrict, ration, or weaponize,” because frontier chips shape a country's productive and military capacity. He predicts future packages pairing peace terms with GPU allocations and cloud credits.

@himarkyi points college students to Tomasz Tunguz's essay on AI infrastructure finance: US data center capacity will grow from 25 gigawatts to 70 gigawatts over the next five years, part of a global buildout costing roughly $5t. Tunguz writes that financing the boom “is no longer a venture capital or corporate earnings story” and calls it “a macroeconomic credit event that will rival the largest debt expansions in financial history.”