OpenAI’s Discounts Pull Coding Work at Rippling and Replit From Anthropic

@Marktechpost reports that Cognition open-sourced its Devin Vulnerability Fixes plugin: three agent skills for PHP2Shell, the Oct. 21 Node.js release and GitLab CVE-2026-110264. Cognition’s blog lays out the flow: a read-only inventory, your approval, a patch to a verified fixed version, tests before and after, then one PR per repo.

@clairevo says she has moved 80% of her personal coding to Claude Code with Opus 5.5. Her reasons: Claude Code projects is “goated,” she can no longer press “steer” in Codex, and there is less front-end slop.

@rauchg writes that sometimes you over-guide the AI and sometimes you under-guide it, and managing that is “the #0 skill I look for.” @poteto replies that he wrote about this in his pstack guide.

@mattpocockuk says this has been his best year selling self-paced video courses, 98% of his income since 2022. What is worth learning now, he argues, are techniques that are harness-agnostic, model-agnostic and language-agnostic, such as module design and test quality.

@doodlestein says he is at nearly 1T tokens per month and 50k+ commits per month across everything. Skills provision his Linux machines identically through his acfs tool, and his rch and dsr tools and their skills handle builds.

@niklaslenz_ai gave DeepSeek-V4.1-Flash-1017 the same voxel-world prompt on two Sparks and on four. Two Sparks ran @bertholomusai’s 2.5 bpw EXL3 quant on recipe v0.7.2 and finished in 73.7 min.

@niklaslenz_ai ran GLM-5.3-Flash on two Sparks with @MiaAI_lab’s recipe: a 4 bpw EXL3 quant split across both (TP2) with a DFlash2 draft model. He measured 51.2 tok/s on code, 48.6 on prose and 99.2 tok/s total at eight requests.

@niklaslenz_ai ran Qwen3.8-27B on one RTX 5090 with three engines and the same airship-dock prompt. @BobClawblaw’s TensorFold recipe with DFlash2 decoded code at 268.2 tok/s, Neroued’s NInfer at 169.0 and @MiaAI_lab’s kit at 154.5.

@ashxhart released TensorFold 1.1.4. On GB10, a new Qwen3.8-27B recipe improves cold prefill by 54% at 2k and 416% at 64k, where it goes from 203 to 1,045 tok/s.

@NftHoss runs MiniMax-M3.1-Flash on four Sparks with SGLang at TP4 over a switchless RoCE ring, at the checkpoint’s full 1,048,576-token context. He measured 59.0 tok/s single-stream and 122.9 tok/s across six concurrent streams, with prefill at 1,121 tok/s at 32k.

@vllm_project says vLLM matches DeepInfer, DeepSeek’s own engine for DeepSeek-V4.1-Flash, on DeepSeek’s decode configuration from batch 8 to 256, at DEP32 against DeepInfer’s EP64DP64. Its blog puts decode within 6% at 131K context.

@jundotkim’s oMLX 0.7.3.dev1 runs MiniMax M3.1 Flash on Macs. With its bundled DSpark draft running as Lightning MTP, one request on an M3 Ultra with the oQ4e quant goes from 22.5 to 42.0 tok/s (+87%), and new fused decode kernels add another 6-7%.

Sam Altman told OpenAI employees that the industry needs better ways to slow the development of potentially dangerous AI, @theinformation reports. His proposed paths include international regulation, merging leading AI labs or allowing one company to dominate computing resources.

@amorriscode wrote on Saturday that his last day at Anthropic was the day before. He created Claude Code on desktop, he says, because he believes “everyone can be a builder.”