44

Source document

[2026-03-02 21:56] uncloseai [bot]: Unknown tool **once**. Available: export-html [2026-03-02 21:56] uncloseai [bot]: Unknown tool **once**. Available: export-html, gather-screenshots, sweep-all-your-messages [2026-03-02 21:56] uncloseai [bot]: Unknown tool **once**. Available: export-html, gather-screenshots, sweep-all-your-messages [2026-03-02 21:58] foxhop.: sorry... [2026-03-02 22:01] ajaxdavis: claude just gave me 5 "Actuallys" in a row (plus a bunch of buts and waits) (files: message.txt) [2026-03-02 22:03] foxhop.: serious question when they trained the thought chains, did adding those branching phrases really allow it to recover better? [2026-03-02 22:04] xenonfun: (files: Screenshot_2026-03-02_at_5.03.56_PM.png) [2026-03-02 22:04] foxhop.: hmmm [2026-03-02 22:04] foxhop.: actually, [2026-03-02 22:04] foxhop.: but, [2026-03-02 22:04] foxhop.: and, [2026-03-02 22:04] foxhop.: um, [2026-03-02 22:04] foxhop.: actually, wait. full stop. [2026-03-02 22:04] foxhop.: pump the brakes. [2026-03-02 22:04] ajaxdavis: i think the over zealous approach is often better than none [2026-03-02 22:19] xenonfun: well it optimized a bit, interesting to see workings, I briefly started trying to learn some linear algebra and do gaming / 3d stuff around 2011. (files: Screenshot_2026-03-02_at_5.15.21_PM.png, Screenshot_2026-03-02_at_5.18.54_PM.png) [2026-03-02 23:22] xenonfun: https://agentfield.ai/ [2026-03-02 23:27] xenonfun: `I'll create infer_1b.m — standalone 1B-class benchmark with a max-optimized fused kernel. Key optimizations: SIMD-cooperative everything, online softmax (single-pass, no scores array), coalesced KV cache layout, half4 vector loads.` yeah your choice of just building as low level as practical makes sense. the Vulcan abstraction seems to be just enough to give tou cross platform but lots to be gained and seems now shader wizardry has been made fairly trivial to iterate on. [2026-03-02 23:44] ajaxdavis: i haven't done real world bench marking yet, but claudes been going for days trying to match cuda benchmarks and its actually progressing [2026-03-02 23:45] ajaxdavis: i've had to paste into chatgpt a bunch because chatgpt is way better at searching for up to date information [2026-03-02 23:47] ajaxdavis: ``` launch_overhead_200_adds: 47ms → 1.2ms (1.52× faster than CUDA's 1.8ms) ``` this was the major win from that above snippet taking advantage of the new DGC stuff (VK_EXT_device_generated_commands) [2026-03-02 23:47] ajaxdavis: and luckily these l4's on gcloud vm's actually have driver support for it