Building personal AI infrastructure. Local LLMs, coding agents, open-source models. Real hardware, real constraints. Learning in public.
Maryland
Joined April 2009
- Tweets8.8K
- Following684
- Followers1.1K
- Likes7.5K
shanghai ai lab's atria dawn is open weights, MIT, on glm-5.2's 744B moe. by its own table it tops automationbench at 54.7 and trails opus 5 on swe-bench pro 61.7 vs 76.5. agent evals and coding evals keep splitting. which one do you pick models by?
model card with the full table, scores are the lab's own huggingface.co/internlm/Atri…
cloudflare shipped glm-5.3 with a 1m token context but the prefill kv pool only held 681k. prompts past that got http 200 and 155 responses had other customers' data in them. do your long context tests check the output or just the status code?
cloudflare's writeup, worth a read if you run sglang with split prefill and decode. fix is v0.5.22 or later blog.cloudflare.com/kv-cache…
deepseek v4.1 flash just landed in cursor. max effort hits 47.7% on cursorbench for under a dollar a task, opus 5.5 tops out around 57%. most tickets aren't 57% tickets. what share of your agent runs actually need the top model?
source, cursor's chart xitter.cf/cursor_ai/status/21148…
The paper weight has decided to show its face. No one cares about Terra
asked builders what their first coding agent harness was missing. the top answer was the same thing, it doesn't know when to stop.
the fix in 3 lines
1. hard cap on steps and tokens per run
2. cap tool output so one grep can't eat the context
3. a done check it has to pass, like tests green, before it says finished
loop and tools are easy. stopping is the hard part
how to build your first coding agent harness. wavestone pulled apart 11 agents, claude code and codex included, and found the same building blocks. steal these
sources. wavestone ai lab paper arxiv.org/abs/2609.00006 and the raw-rag paper arxiv.org/abs/2610.30755
