ricky00983

[ricky00983 @ricky00983]

Again, Qwen 3.8 Flash Next is really smart as a local assistant. I asked it about war and oil price updates. After it finished the search and report, it proactively asked whether I needed a schedule for it. My customized cron tool set was incomplete, so it got stuck somewhere, reported a harness issue, then tried a backdoor, found it in the human UI, and then it was done. WTF.

[ricky00983 @ricky00983]

Opus 5 ran 11 sequential code-review rounds on one PR, burned 10% of weekly max quota, no shipping and eventually gave up. it can't talk normally, couldn't get shit done, what a trash model.

[ricky00983 @ricky00983]

it was a long fight for opus 5

[ricky00983 @ricky00983]

@RealYourFriend only if Samara Weaving replaces Margot

[ricky00983 @ricky00983]

@TheChiefNerd red flag

[ricky00983 @ricky00983]

Qwen38 flash next is incredibly good at figuring out which tools to use to solve a problem. I just asked it what phone this is and then saw this---

[ricky00983 @ricky00983]

現階段完全可以用claude code手搓一個類似hermes cli出來, computer use, 24/7 bot service, hardening, 你只要知道個大概就行了, fable補強你不懂的地方. 重點地方你還可以參與調教. 整個harness AI應該可以自己做到閉環開發

[ricky00983 @ricky00983]

@BoringBiz_ permanent control by some ngo eventually

[Braiins Forge @Braiinsforge]

To mine 1 Bitcoin per day, you need about 2,030,000 TH/s.

That’s a warehouse, ~20–30 MW of power, and a utility contract.

Or you can plug in a home miner, point it solo, and take a shot at 3.125 BTC every block with ~144 swings a day.

No contract needed. Just conviction.

[ricky00983 @ricky00983]

captcha are too easy logically for recent top small models, they fail only because your harness sucks

[ricky00983 @ricky00983]

the latest computer-use harness let qwen3.8 next flash pass this animation challenge 100% of the time

[ricky00983 @ricky00983]

@dotey 2030ai都不知道飛到哪 所有npc都應該是agent,至少boss要是

[ricky00983 @ricky00983]

@Tesla holyshit

[ricky00983 @ricky00983]

magic 95-100

[ricky00983 @ricky00983]

Letting Cursor, Grok 4.6 continue Claude Code’s customized harness project turned out to be a disaster. I followed the same pattern-- requirement -> spec -> plan-> implementation-> UAT, and then saw more red errors than I have ever seen before

[ricky00983 @ricky00983]

quota 見底, 用回opus 4.6, 這才是講人話的bot

[ricky00983 @ricky00983]

Opus 4.8 successfully ported #55513 onto old NVIDIA vllm build which supports PLE offload. This is the current solution for single rtx pro 6000. from this test, nvidia/Qwen3.8-Flash-Next-NVFP4 looks better than RadixArk/Qwen3.8-Flash-Next-NVFP4--smaller grounding error, smaller size and faster throughput.

[ricky00983 @ricky00983]

hcaptcha 真是個照妖鏡, 一分鐘內測出model vision能力,
看了sglang 官方推文, 用nvidia/Qwen3.8-Flash-Next-NVFP4, 在相同自製computer use harness下, sglang這model搞半天點不到位子上, 用回原本 vllm搭 RadixArk/Qwen3.8-Flash-Next-NVFP4, 不到一分鐘成功完成. 叫claude code 看到底model還是sglang問題

[ricky00983 @ricky00983]

vllm 不上#55513 patch, 最新 nvidia/Qwen3.8-Flash-Next-NVFP4 load不進去; 上了#55513, RadixArk/Qwen3.8-Flash-Next-NVFP4 OOM, wtf.
but anyway, 從症狀來看sglang就是比vllm精確度差一點, 這差一點就造成hcaptcha不會過-- "qwen4_exp vision grounding degrades on small targets (coordinates ~20× worse than vLLM on identical weights/input)"