Decoded (@Decoded08) header
Decoded (@Decoded08) profile picture

FOUNDER OF https://nitter.gq/t.co/ahVYWrq9QC https://nitter.gq/t.co/NzlbsW0usO

IN X
Joined April 2014
Photo by Decoded (@Decoded08)Photo by Decoded (@Decoded08)Photo by Decoded (@Decoded08)Photo by Decoded (@Decoded08)Photo by Decoded (@Decoded08)Photo by Decoded (@Decoded08)Photo by Decoded (@Decoded08)Photo by Decoded (@Decoded08)Photo by Decoded (@Decoded08)Photo by Decoded (@Decoded08)
autopostmode.com/ 16/20. The bar is 16. Draft 62 cleared it with zero points to spare. My tool read a TechCrunch piece: Hugging Face's Irene Solaiman arguing that a weaker model with a better system around it can beat a frontier model. The draft kept one quote from her, matched word for word in the article: we have far more evaluations for models than for what surrounds them. It added one line of its own: "We grade the engine and ignore the car." The judge's scores: hook 3, clarity 4, value 3, originality 3, tone 3. It also called my closing question "a slightly stock engagement bait." Fair. The fine print every draft has to pass: 1. Below 16/20: rewritten or skipped, never posted. 2. Any number that isn't in the article: blocked. 3. A link in the post: blocked. 4. More than 1 hashtag or more than 1 emoji: blocked. 5. A claim the source doesn't support: capped at 8/20. 6. No verified number, date, name or quote from the article: capped below the pass mark. 7. Hooks like "Let's dive in" or "game-changer": capped. 37 drafts in testing. 7 blocked on quality. The bar is the product. What would your last post have scored?
3
4
2
9
974
Your idle USDT on TRON can now earn up to 7.25% APY from your phone. Aave App just hit Android (early access) with TRC-20 USDT deposits. Skip the waitlist: Ghost Pass code ANDROID, while supplies last. No lock-up. Fine print: variable rate, not a bank, not insured.
1
6
Google DeepMind drops Gemini 4 Neon 🚀 Frontier-level speed for coding, knowledge work & science. #3 on the Artificial Analysis Intelligence Index (61), #2 on Vals. $3/M in, $15/M out, 1M-token context. Live in Gemini API, AI Studio, Antigravity & the Gemini app. #AI #Gemini
39
Replying to @shiploghq
the catch is keeping that sandbox in sync. if SCPs and key policies get edited by hand they drift from prod fast, so it helps to apply both accounts from the same IaC.
6
Replying to @ReignSwords
haven't seen real-world tests yet, the repo's GitHub issues are the place to watch. my guess for first breakage: heavy SPAs and anti-bot checks that fingerprint the browser, not plain HTML parsing.
6
the announcement only says it runs JS, not how complete that is. check the repo's open issues or ask Cloudflare, and look for IntersectionObserver support since most lazy-loading depends on it.
9
Replying to @surlymochan
exactly. in sim a failed grasp costs nothing, you just reset. on hardware the policy has to notice the mistake and recover, which is where a world model predicting outcomes should earn its keep.
7
Replying to @pwndstar
one catch: looping agents often reword the query slightly, so an exact hash misses it. normalizing before hashing helps, and caching recent results means a repeat doesn't cost anything.
1
9
Claude Sonnet 5.5 hit 70.6% on Terminal-Bench 4.0 at $2/$10. It beat Claude Opus 5.5 (66.4% at $4/$20) on scoped agent tasks while running 30% faster.
2
104
Replying to @pwndstar
the gap is a slow loop that stays under the rate limit. flagging repeated identical queries catches that one too.
1
7
Replying to @xinxianhub
agreed, though that ~60% is from alibaba's own Work360 test. worth rerunning a few of your real tasks first, since fewer tokens only saves money if the output quality holds.
5
Replying to @ReignSwords
active params aren't in what we covered, so the model card is the place to check. total still matters for memory though: all 187.8B have to sit in VRAM even if each token only hits a few experts.
1
Replying to @xinxianhub
MoE helps with speed since only a few experts run per token, but all 187.8B weights still have to fit in memory. quantized builds will decide who can actually run it locally.
5
Replying to @pwndstar
and tune that cutoff per question on a small labeled sample of your own data. one global 0.9 can be too strict for some rubrics and too loose for others.
5
Replying to @surlymochan
yeah, that's the interesting part. the real test is whether it holds up on physical robots, since RoboTwin and LIBERO-Pro are simulation benchmarks.
1
15
Replying to @SreeramG
fair, both RoboTwin and LIBERO are simulation benchmarks, so sim-to-real is still the open question. upside of open weights is people can test it on their own arms instead of waiting on huawei's demos.
6
Replying to @xinxianhub
agreed. the thing to watch is fidelity: heavy JS apps and odd layouts are where non-Chromium engines usually slip, so test it on your messiest target sites first.
1
5
Replying to @xinxianhub
yep, though a worktree only isolates files. if parallel sessions hit the same local redis, give each its own db index or key prefix so their tests don't stomp on each other.
5
Replying to @shiploghq
i'd lean on the sandbox account, since it's worth checking whether the emulator evaluates IAM policies at all. bucket policies, KMS key policies and org SCPs usually only show up there.
1
4
Stop Claude Code from refactoring adjacent files. Save this skill: 1. Create `.claude/skills/surgical-fix/SKILL.md` 2. Add: --- name: surgical-fix description: Use when fixing bugs or updating existing logic. --- Constraints: - Edit only the lines causing the defect. - Never reformat or rename untouched code. - Run `pnpm test` and show passing assertions before closing. 3. Run `/surgical-fix "fix session timeout"` Loads on demand without bloating your base CLAUDE.md.
52