XCSme

12 ms·

1. 331.▲
   by XCSme 6mo ago

   Oh wow, scrolling through the page on mobile makes me dizzy
2. 332.▲
   by XCSme 6mo ago

   And so the bubble keeps bubbling...
3. 333.▲
   by XCSme 6mo ago

   Oh yeah, I just noticed, like 3x the reasoning tokens.
4. 334.▲
   by XCSme 6mo ago

   In my tests[0] it does only slightly better than Kimi K2.5. Kimi K2.6 seems to struggle most with puzzle/domain-specific and trick-style exactness tasks, where it shows frequent instruction misses and wrong-answer failures. It is proba
5. 335.▲
   by XCSme 6mo ago

   Yeah, I noticed that, HN doesn't let me delete my comment. The other release, Qwen-3.6-Max is the one comparing it to 4.5
6. 336.▲
   by XCSme 6mo ago

   A bit weird to be comparing it to Opus-4.5 when 4.7 was released...
7. 337.▲
   by XCSme 6mo ago

   (commented on the wrong thread, HN doesn't let me delete it :( )
8. 338.▲
   by XCSme 6mo ago

   I loved implementing the Rabin-Karp algoritm, such a fun and celever solution.
9. 339.▲
   by XCSme 6mo ago

   I tried it on their own website: We couldn't scan this site isitagentready.com returned 522 <none> The site appears to be experiencing server errors. This is not an agent-readiness issue. Try scanning again later.
10. 340.▲
    by XCSme 6mo ago

    I am not sure it's a mistake, this might be their new "adaptive reasoning" + hidden reasoning trace, so we can't verify.
11. 341.▲
    by XCSme 6mo ago

    I was initially excited by 4.7, as it does a lot better in my tests, but their reasoning/pricing is really weird and unpredictable. Apart from that, in real-life usage, gpt-5.3-codex is ~10x cheaper in my case, simply because of the ca
12. 342.▲
    by XCSme 6mo ago

    Oh, and also, the "none" and "medium" variants performed the same (??)
13. 343.▲
    by XCSme 6mo ago

    The reasoning modes are really weird with 4.7 In my tests, asking for "none" reasoning resulted in higher costs than asking for "medium" reasoning... Also, "medium" reasoning only had 1/10 of the reasoning
14. 344.▲
    by XCSme 6mo ago

    > Instruction following. Opus 4.7 is substantially better at following instructions. Interestingly, this means that prompts written for earlier models can sometimes now produce unexpected results: where previous models interpreted instru
15. 345.▲
    by XCSme 6mo ago

    I guess you were as confused as me, as I only asociate BackBlaze with B2, I haven't used any other of their services.
16. 346.▲
    by XCSme 6mo ago

    Thank you! One of the worst is TikTok, even as a developer, when someone sends me a TikTok link and I have to visit it, I get stuck in the browser (same with the app but I uninstalled it), and it feels almost device-breaking the way they tr
17. 347.▲
    by XCSme 6mo ago

    Initially I thought this was about their B2 file versions/backups, where they keep older versions of your files.
18. 348.▲
    by XCSme 6mo ago

    It just describes what's in the photo and then some completely wrong/random facts about self-esteem, income, religion, etc.
19. 349.▲
    by XCSme 6mo ago

    I guess writing code is now like creating punch-cards for old computers. Or even more recently, as writing ASM instead of using a higher level language like C. Now we simply write our "code" in a higher language, natural language,
20. 350.▲
    by XCSme 6mo ago

    Releasing version 9.0 of my self-hosted analytics app[0]. I will finally add an in-app cron job editor, so you can easily schedule clean-up jobs, data retention settings, newsletters/summaries, etc. [0]: https://www.uxwizz.c
21. 351.▲
    by XCSme 6mo ago

    If it's relevant to the discussion, I hope not. I've spent probably over100 hours working on this benchmarking/site platform, and all tests are manually written. For me (and many others that reached out to me) are not useless
22. 352.▲
    by XCSme 6mo ago

    General intelligence (not coding) comparison: https://aibenchy.com/compare/z-ai-glm-5-medium/z-ai-glm-5-1-...
23. 353.▲
    by XCSme 6mo ago

    The (none) version especially shows considerable degradation.
24. 354.▲
    by XCSme 6mo ago

    GLM 5.1 does worse than GLM 5 in my tests[0] (both medium reasoning OR no reasoning). I think the model is now tuned more towards agentic use/coding than general intelligence. [0]: https://aibenchy.com/compare/z-ai
25. 355.▲
    by XCSme 6mo ago

    Gemma 4 is great: https://aibenchy.com/compare/google-gemma-4-31b-it-medium/go... I assume it is the 26B A4B one, if it runs locally?
26. 356.▲
    by XCSme 6mo ago

    I tried using Astro for https://aibenchy.com , initially it went great, but then I got into static-website limitations (such as dynamically generating all comparison pages, which would been generating N^4 pages, where N is the nu
27. 357.▲
    by XCSme 6mo ago

    I don't have coding tests yet, will add soon
28. 358.▲
    by XCSme 6mo ago

    Good question! I might add them, but there were multiple reasons: 1. Most variants on HIGH/XHIGH provide only marginal improvements in accuracy, but at drastically increased latency and cost. One special example is Gemini 3.1 Flash Lit
29. 359.▲
    by XCSme 7mo ago

    It does quite well on my limited/not-so-scientific private tests (note the tests don't include coding tests): https://aibenchy.com/compare/google-gemma-4-31b-it-medium/go...
30. 360.▲
    by XCSme 7mo ago

    3.6 Plus seems to be simply a refined/more consistent 3.5 Plus: https://aibenchy.com/compare/qwen-qwen3-5-plus-02-15-medium/...