verdverm

18 ms·

1. 661.▲
   by verdverm 3mo ago

   https://vllm.ai/blog/2026-07-27-k3 Original full blog on their own domain
2. 662.▲
   by verdverm 3mo ago

   Yes, Ai can code, kind of, but it cannot make good design
3. 663.▲
   by verdverm 3mo ago

   Trash bin? ~1M LOC for what? If you cannot review or test it, then why would you trust it? This is why it's better to stay human in the loop during the process. Sounds like you don't think it should take 1M LOC. It's likely y
4. 664.▲
   by verdverm 3mo ago

   Looks like you are using Ai for this kind of thing based on your submission history. I don't think people are looking for proxies to design slop
5. 665.▲
   by verdverm 3mo ago

   It's best not to go back and forth with someone like that. They did not answer my only question in their pass-by commentary, so I feel no need to respond to the quotes they pulled from the article (not their points), on top of the bett
6. 666.▲
   by verdverm 3mo ago

   I'm not sure how much we can trust software that hasn't been security audited by Ai going forward. Humans are not that good at writing code, we are going to need help that relentlessly looks for issues. We should certainly avoid b
7. 667.▲
   by verdverm 3mo ago

   you can use SearXNG with an API call, here is an example https://github.com/verdverm/gmd/blob/main/pkg/web/providers/...
8. 668.▲
   by verdverm 3mo ago

   LLMs are unreliable, which is why I use the plan pattern 1. describe what you want in a paragraph or two 2. tell it to write a plan.md, iterate and review multiple times 3. implement and test, iterate and review multiple times Putting in mo
9. 669.▲
   by verdverm 3mo ago

   > Please don't comment on whether someone read an article. "Did you even read the article? It mentions that" can be shortened to "The article mentions that". https://news.ycombinator.com/newsguide
10. 670.▲
    by verdverm 3mo ago

    1995 was a different leader in a different time. Let us not forget at one point, America forced walked people on The Trail of Tears. American Amnesia is very potent and powerful, we easily forget about last week, year, century When did we g
11. 671.▲
    by verdverm 3mo ago

    with the president trying to sell early access to his market moving "truths"
12. 672.▲
    by verdverm 3mo ago

    1. Everyone has been doing distillation for a long time 2. You ideally want outputs from multiple models, not a single one 3. Distillation (or a model trace) is insufficient on its own (a) you need a sufficiently strong base (b) crafting RL
13. 673.▲
    by verdverm 3mo ago

    considering all the other Ai snafus from this admin, I'd put my money on that being more likely
14. 674.▲
    by verdverm 3mo ago

    I used to avoid all the new features and stuff, but then I wanted an EV. Picked up an Ioniq 6, which still has many physical buttons and such, but boy was I missing out on a lot of QoL and safety features keeping an old car for so long!
15. 675.▲
    by verdverm 3mo ago

    "hey Rosie, the cat yakked on the rug again" (real-time mess handling)
16. 676.▲
    by verdverm 3mo ago

    I think ~car is a more likely range, there's a lot of value in cooking / cleaning, Americans buy cars pretty frequently
17. 677.▲
    by verdverm 3mo ago

    E is for everything TTD
18. 678.▲
    by verdverm 3mo ago

    *hands on a peace of the pie
19. 679.▲
    by verdverm 3mo ago

    Turing was testing our gullability, v2 is a preference test
20. 680.▲
    by verdverm 3mo ago

    Alternative source: https://www.france24.com/en/sport/20260730-uefa-to-boycott-f...
21. 681.▲
    by verdverm 3mo ago

    I'm (not?) surprised how many Democrats will stick with a candidate as well, 40% seems to be the floor for anyone on any question about their continued support. Bernie defending Platner's tattoo was a deal breaker for me
22. 682.▲
    by verdverm 3mo ago

    Jack Ma's perspective would be interesting
23. 683.▲
    by verdverm 3mo ago

    I don't use compaction, I keep my context limited, my sessions fresh, and have agents write markdown files as needed
24. 684.▲
    by verdverm 3mo ago

    you can turn those on through the existing menu system
25. 685.▲
    by verdverm 3mo ago

    ADK seems like a better thought through SDK family, I use Go, Py is most mature, also TS and Java. The docs still need a lot of work to cover all the languages, it's pretty hit or miss right now, but one can also have a coding agent lo
26. 686.▲
    by verdverm 3mo ago

    We have guardrails for input/output and nudity classifiers have existed for a long time, can they not run these things? eg: https://huggingface.co/zentropi-ai There's even work to detect harmful content retraining
27. 687.▲
    by verdverm 3mo ago

    the art of reinforcement learning
28. 688.▲
    by verdverm 3mo ago

    late training is where a lot of capability gains come from too
29. 689.▲
    by verdverm 3mo ago

    How much do we assume human written code for training the original models is free of human made slop? We all know we all take shortcuts and have code we are not proud of. If they are I deed averaging machines, is it possible their output is
30. 690.▲
    by verdverm 3mo ago

    Bernie Sanders and Donald Trump have plans that would put lots of public money into the biggest Ai companies (sovereign wealth fund). They are getting a lot of money through government contracts, which is ultimately funded by the public thr