wuschel

7 ms·

1. 61.▲
   Microsoft Paper: LLMs Corrupt Your Documents When You Delegate (Arxiv.org)(arxiv.org)

   7 points by wuschel 6mo ago|2 comments
2. 62.▲
   by wuschel 6mo ago

   Why a retrain? What are you using the model for?
3. 63.▲
   by wuschel 6mo ago

   This is not true for some cases e.g. there are stark differences in the correctness of answers in certain type of case work.
4. 64.▲
   by wuschel 6mo ago

   How is the quality of model answers to your queries? Are they stable over time? I am wondering how to measure that anyway.
5. 65.▲
   by wuschel 6mo ago

   > but the quality of output will not be usable Making the the right pick for model is one of the key problems as a local user. Do you have any references where one can see a mapping of problem query to model response quality?
6. 66.▲
   by wuschel 6mo ago

   Great project! Thank you for that!
7. 67.▲
   by wuschel 6mo ago

   Depends what type of models you look at. There were many German designs that were much less prone to technical breakdowns due to pragmatic and mission focused design choices e.g. many of the Jagdpanzer ("tank destroyer") class lik
8. 68.▲
   by wuschel 6mo ago

   What near SOTA open models are you referring to?
9. 69.▲
   by wuschel 6mo ago

   Hi, You mentioned switching out LLM backends. Would you think that lesser models e.g. one of the recent Qwen variants would work for your use case?
10. 70.▲
    by wuschel 6mo ago

    This is a precious tool that unlocks freedom of choice by reducing the barrier to leave an ecosystem. Thank you for that!
11. 71.▲
    by wuschel 7mo ago

    Seconded. Would love to hear about “best Fraunhofer practices” and fort hand experience. What could one improve how the operate?
12. 72.▲
    by wuschel 7mo ago

    Sourcehut comes to my mind: https://sourcehut.org/
13. 73.▲
    by wuschel 7mo ago

    One should be able to define a PIN number that wipes your phone.
14. 74.▲
    by wuschel 7mo ago

    USD / kg?
15. 75.▲
    by wuschel 7mo ago

    Thanks for your comment! I thought that Chinese local governments receive central government tax revenue shares and sell/develop land property as main income source. May I ask where you got some of the information from (e.g. re ownersh
16. 76.▲
    by wuschel 7mo ago

    What do you mean exactly?
17. 77.▲
    Nvidia and Thinking Machines Lab draw multi-year chip deal(siliconrepublic.com)

    3 points by wuschel 7mo ago|1 comments
18. 78.▲
    by wuschel 7mo ago

    Snippet from article: "Huang has previously said that one gigawatt of AI data centre capacity costs up to $50bn, which would place the latest deal’s value to several billion dollars. In addition, the company has also made a “significan
19. 79.▲
    by wuschel 8mo ago

    I wonder: How does Rust and Haskell compare to Go when it comes to LLM code generation? I was always thinking that the compiler induced feedback loops could give these languages an edge? What would be the best language properties for LLM as
20. 80.▲
    by wuschel 8mo ago

    Wow, did not know about the regional blocking of git servers! Makes me wonder what else is kept from the western audience, and for what reason this blocking is happening. Thanks for sketching out their approach and for the URI.
21. 81.▲
    by wuschel 8mo ago

    There is an post describing the possibility of an organised campaign against archive.today [1] https://algustionesa.com/the-takedown-campaign-against-archi... How does the tech behind archive.today work in detail? Is there
22. 82.▲
    by wuschel 8mo ago

    I found the metaphors / analogons rather disturbing e.g. metabolism, etc. Does the human body have a Chinese assistent with an oxigen tent at 8000 metres?
23. 83.▲
    by wuschel 8mo ago

    I think the problem is that there is not much data presented in this article - or other articles of that kind. Systems are resilient - until they are not, and exhibit sudden factures and fast collapse. The question is not whether Russia col
24. 84.▲
    The New Way to Build a Startup (YC Video, YouTube)(youtube.com)

    1 points by wuschel 8mo ago|0 comments
25. 85.▲
    by wuschel 8mo ago

    Perhaps there is an incentive to go back to OS that can operate with 640KB RAM ... /s
26. 86.▲
    by wuschel 9mo ago

    > I'm convinced it's because that first ChatGPT release was probably trained on data almost entirely untainted by other LLMs, and it may no longer ever be possible to obtain such a dataset again. Interesting statement. But woul
27. 87.▲
    by wuschel 9mo ago

    Hi, would there be a a way to contact you? I have an email in my profile. Would love to exchange some thought on that. China is not only a strong player in biotech. Their capability in chemical R&D and market transfer is very strong, to
28. 88.▲
    by wuschel 9mo ago

    What could be the potential risk of not being compliant to the software license at hand e.g. let us say we would sue a GPL violator?
29. 89.▲
    by wuschel 9mo ago

    Interesting. Where is it possible to read up on this?
30. 90.▲
    by wuschel 10mo ago

    Could you elaborate a bit on your example? What do mean by "that sort of pattern matching" and the argument of "latent space perspective?" Thanks!