tucnak

30 ms·

1. 571.▲
   by tucnak 2y ago

   Regularise is a really good choice of word :-)
2. 572.▲
   by tucnak 2y ago

   The LIMA paper, I think, would be a good place to start https://arxiv.org/abs/2305.11206 You can create inputs for DPO/ORPO synthetically which is a huge one as previously it would require gigantic investments ht
3. 573.▲
   by tucnak 2y ago

   This is not an accurate assessment; the forward-pass is nontrivial, i.e. you're always adding new information. When they say "synthetic" datasets, nobody is suggesting that the past model is used to invent it completely. What
4. 574.▲
   by tucnak 2y ago

   I often wonder if LLMs would have made Wittgenstein cry...
5. 575.▲
   by tucnak 2y ago

   > We both learned something about the limits of my patience that day. Let me guess: you did absolutely nothing?
6. 576.▲
   by tucnak 2y ago

   If it weren't for crypto, zero-knowledge technology would have been lame
7. 577.▲
   by tucnak 2y ago

   My perception of his intelligence is that he's based and doesn't care what you think.
8. 578.▲
   by tucnak 2y ago

   Unless the hardware employs some variant of encrypted RAM, see cool paper by the NSA from a few months back that includes benchmarks in FPGA and silicon https://eprint.iacr.org/2024/1240
9. 579.▲
   by tucnak 2y ago

   I'm led to believe this is mostly because "known unknowns" are not well-represented in the training datasets... I think, instead of bothering with refusals and enforcing a particular "voice" with excessive RL, they
10. 580.▲
    by tucnak 2y ago

    Thank you, it's always interesting to hear physicists talk about this stuff :-)
11. 581.▲
    by tucnak 2y ago

    I'm sure Wolfram is for real; it's just funny how supposed scientists would jump out of their trousers to criticise him, all the while happily ignoring fraud in their immediately field, their own faculties, etc. Talk spineless. Th
12. 582.▲
    by tucnak 2y ago

    That sounds right, I was judging by mac studio I own that is passively cooled.
13. 583.▲
    by tucnak 2y ago

    Thank God we have scientists rallying against Wolfram, Penrose, et al. It would be great if next-up, you guys actually had the guts to challenge your _actual_ peers, who are turning tricks of mainstream scientific literature at the highest
14. 584.▲
    by tucnak 2y ago

    All of Google is acquisitions
15. 585.▲
    by tucnak 2y ago

    thia is the best assay i read on notion ever
16. 586.▲
    by tucnak 2y ago

    > AI workloads would drain the battery and cook your eggs in a matter of minutes, assuming you can run them M2 Max is passively cooled... and does 1/2 of 4090's token bandwidth in inference.
17. 587.▲
    by tucnak 2y ago

    And how many of these 10M are bots now? I reckon in the high 20s percent.
18. 588.▲
    by tucnak 2y ago

    Is this a joke? You don't have to put everything in 1 table.
19. 589.▲
    by tucnak 2y ago

    What's shitty about stating the obvious? Or do you really think that being "nice", i.e. giving lamers pointless validation that they don't deserve would make anything better? People need to learn, and Internet culture ha
20. 590.▲
    by tucnak 2y ago

    Quality comment. Somebody needs to remind these lamers what home-computing is all about. I don't know why people are upset, it's a really accurate assessment of OP.
21. 591.▲
    by tucnak 2y ago

    Note: the latest OpenWrt on rack-grade routers is significantly more secure than proprietary, backdoored firewall appliances. I haven't really used Netgear/pfense but I reckon it wouldn't be much of improvement. VLAN are perh
22. 592.▲
    by tucnak 2y ago

    Why, yes?
23. 593.▲
    by tucnak 2y ago

    Our military's comms are over Matrix/Element, Whatsapp, Signal (in the descending order of significance.) Officers are trained to avoid Telegram just like the other known-to-be-compromised communication channels. Nobody who's
24. 594.▲
    by tucnak 2y ago

    So what is the alternative? Nomad?
25. 595.▲
    by tucnak 2y ago

    I really wish Supermicro PSU's were accessible over PMBus without their proprietary IPMI utilities[1] but they're not. Moreover, it's x86-only so there's no way to interface it from ppc64el. Had it been open source, it c
26. 596.▲
    by tucnak 2y ago

    In the sense that it's not used for military communications.
27. 597.▲
    by tucnak 2y ago

    Interesting that this comment is down-voted. I think it's worth noting that Telegram LLC—as organisation itself—is anomalous to the point where it's consistent with Russian "cover" doctrine. For example, you wouldn'
28. 598.▲
    by tucnak 2y ago

    Is discrete maths a good starting point if I was interested in cryptography? Certainly better than analysis right?
29. 599.▲
    by tucnak 2y ago

    American way of life
30. 600.▲
    by tucnak 2y ago

    These systems already exists in Ukraine, and no, they are not expensive. Simply running SAM would already be more expensive.