silverqx

eol ~ eof « tsr

[silverqx @silverqx1]

@slater57649 from the ride directly to chiropractic 😅

[silverqx @silverqx1]

@Exocija stop spreading fear unnecessarily; there is already enough tension

[silverqx @silverqx1]

@DavidOndrej1 how good is it in your workflows? I would like to see your YT test or your opinions on this model 👋

[silverqx @silverqx1]

anthropic cries other distill them, what even may not be true fully, but they will at sure implement MLA and kvcache compression developed and open-sourced by deepseek, oai is doing it right now. Anthropic is more conservative in pushing new changes to their models, so it will take 1y they implement this and be cheaper

[silverqx @silverqx1]

@shaunmmaguire thats all they have, shorts, flip-flops, and alah 😅

[silverqx @silverqx1]

@jun_song and all that works with 1M ctx with kvcache hit after 20h, its like black magic compared to all other AIs, its not frontier but for everyday cheap work its already enough

[silverqx @silverqx1]

@jun_song should be mandatory to describe exact params for subscriptions; users should know what they are buying, every product must have an exact composition label on it, only closed-source AI providers are hiding it and selling things upside down as they want

[silverqx @silverqx1]

@elder_plinius most fun I have ever experienced was 3y with c++23, it has very sophisticated low-level std lib, but am not using it actively from then, development takes too long

[silverqx @silverqx1]

@jun_song 🐋 is the only provider that has MLA with kvcache TTL more than 24 hours backed up by nvme, kvcache has 3.5GB for 1M ctx, compared to other providers where kvcache has 350GB and ttl 5-30min, no other provider has that, dspark is another thing other providers dont have as well

[silverqx @silverqx1]

it can also be that they need to slow down open source progress because they have too much money in and need a few years to get it back or be somehow profitable, but open source is progressing so rapidly that in a few years users will be able to run 5.6-like models on consumer HW, and these companies will have problems

which would mean they are only protecting their own investments; somebody should ask these CEOs this QA

[silverqx @silverqx1]

@RaoulGMI @AnthropicAI reasoning about massive constitution and guardrails pushes it off the track, poisoned ctx

[silverqx @silverqx1]

@beyoncegarden he was like wth 😅

[silverqx @silverqx1]

its not about releasing cuda or ms office as open-source, its about they support open-source and not telling its dangerous or a security threat or similar nonsense, or not limiting or prohibiting it, the better analogy here is nvidia vs nouveau gpu drivers, or ms office vs libreoffice

[silverqx @silverqx1]

@maria_rcks almost all essential info were exposed, if a few words land clearer, we would have had a full review 😅

[silverqx @silverqx1]

@Lila_is_onX Sonet 5 is empty shell, there is nothing, its even unpleasant for me to write to it, I dont feel good querying these AIs, even their English is terrible, very unnatural. Some paragraphs don't even make sense to me

[silverqx @silverqx1]

@MatthewBerman unbiased attitude, you always explain best 👍

[silverqx @silverqx1]

I think it can be even that way China does progress, release everything open-source, all papers, and US then implements these new improvements, we can expect US will do MLA and kvcache improvements in the next round, once TTL for kvcache will increase from anthropic 5min and oai 30min to higher values, we will know it happened

[silverqx @silverqx1]

@zimm3rmann @max_spero_ you have no idea right?

[silverqx @silverqx1]

@CodingGarden 👍

[silverqx @silverqx1]

@Angaisb_ until you realize how much it cost him, it was not a few thousand $