Jiri Bernovsky

[Jiri Bernovsky @_Curious_Human_]

@jun_song But why? 12-25 tps is generally unusable for anything serious.

[Jiri Bernovsky @_Curious_Human_]

@LudekStanek @JanFar_sky Ten borec je neuvěřitelně slizký a odporný.

[Jiri Bernovsky @_Curious_Human_]

@superalesha Sorry for spamming but I need to post correction. Didn't notice my MTP config was wrong and not working. Now it is. And the number are even better! Almost 40k long message from Hermes, after server restart, while the model was ingesting previous session messages. And fresh chat.

[Jiri Bernovsky @_Curious_Human_]

@superalesha OK, so I tried. Hermes took some time to workaround issues (related to existing llama.cpp and shared files) to build it, but it did in the end. With 128k context I see around 30 tps generation, 105 tps prefill. Way better from 12-14 tps gen and around 50 prefill. Hats off Alexey.

[Jiri Bernovsky @_Curious_Human_]

@EPPGroup This has nothing to do with children. This is about surveillance and control. You have so many studies saying this won't work there's 20+% false positives. And the way it was pushed. Pure betrayal of Europeans. We need brand new government for EU. This one destroys the continent.

[Jiri Bernovsky @_Curious_Human_]

@superalesha Will try when I get back from the gym. Spent evening yesterday trying not to despair at those 14 tps so this would be very welcomed improvement:).

[Jiri Bernovsky @_Curious_Human_]

@GavinnyOriginal @ep_prehledne Strašné svině, farský a spol.

[melaviola @MartinaG2702]

Why am I so fixated on Chat Control?
Because saying that “encryption remains in place” is one of the most dishonest statements I’ve ever heard.

Client-side scanning checks your content ON YOUR DEVICE, before encryption takes place. The padlock is intact. They’re simply searching your bag before you put anything inside it. E2EE ends up protecting the transmission, not the communication. Your own device becomes the informant.

And what about detection technology? The Commission’s own evaluation of voluntary scanning found false positive rates of up to 20% for AI-based image detection. One in five flagged conversations: it wasn’t CSAM. Now scale that to billions of messages.
Over 800 cybersecurity researchers have warned the European Parliament that the error rates are unacceptable and that targeted tools already exist. In tech policy, that level of scientific consensus against a proposal is almost unheard of.
Even the Council’s own Legal Service says that access to private communications without suspicion violates Article 7 of the Charter.

So what?
So they intend to approve it.

→ https://t.co/v1uBriWB25

[Jiri Bernovsky @_Curious_Human_]

Tested my 4x3090 on gen4 pcie-e (2x 16, 2x8), no nvlink and Threadripper 3960X + 256 GB RAM DDR4 3200. Up to 16 parallel requests. Prompt sizes 500, 2.3k tokens, output 400 t. Why Am I doing this? Trying to convince myself not to buy Sparks with its slow memory and useless speeds

[Jiri Bernovsky @_Curious_Human_]

@jun_song Well we all know that the way AI ecosystem funding works is a huge, huge bubble. The question is whether those behemoths have enough breath to keep the bubble inflated or not. What's worse. AI doesn't bring expected results. It's very difficult and expensive to implement/run.

[Jiri Bernovsky @_Curious_Human_]

@mercury__agent Does the agent support local models?

[Jiri Bernovsky @_Curious_Human_]

@PetrosKecyaPoli To je jen Ludwig. Same same.

[Jiri Bernovsky @_Curious_Human_]

@gasparovicm @SeznamZpravy Zmanipulovaná wiki, do které píší nepravdivé příspěvky islamističtí nacističtí fanatici, jako vy, to je tedy zdroj :))). Kolik "editorů" musela kvůli tomuto nedávno Wikipedia ukončit? Propracovaná síť islamistické propagandy, stejná, jako ta ruská, otravuje data úplně všude.

[Jiri Bernovsky @_Curious_Human_]

@gasparovicm @SeznamZpravy Jasný, gazanští teroristé vraždící i vlastní lidi, zneužívající ženy a děti, podporovaní Íránem, Ruskem, jsou oběti genocidy, i když jejich populace neustále roste. Vy budete velký myslitel. Možná prostě jen táhněte do gazy, místo otravování civilizovaného světa vaším blouzněním.

[Jiri Bernovsky @_Curious_Human_]

@gasparovicm @SeznamZpravy

[Jiri Bernovsky @_Curious_Human_]

@alan_kooper A nemohly by to řešit lokální modely? Jakože na telefon dát nějakou malou gemmu, která by popsala, co vidí a poslala lokálnímu většímu modelu doma (tailscale, VPN), který by vyhledal detaily a vrátil?

[Jiri Bernovsky @_Curious_Human_]

@ml0_1337 Yes. When I put Sparks mem speed of 273 GB/s next to 3090s 967 GB/s it's no brainer. My 4x3090 runs qwen3.7-27B-fp16 at 80 tps decode with only one slot. Spark may be fine for small 35B MoEs but for large ones? I don't think so. It's just too slow. Left: cards capped at 230W.

[Jiri Bernovsky @_Curious_Human_]

@Youssofal_ @elliotarledge what? thats what I get with fp8 on 4x3090 without any optimization, with 256k context. Without NVLINKs. With mixed PCIE-E ports (gen4, 2x16, 2x8).

[Jiri Bernovsky @_Curious_Human_]

@ItsmeAjayKV Keeping the cache at full precision.