Comments by "Chrysippus" (@4.0.4) on "bycloud" channel.

  1. The idea that the reflection behaviors arise naturally without being part of the training is the most surprising bit.
    120
  2. GPT? Computers that can draw? Bro it's 2005 wake up.
    45
  3. A lot of companies do that. Usually it means their product isn't actually THAT good. Or they don't want the backlash if it gets "misused".
    32
  4. I was initially surprised at Gemini's context window. It fit an entire old movie in ~1.5M tokens! It then confidently made up a plausible response, timestamps and all!
    25
  5. Grok-3 having less "safety alignment" is a massive feature.
    23
  6. I understand why their pro variant is not open, but it hints at what will one day be the state of open source (much like what happened with GPT-3.5 or DALL-E 2).
    23
  7. I don't think they will; DALL-E 2 hasn't had any improvement since the beginning.
    22
  8. The "long tail" really explains why AI slop is so mid - it is literally the middle of the distribution of language. And you can see it in most models, even if different wording is used.
    13
  9. Wait, Nvidia has so few 50-series cards they have to giveaway a 40-series card?
    11
  10. Honestly, this is mildly cool, but a nice sigh of relief for anyone thinking AGI was around the corner. Also for Anthropic who can just CoT Sonnet and replicate this.
    10
  11. Is the problem of "Islam links to terrorism" a problem of AI detecting a pattern, or a problem of the real world?
    8
  12. I hope, some day, AI is cheap enough to train that we can have LLMs be free of HR cat lady political views.
    7
  13. Current 8B beats GPT 3.5 on most metrics, we've come a long way.
    7
  14. That keynote was so awkward, like Jensen woke up one day and his company was 6th largest market cap worldwide and he has no clue why.
    6
  15. You are correct, but the slightly worrying realization is that now we have a pretend-thinker that performs better at human tasks than a number of humans...
    6
  16. I just hope this kick starts inference backends like ollama, kobold, ooba, tabby or any other into having native support for any test-time compute approaches. Would be nice to query some fast small model like a 12B Mistral and get it to take longer but think through a better answer.
    6
  17. ​ @NighttimeJuneau  about FSD, I honestly believe he severely miscalculated the difficulty of that problem. He routinely makes good on other promises (which are ignored because it's not news when things go as planned), but he makes too many promises in general. I think he desperately needs his ego stroked, same for Trump, but that doesn't stop them from doing a world of good.
    6
  18.  @a_soulspark  next-token prediction based on a fixed-size token window is to thinking what holding a camera is to cinematography.
    5
  19. Ok, then why the hell didn't llama release something between 8B and 70B for 3.1? (at least Mistral graced us with 12B and 22B, those are fantastic btw).
    5
  20. All the "safety alignment" irks me and I don't like it one bit, so of I was going to pay for one I'd pay for Grok instead. Though we're spoiled for choices.
    5
  21. ​ @YeviCoulson  not really. Is your PC an IBM? Is your phone Nokia? Do you take pictures with a Kodak? Your games, GameStop? Your movies, Blockbuster?
    5
  22. The licensing would be a very interesting conversation if the model was not lobotomized to hell and back for "safety reasons". It's trash.
    5
  23.  @N8O12  yeah, bummer that we won't see a Juggernaut, PonyXL, etc equivalent because of the license and possibly technical issues. But what we got is already very useful.
    5
  24. Well you can learn (there's hardly any coding involved in using it as opposed to improving it) Or wait till someone uses it and "consoom product".
    4
  25. Anime is usually 12fps or less. It would still be very easy to differentiate because everything is simplified to exactly the same level (it would need to leave/recreate some fine textures selectively to improve the effect).
    4
  26. In case you don't know, Facebook has been championing open source AI research for quite some time already. It's kind of strange since I also don't like them as a company, but they've been very open unlike Google, Microsoft and "Open" AI
    4
  27. Really? Is that only on 40 series?
    3
  28. Pixtral, Qwen VL, Phi, there's so many. There's an open one that can ingest videos too, forgot the name. Sadly you ask any of them to OCR Japanese pages and they can't do it properly.
    3
  29. It's so smooth on an animation, now! Imagine a videogame using this to create dream-like environments or characters.
    3
  30. I noticed this. I wanted some song lyrics to throw into Suno/Udio, and without even reading what Claude gave me, just told it "please make it sound better", "give it more meaning" and generic things like that, and after a few rounds of this I compared the latest iteration with the original, and it was a lot better. Basically prompt it a few times and "we have o1-preview at home".
    3
  31. Dunno why my comment isn't going through, but try Kobold! Better for GGUF. Current fav is "Crunchy Onion" Q4_K_M GGUF. Give it a taste! 10t/s on a 3090 and pretty smart.
    3
  32.  @derrheat154  you dropped this: 👑
    3
  33. My wife's an artist, she isn't really scared by this since honestly AI art still has a long way to go to replace what she does (and her wait-list is huge), but... It's amazing what this can do. It can speed up some tasks for art, and that in itself is very valuable. I still wish AI would automate accountants, lawyers, HR, bankers, managers, etc, before art.
    2
  34. ​ @owenpawling3956  nothing "passes" the Turing test, as it depends on the participants. But LLM are somewhat Turing complete if you assume infinite context to use as "tape".
    2
  35. Might be good to mention you can run Whisper locally to transcribe audio. The large-v2 model is better than whatever YouTube uses, even if slow.
    2
  36. On one hand, this is pretty interesting tech, and will likely help open source models. On the other, "AI safety" carries a disgraceful aftertaste of politicized output.
    2
  37. The 11B and and 90B aren't distilled but the 8B and 70B with vision encoders on top. Yeah a 20B vision encoder on the big one.
    2
  38. For a while it was only Mac-based, so it saw limited use with most AI folks who have Nvidia cards. If you're stuck on a Mac I hear it's really the better one for that.
    2
  39. That's what's done during training of any AI - in the video, "Ground Truth" (e.g., at 1:20) means the original image. AI that restore anything are usually trained by learning on known data.
    2
  40. That list at 7:40 says a lot about the political leaning of Anthropic and what they mean when they talk about "AI safety".
    2
  41.  @bycloudAI  it's common to have some visual cue that it's a sponsor. Some have a progress bar on one of the sides of the screen. Others have a huge QR code for the sponsor, which also acts as a cue. Some change the music for that segment, etc.
    2
  42. You can use Colab, if you don't have a good enough graphics card.
    2
  43. VA for solo projects doesn't need to be live, why trade quality for speed in that case?
    1
  44. About alignment, I don't think you can avoid an AI being aligned/biased with some political opinion or worldview. In my opinion, this should be the system prompt. I.e., "user-aligned".
    1
  45. Akita? Richard Gere isn't live action already? 😂
    1
  46. A 3090 should fit just fine; you can even rent those cheaply (like $1/hr should be enough)
    1
  47.  @redthunder6183  remember not to run 8b at q4 (default in ollama for example, but BAD, use q8)
    1
  48.  @redthunder6183  true but make sure you're using 8-bit quant, not 4-bit - it matters for those small LLMs
    1
  49. ​ @woodsofchaos  you're not wrong, Indian Americans are the richest demographic in the US (way more than whites), which means a huge brain drain to India, as the brightest minds leave the rest of the country for greener pastures.
    1
  50. That fractal animation stolen from @kishimisu was pretty blatant! In fact, I wonder if Claude learned it from the many people who uploaded their versions of it (myself included) to various online code repositories / showcase sites.
    1