Comments by "Chrysippus" (@4.0.4) on "AI Search" channel.

  1. Biggest surprise to me is that the thinking and reflection, wasn't trained. It just emerged from training. Only accurate answers and presence of <thinking> were rewarded!
    77
  2. I know "Code (coming soon)" usually means it won't be coming anytime soon, but I really hope it does this time. Even if it doesn't; this is 8B!!! At some point, we'll run this locally!
    61
  3. ​ @AB-cd5gd  if movies were well written, I'd agree with you. Don't get me wrong, talented writers could make much better movies than AI. Sooo... Hollywood is doomed.
    28
  4. I find it wild how Google's AI seems great for reviewers and benchmarks, and then you try it out yourself...
    24
  5. Honestly, the 70B is probably not too far behind the 671B. For example llama 405B isn't several times smarter than llama 70B.
    11
  6. If cars get commercialized, people might run over each other and drive drunk. I don't think cars should be accessible to the general public, since I believe complete safety is the only thing that ever matters. Anyone who doesn't see it my way just doesn't get it. #SayNoToCars.
    9
  7.  @chad_usa  just a regular home computer, I see.
    6
  8. Low bar these days
    5
  9. ​ @akonako5307  no, they won't upsell you on the camera sensor. But the VRAM and AI chip :)
    3
  10. ​ @999satyam  $700+ for 16GB isn't a good deal, it's just a bit less terrible than Nvidia's.
    3
  11. It sounds like the movie Her. I personally don't like that voice, but hopefully it can be fine-tuned or Lora'd
    3
  12. Of course, everything side A is good, and everything side B is bad. This is such a nuanced and intelligent view of the world. Let's ignore Mistral, Meta, Stability AI, BFL, HuggingFace, etc.
    3
  13.  @kamik111  I'd like Japanese culture to keep existing, for example. I assume you do as well.
    2
  14. No link to code, and probably they're never going to release the code.
    2
  15. I don't think AI is movie grade yet (might take a while), but it will absolutely revolutionize 🌽 and memes. Don't get me wrong, it's amazing even if it looks like shit.
    2
  16. ​ @marwanhbt5  not the one in the video. There is a 14b model that fits in a 3090 and a ten times smaller one. But, the 14b has gguf quants!
    2
  17. For anime it's crazy how all of these are so behind even SD 1.5 models, let alone SDXL finetunes, which is heartening because it means open source will always have room to compete. One workflow to get the best of both worlds is to generate with Flux, and get ControlNet to transpose those details into something else with better style.
    2
  18. Doesn't look good... Maybe for memes it's somewhat usable.
    2
  19. @ronilevarez901  I have not provided any info for downloading llama, since I never used the basic un-tuned model and it's open source. Also I did have to provide info to try out some Chinese AIs, so this isn't one-sided.
    2
  20. You are massively over-estimating the quality of what Rockstar will be able to produce. Companies have been changing a lot due to politics.
    1
  21. ​ @kamik111  there aren't that many people in civilized countries, and those are both decreasing and getting replaced. If you take a billion Somalians and put them in Japan, they won't suddenly start making anime.
    1
  22. ​ @kamik111  not in all countries. Some places like India, China, etc have overpopulation, others are slowly vanishing.
    1
  23. ​ @kamik111  the combined total of all humans doesn't matter if you want cultures to keep existing (e.g., Japan)
    1
  24. Yeah it uses the framework he made and popularized. It's nowhere near the same thing though.
    1
  25. @Worldmaster-mq3vr let's not forget Samsung was making any white circle on a black background into a crisp HD full moon 🌝
    1
  26.  @shabadooshabadoo4918  there are plenty of good upscalers, from pretty fast like Siax to incredibly slow like SUPIR.
    1
  27. ChatGPT o3-mini reasoning is somehow really bad for creative tasks compared to DeepSeek-r1. It's subjective so I encourage everyone to try.
    1
  28. DiffVSR is the one I cared about the most but the code link still goes to a 404 page... is this the new "Code (Coming Soon!)"? Just have the "Code" link, technically pointing to a GitHub URL?
    1
  29. Prompt: "Battery material, ((high power density)), cheap, lightweight, masterpiece, best quality, trending on arXiv, by John B. Goodenough and Stanford R. Ovshinsky"
    1
  30. ​ @74Gee you might pick an used 3090 for that, not a 4090
    1
  31. Out of these, SORA is surprisingly not even in the top 2 of what I found cool. Trellis was the coolest, MMAudio second (both have code!!).
    1
  32. Microsoft already owns OpenAI...
    1
  33.  @darekmistrz4364  only accurate responses were rewarded, not all the anthropomorphic reflection during thinking. That's emergent.
    1
  34. They're really trying to make an AI gf, huh?
    1
  35. As opposed to the very coherent, lucid, aware, Joe Biden.
    1
  36. What do you mean? Is the code not there?
    1
  37. 36:31 - kinda impressive just how far behind Meta is considering how much money they burn. I hope Llama-4 is a banger.
    1
  38. X Portrait 2 seems like a perfect tool for memes. Hope they do release the code! Keep us posted.
    1
  39.  @marwanhbt5  yeah to be clear, it's "Wan2.1-T2V-1.3B". It's a 1.3B model, very small compared to the 14B one.
    1
  40. @ wait an AI decade (a few months in human terms).
    1
  41. It's gonna be wild when they have LoRAs.
    1
  42. I'm glad that this is being developed, even if it's still at a point where I wouldn't even enable it if it was as easy as a toggle, let alone dig into code to get it working.
    1
  43. At some point in the future, effects like Pika's will be simply LoRAs for open source video models.
    1