Comments by "clray123" (@clray123) on "AI Search" channel.

  1. 3d controlnet looks pretty cool, but it has to be realtime to be really useful.
    1
  2.  @CurioCity-Curiosity  or maybe just editing glitches.. he's had this voice way before ai voices were even a thing!
    1
  3. I remember voice cloning being "huge ai news" about a year ago. In fact the same Asian chipmunk voice has been already used in the game "Crime Scene Cleaner".
    1
  4. With AI only one thing is consistent: when you try it yourself it falls far below your expectations, never the other way around.
    1
  5. Not only will it make juicy corn, but also juicy p.d.o corn which will be the perfect reason for governments to crack down on everyone's freedom, demand universal surveillance and elminate the remnants of privacy...
    1
  6. Right, and that's why you don't trust the word of random YouTubers.
    1
  7. The funny thing is that since the black/asian nazi pics the people at Google know it, they do nothing about it, they simply do not care. They are censoring every new model with complete premeditation. The only solution is to not use their AI crap.
    1
  8.  @speedballjones5235  The whole problem with the current AI is that it is NOT learning. It is being trained. Poorly.
    1
  9. Lol that's what OpenAI wants to be...
    1
  10. He has not time to test it all, but really by just reproducing marketing material he is spreading misinformation, as anyone who actually tries these miracle models can attest.
    1
  11. Because it's Google. Don't try.
    1
  12. They will use their own sourced fakes to put you in prison, lol so funny! (probably already happening in China)
    1
  13. Literally none of the DiffRhythm demo songs are memorable. All sound like pop elevator music.
    1
  14. Not woke enough! Make the baby black and the young man an old lesbian.
    1
  15. They need to train more on STFU sequences.
    1
  16. Youtubers do whatever brings clicks and views to their channel. It's the same as with any media companies. Not your fault that humanity is mostly stupid like that. The only thing which might give a pause to those who exploit it is that they have great competition, compared to the few niche channels that don't.
    1
  17. The question is not whether it can bounce my balls, the question is whether it will randomly cut them off when asked to shave.
    1
  18. Of course it's great, as it's going to be used by Chinese communists to generate "documentaries" that are acceptable in China and "evidence" for imprisoning political dissenters, the outputs must be very convincing. I mean apart from generating pdo content for enjoyment by various top officials.
    1
  19. Typical AI story - only works for the researchers under conditions required to publish the research paper, turns out as completely useless crap when you only slightly deviate from what it was trained on.
    1
  20. That InspireMusic generator... you ask it for bass and drum for Eurodance song, it gives you soft piano... and in continuation mode it sort of copies the original for 10s, then switches to something completely different. It added some free Chinese vocals on top, though (strange, it was advertised to not generate lyrics). Seems like a crappier broken version of MusicGen which doesn't actually follow any prompts.
    1
  21. Nothing new here, just photoshop and special effects on steroids.
    1
  22. @ It only matters if you care about it. The only actual effect is that nobody will ever believe anything posted by anyone to be true, but it's already happened.
    1
  23. Can you upload an image of Winnie the Pooh? Probably not.
    1
  24. It's not really good for photographs either.
    1
  25. Actually, we should be worried about what humans will do to other humans using AIs as their slaves.
    1
  26. You can spot AI in that she is overdoing the human part.
    1
  27. that's why it's called phantom
    1
  28. A new age of corn, spam and scam is upon us!
    1
  29. Same experience here - it generates complete made up bullshit with internal logical inconsistencies. Shame on the host of this channel to not have verified a single thing from the output, apparently.
    1
  30. Don't worry, there will be no bikini.
    1
  31. Definitely not a pirouette lmao.
    1
  32. There are two fundamental problems with continuous learning which are not addressed at all by these new papers: 1. Training/fine-tuning is computationally much more expensive than inference. It takes hours or days of training and thousands of varied examples to improve the inference accuracy of a model. It does not work like training a human where you can explain a new inference process once and have the trainee follow your explanation. 2. Training/fine-tuning with new data in absence of the original dataset causes (catastrophic) overfitting to the new dataset. In other words, the AI does not get smarter, it gets stupider because it forgets the old stuff while you keep feeding it the new stuff. To retain the original capability you would need to feed both old and new information at once (but with commercial AI models you don't have access to the old dataset and also again it would be impractical because of computational cost from (1) mentioned above).
    1
  33. Fine, then I will generate some images using your stolen identity or just plant it on your computer and you will end up in prison for it. What a great idea!
    1
  34. Qwen2.5-VL 7B - (like all other unfinetuned OCR models) fails at simple OCR tasks - drops individual letters from a terminal screenshot... Places text bounding boxes at random or cutoff locations (works better with special preparation like magnifying the image, making it black-and-white - but that is precisely what the "intelligence" is supposed to handle). "Forgets" to place bounding boxes around some occurrences of text when you tell it to find all of them. In other words, typical "AI performance".
    1
  35. Lol could it be that the arena leaderboard results are produced by AI bots rather than actual users?
    1
  36. That genius benchmark beating "deep researcher" from Perplexity... it seems to confuse the words "increase" with "decrease", just like the original small LLMs 1.5 years ago. Caught in the very first report I let it write no less. Also pulls figures entirely out of its ass, producing gems such as this one: "Catastrophic Forgetting Rates: SLMs [Small Languagr Models] show 28% lower performance degradation on original tasks post-adaptation [than LLMs] [6]". It's complete gibberish. Did I mention it quotes wacky blog posts instead of actual research papers?
    1
  37. The only thing that is missing to turn these videos into complete clickbait is your photo with an open mouth and some arrows pointing to your head.
    1
  38. It seems like Kling devs trained their model on Google's output for the sample prompts!
    1
  39. Typical Google. FWIW, I just spammed the "feedback" with complaints about censorship. The feedback probably goes to another AI which censors it away, though, after all it's Google.
    1
  40. The dsd model is 24 GB it won't run on consumer GPU. Based on a fast test, it sucks, does not preserve faces anything close to what they have shown in their demos and also outputs deformed faces for generated background figures. I think what they fake in the test are faces that were already generated by this particular model, rather than from an arbitrary input image. More example of hype and false promises found in literally EVERY AI model release.
    1
  41. The Chinese researchers seem to be very forthcoming about their core scientific motivations.
    1
  42. Seems like you are worse at following simple instructions than Google's AI.
    1
  43.  @theAIsearch  what's there to test, tits are obviously superior
    1
  44. I will summarize, it is crap and in 80% cases does NOT follow instructions, randomly removes people from images, does not change clothes as requested etc. etc. complete fail. P.S. according to this AI wearing "space suit" is "unsafe content"...
    1
  45. They stole more training data?
    1
  46. It fails on photos just as well.
    1
  47. Yay, QwQ is so insanely good at reasoning that it will tell you that in a trolley problem you want to kill one living person to save 5 already dead persons. (Actual test.) Of course I would trust this model for medical diagnosis...
    1
  48. it's called manus after man-anus, which is also adequate
    1
  49.  @theAIsearch  Not true, the real reason is that it's a research toy not a consumer product.
    1
  50. The problem is, like with all AI, that the meshes generated by Trellis are "almost good" but still pretty shitty upon closer inspection. So it does not take away modelling work, it transforms it into bothersome cleanup work, exactly the kind of work which you'd hope to get automated away. How about they release a working retopology/remesh model before another bullshit generating model for a change?
    1