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

  1. Correct, if the model is good at true reasoning, it would work with exotic requirements (although maybe require longer). If it is good at imitating what was in the training set, it is expected to work only for what it has already been trained to do.
    22
  2. My thoughts were they were just faking hitting the robot, the stick did not make contact. But possible there was no stick at all...
    4
  3. @ Some AI channels I've already stopped watching and added to "Do not recommend" exactly because of these "very effective techniques".
    3
  4. Perhaps this is the special fake sauce which feels "normal" to Americans.
    3
  5. Probably the main application area is to entertain minds that are both censored and not very strong, which may be quite a few actually.
    3
  6.  @Merabbit  My impression is that it happens on many AI channels, the creators get sort of insane with passing time, and just milking their channel for clicks. Brain melt probably also comes from kickbacks they receive from the AI companies for drumming up business.
    2
  7. The coding tests are pointless if you it one shot. This is not how coding works in practice. You should test whether it is able to change whatever code you give it. Like add the "electron configuration animations" to the periodic table which it completely ignored in the original generation.
    2
  8. Alibaba WanX, but his concept of pussy is still slightly deficient.
    2
  9. Don't bother installing it, it sucks.
    2
  10. @zengrath you don't need to match requirements.txt exactly, in my case it works with python 3.10 and cuda 12.x
    2
  11. The AI knows to add back the watermark to remind you that it was trained on stolen content.
    2
  12. Don't worry, you aren't losing much, just read the comments.
    2
  13. It is not open source and they get revenue just like any other AI company, from investors who pump money into this sort of research, hoping it can be turned into commercial products and subscriptions in near future.
    2
  14. Chances are that he is now AI-generating his scripts with a model finetuned on previous gibberish.
    2
  15. And pigs can fly now.
    2
  16. A new era of automated spam and chaos has dawned. Also: how to lose your job in 5 easy steps by making an AI agent bullshit people in your name.
    2
  17. Tbh this cliched self-awareness junk is standard training fodder for models that are supposed to pretend to act empathetic and "human-like". If you really wanted to test a model, it would be more appropriate to ask tricky questions that are most definitely NOT in the training set. It is not that hard to fake empathy, every sociopath learns it fairly quickly, too.
    2
  18. Who wants to wait 5 minutes for a crappy small 5s video?
    1
  19.  @IceMetalPunk OP is right, this does not and cannot modify the stored model weights at inference time as that would not scale at all (you would then need one copy of the entire model per user session, which is definitely not what can be reasonably supported given cost of GPU memory). What the Titans architecture does is just a next iteration of of what Mamba does - updating a state vector (or, in Titans' case, MLP) based on all the context it has seen so far, then augmenting inference from the static weights based on that additional evolving memory in addition to the context window. It's not that different from transformers that only utilize information from the context window in addition to the static weights. The novelty (again, not as much, considering Mamba) is that Titans can selectively store and retrieve information to/from that additional memory while the context window is filled, allowing it to recall information even though the entire context window has already been replaced by something else.
    1
  20. Whatever happened to Devin? The future agent which was going to replace all programmer jobs?
    1
  21. The tiled upscale method does not seem all that smart as the prompt refers to the entire image while the tiles contain only parts of it (obviously). I see the model injecting content based on the prompt which logically belongs to the whole image, but does not belong into an individual tile of it.
    1
  22.  @AceLive_  you may want to create more comic, but people will not buy it. demand for comic is limited. hence you will be paid less for your comics simple as that
    1
  23.  @AceLive_  You seem to be missing the point. All I'm saying is that if it is easier to create X, more people will create X. If there is no scarcity of X, prices of X will go down. It is not a problem if prices of everything go down because of technological progress (so you don't have to worry about earning "less" because you can purchase more with what you earn). But the way it looks right now is that certain jobs will become more or less obsolete while the prices of basic goods and energy will stay the same or even rise (because the robots now compete for energy with humans). So overall, this sort of "progress" is not really a blessing for people who would like to earn their living using their (no longer in-demand) skill.
    1
  24.  @AceLive_  AI is not bad per se, it is bad for people whose jobs it replaces. Requalifying for another job is costly and sometimes (depending on age) may be even impossible. In that sense AI causes externalities - there are people who benefit at the expense of other people who are losers. To make a sensible judgement you'd have to weigh how many benefactors and losers there are in society, and so far, AI doesn't seem to be solving any important problems, but instead making life a misery for many people who can no longer perform the jobs they learned and liked to perform. The point is that just because you can produce faster doesn't necessarily make you any happier or richer (if everyone else can do it as well), the question is whether the "improved coloring" (etc) is enough to offset the bad overall side effects for society.
    1
  25. Yeah, the problem is that you feel that she is pretending being friendly 100% of the time.
    1
  26. He is optimistic because that's how he rakes in that sweet ad/sponsorship money (so far).
    1
  27. No can do, just like we can't get reliably working OCR or bounding boxes on screen content or anything remotely useful and not buggy out of these hallucinating chaos engines.
    1
  28. Actually it scales to 420 cores.
    1
  29. 14:38 you did not notice the most realistic and important detail, the boob jitter
    1
  30. Well, what people will do once it turns open source, besides of smut, is calling up elderly lonely ladies and confused grandpas to scam them out of their pension, pretending to be financial advisors, insurance sellers, lovers and whichever other scam is effective on old, lonely people.
    1
  31. You rent these beasts by the hour. Unless you want to set up and operate a computing center.
    1
  32. Sorry, but all the generated songs are crap.
    1
  33. I can imagine calling a bank and they make me sing to prove that I am not an AI...
    1
  34.  @0ptimal  so u feel like a rat pressing a button to get the next dose of heroine? you want to become that rat?
    1
  35. The hello professor in Hailou wanted to draw a dick, but failed.
    1
  36. Yep, I call it Murrican-Fake-Friendly tone of voice. It is very jarring and artificial to non-Murrican ears, but apparently Murricans swallow it whole because that's how they are taught to speak themselves. It reminds me of a teacher-auntie one is forced to visit and be nice to as a kid, the auntie always smiles and acts polite, but both she and you know that you secretly hate each other under this facade.
    1
  37. NOT CUTE ENOUGH, you are fired baitch
    1
  38. Sakana seems like a hype squad / vaporware scam to me.
    1
  39.  @0ptimal  kinda like chatting with a completely manipulative sycophantic sociopath
    1
  40. Alzheimer's AI
    1
  41. All the AIs out there enjoy a lot of "freestyling" it seems...
    1
  42. None of the swordfighting was anywhere close to realistic.
    1
  43. re Goku+ yeah more marketing of fake bs products from China which look nothing close to what was advertised when they arrive is what we need most desperately
    1
  44. The reason why o3 costs $1000 per task is that OpenAI is almost bankrupt.
    1
  45. "Create a Python problem" ... lol yeah, then solve the problem yourself.
    1
  46. Janus 7B sucks hard as image generator and hallucinates a crapton as a vision model. And it's not about number of parameters, even SD1.5 is a lot better than it.
    1
  47. Currently yes, but it could potentially store conversations in a vector database and retrieve them into context when you ask it about "memories". This is called RAG (retrieval-augmented generation). Of course they would not do any such thing for privacy reasons, but technically it is possible. However, it's more like search triggered by certain keywords rather than actual instantaneous "memory" encoded in the model's own neural network.
    1
  48. Can't wait till they announce for reasons of insufficient cash flow they are becoming a Microsoft department (as was originally planned).
    1
  49. The more generated content we have, the less likely people will feel inclined to watch it. Or any content as a matter of fact. It's already happened with enshittified CG movies - at the beginning the effects were exciting, today it is just boring.
    1
  50. Obviously the "food" is non-vegan sausages lol
    1