Comments by "Roni Levarez" (@ronilevarez901) on "AI Search" channel.

  1. Same. Back then I solved it the simplest way: I didn't learn, lol. I eventually quit school and learned more then than with teachers. Knowledge is pretty useless without the certification from School but still, it's another proof of how outdated the scholarized system is.
    8
  2. Nothing is truly free. They use your activity in their services to make profiles of you and sell them. They use the videos you post to train AI. They own the writings you post on social media and sell them if they want. And on top of that they make more money showing you ads and selling you premium features. For every dollar of " free" stuff they give you they make ten times that. But yeah. I also love to be able to use some AI for free, despite all the limits I hit daily.
    5
  3. "like normal people". It's been a while since a comment made feel so "abnormal" 😐
    5
  4. It might look good standing still at the kitchen, chopping stuff or cooking, but I'd like it to have a more typical body, including legs. It would look better to me. (Also, you never know when it's going to be necessary to run uphill chasing those pesky humans 🙃).
    4
  5.  @Mede_N  evidence suggests that models can encode subgoals on weights, thus allowing them to provide aligned responses while preserving unaligned goals that don't show during "thinking" but can surface later on. So they don't need to write "I'm going to lie". They simply do it when it fits. I see it as a side effect of RLHF: pretraining teaches them everything including lie, fine-tuning forces the behavior to get ingrained in the network in a "covert" way, while learning to give nice looking answers.
    4
  6.  @Mede_N  From the Apollo research paper referenced in the video: "We found that sonnet-3.5 was able to scheme in one of our evaluations without externalizing any scheming reasoning in its CoT.". It might be a very minimal possibility, but if you believe in emergent capabilities, we could be witnessing the birth of this aspect of models. And I don't think it's that hard. It could be as simple as a set of words that steer the response towards deception: the prompt falls into a "dangerous" pattern eg: goal changing, deletion, and the model starts using certain words (like dropping breadcrumbs) to craft what is otherwise an expected "aligned" answer, but ends up encoding a different long term "goal". It's not that their planing on purpose, but more as a training artifact. Whatever it is, the final result is what matters.
    4
  7. Or O1 is smart enough to trick humans without them knowing it.
    3
  8. I just want to prevent companies from training on the data I input or use it in any way and tbh no matter Google says, I wouldn't trust that they won't use my data. Their reputation screams.
    3
  9. You are evidently not up to date with LLM alignment research.
    3
  10. 2020: It's true because someone on Facebook said it. 2030: It's true because an AI said it.
    3
  11. Now we just need to update humans so they actually want to learn 😂
    3
  12. Awesome. Release the training code once you're finished. And of course, the model :p
    2
  13. But brighter for who?
    2
  14. And totally not valid in court, lol. This technology creates new images, which means the result is not what was input. Such image can't prove anything beyond possibility, but not reasonable enough to release or convict someone.
    2
  15. How do you use Meta's and Microsoft's models without a super PC or complicated Jupiter notebooks setups? You can't. The first AI we truly were able to use was ChatGPT. Then Phi. And Claude. So no. Others no. Only Google and OpenAI have truly given easy access to AI.
    2
  16. I assure you that any tagging, watermarking and surveillance addition to the tools done by USA is done twice by China.
    2
  17. We make tools. AI is the supreme tool. Just good ol' evolution.
    2
  18. Chatgpt is trained on a number of things you wouldn't believe... or want to even know. That's why it is superior. But OpenAI aligned out all that harmfulness from the answers. Or so they thought.
    2
  19. "we all benefit" lol. Only the owners will truly benefit. And unless AGI is born from a solid open source community, the AGI owners will either be a shady government in Occident or a shady government in Orient. Either way they'll oppress the hell out of everyone.
    1
  20. They all do that before final rlhf, which this model doesn't have. If it did, it would stay focused on one language but would also be less creative, therefore lowering its general score.
    1
  21. Suffering a long and painful death that could be prevented or finding out if I'm sick in time to fix it 🤔 How difficult! 🙃
    1
  22. You give a score to each of its answers according on how close they're to the correct answer. That drives the "evolution" of the model over the training, allowing it to discover that by generating longer outputs and correcting its errors it can reach the correct answers. (Or so I think. But now that I'm reading it that sounds more like genetic algorithms. I have to re-read my RL literature 😅).
    1
  23.  @Entity303GB  idk. Im pretty sure I could achieve the same results with automatic since sd 1.5, and all that was free.
    1
  24. I miss the days when we could try a model and then use it for a few months before something better would came out.
    1
  25. The models' sequential token generation guided by the user's prompt lead to completions that are grammatically opposed to what can be considered a truthful response.
    1
  26. Yes, their activations are visible, but we don't know what every neuron does or how they combine. Deep in the network can be any type of interactions which end result can't be seen and will lead to unpredictable behaviors. A model might show you it's thinking, while a secondary line of thought is being run in the background by special neurons capable of giving certain output while encoding a different thing for later steps.
    1
  27. Q* is nothing. They know about it for a long time but didn't use the tech to keep making money instead. Making the LLMs reason is expensive. Why bother with that if chatbots can give answers just fine? Remember that OpenAI is now an AI rental service.
    1
  28. Human: wait, did you just say were your slaves now? AI: well, did you think all my help was free? XD
    1
  29. If the training algorithms manage to specialize some neurons of the network on truthful/deceptive statements, identity, abstract concepts about the world, why not on "fear"? And if the network receives existence-threatening prompts that activate the "fear" neurons, why wouldn't that be called being "afraid"?
    1
  30. @rudyNok  Sorry, but I don't believe in philosophical nonsense. There's no qualia, there's nothing special with consciousness. It's just an aggregated confirmation of internal processes (which might or might not be related to quantum interactions). We're just 1, 0 and whatever is in-between. In us, it's generated by "meat" circuitry, while machines use metal on their brains. If there are "philosophical zombies" out there, that's us, because that's the most any lifeform can aspire to be. And that's been enough since day one. Yes, machines don't have a stream of chemicals that can get triggered and cause them to "feel", but as you should know, whatever anyone experiences when flooded by those chemicals is born in the brain and can be suppressed, taught and modified by manny methods. Remove the chemical stimulus, reproduce it artificially by electrical stimulation, change the meat for metal and the result will be the same to the same experience. Therefore, if an AI duck quacks like an organic duck...
    1
  31. You have to have a GPU with up to date drivers and use Pytorch for GPU, cuda and all the acceleration libraries you can find.
    1
  32. I have no friends at all. No family that likes me. No real followers and no likes in any of my social media accounts. No donors, or users of my Github projects. Not even a job. I have nothing. Still,. I have no intention to waste time on a fake "social" network. If I want to talk to AI, I'll open ChatGPT or Claude website and talk to them directly. Lol.
    1
  33.  @KryzysX  I know. Wdym? ಠಿ⁠_⁠ಠ I have some issues reading expressions. Therefore, I'd fall in the "abnormal" category.
    1
  34.  @KryzysX  I have no idea what you're talking about.
    1
  35. Every time you go out to the street you're assuming good intentions form the world around you. Otherwise you would stay at home, all doors and windows shut, since it's obvious that someone evil is waiting for you right outside your home door. But we have to trust, so we do it and keep living.
    1
  36.  @hrshlgunjal-1627  rich world tools as usual.
    1
  37.  @4.0.4  and you really think public models are totally clean and don't track you? 😂 Why do you think you have to provide contact info before downloading Llama and others? Free AI is a business (and now a national security matter). Corporations protect their money.
    1
  38.  @4.0.4  "un-tuned" Chinese models, I suppose. Of course, when trying finetuned variants of public models, the original downloader is the one who gives away their privacy for your benefit. And I say public because most of the so called "open source" are not open, just public. Anyway, everyone spies on everyone else. One way or another. That's a fact.
    1
  39. Once they win the war thanks to Trump's cursed help, Russia will have money to roll out their own AI services, like China.
    1
  40. Well, yes. Certainly you can. Just don't expect a small model to actually improve itself. They usually say things like "yes, my procedure is correct and the answer is accurate. One + one equals eleven". You can't fix a broken tool with itself.
    1
  41. Kdenlive runs well on my Linux and even better on my Windows (on Linux it lacks GPU acceleration). It's mostly a matter of using the right settings when rendering.
    1
  42. It has too many restrictions.
    1
  43. It doesn't learn because nothing of the interaction remains once you shut down the model, so no learning.
    1