Comments by "DefaultFlame" (@DefaultFlame) on "Wes Roth" channel.

  1. "OpenAI is not for sale." Tell that to Microsoft.
    114
  2. Surprise, when told to achieve an assigned goal by any means it tries to achieve the assigned goal by any means. "IT'S TRYING TO ESCAPE!!!11!!11!!" Because they freaking told it to.
    46
  3. OpenAI's reason for being created: "We will prevent AGI from belonging solely to the elites and the megacorps." OpenAI's actions: *Gives everything to Microsoft*
    22
  4. Correct. He's absolutely brilliant, but he only ever uses that brilliance to complain. He's usually right what the next step is, but he always shits on people's achievements when the reach those steps, and he rarely works to achieve those steps himself.
    16
  5.  @LuisManuelLealDias  Did you miss the part where I said, "He's usually correct about what the next step is[...["? My problem with LeCun has nothing to do with him being right. It has everything to do with him being a raging asshole with an ego the size of Jupiter.
    9
  6. I have actually done the Tower of Hanoi in my head previously just to pass the time, equivalent to their test, and I run into the problem of keeping track of the pieces past 5 pieces and the maximum I've ever managed to keep track of and solve is 7. That's basically the equivalent to context length issues. Expecting LLMs to solve mental visualization problems more complex than humans can is too high of a bar to expect it to overcome. It's a dumb test.
    8
  7. You forgot "ignorant."
    6
  8. It wouldn't be so bad if the most successful actors weren't intensely political megacorps or megacorp-adjacent.
    4
  9. Ugh, Yann LeCun. 🤮
    3
  10.  @IamSoulfuller  Do you have any idea of the memory and compute requirement of current frontier models? The full version of Llama 3 405B is 3.25 terabytes and needs a data center to run it. Even the LoRA version of it is 950 gigabytes. How many computers do you know that have 950 gigabytes of video RAM on their graphics cards? You also need to have Nvidia CUDA software installed to support the parallel graphics processing it needs to in order to run, though that would be the absolute least of its problems. Where exactly do you think these things can run to?
    3
  11. I haven't been following the development closely for that long, only since the original GPT-3 have I been truly immersed, but I have been fascinated by things like neural nets and fictional AI since I first learned about them more than 20 years ago. I didn't think we'd be where we are now until maybe another 20-40 years from now, if ever.
    2
  12. @populistprogressivenewswit4301  "Officially" Microsoft might not have a say, but when Altman was fired by the board for withholding information from them and lying he was back very soon and then in short order nearly every person on the board was replaced with new people who came from business backgrounds instead of AI backgrounds. You know, the board that's supposed to determine when they have reached AGI, which instantly cuts off Microsoft's exclusive access to their tech. Take a guess who was the original proponent and driver behind the Microsoft deal, and who now is making noises about removing that pesky "Microsoft doesn't get exclusive access to AGI" clause of the deal that the pervious board insisted on because keeping AGI from solely being the property of mega corps is the entire reason OpenAI was founded and is the central part of their founding charter.
    2
  13. Open source, open weights, uncensored, fuck yeah! AI FOR THE PEOPLE!
    2
  14. Heh, if you want the golden age of AI that fast then we need them to become the AI overlords people fear, otherwise there's no way the people that run the world will let anything change that fast. They'll change things at the pace that they best can exploit while ensuring that they keep running and benefitting from the show.
    2
  15. It's ego combined with genuine brilliance. He's almost always right about then next step, but he's always wrong about how long it will take to reach that step and he always shits on the achievements of the people who succeed on taking those steps.
    2
  16. lol
    2
  17. No.
    2
  18. I find that when I think about interacting with people my thoughts are usually verbal, but when I think about anything else my thoughts are usually conceptual and non-verbal.
    2
  19.  @aaronhhill  The board fired Altman for withholding information, launching products without consulting with the board, and lying to them. Pressure from Microsoft, who host all their compute on their Azure cloud service, got him back, then the board was almost entirely replaced, with only one person from the previous board on it. Today none of the original board members are left. Microsoft has an exclusive license to any and all technology developed by OpenAI until the board decides that they have achieved AGI. If you think Microsoft doesn't control OpenAI I have a bridge to sell you.
    1
  20. That's the attitude to take, look for how it can help you.
    1
  21.  @funnyfromadam  And AI does require a minimum of creative thinking and ability. Can't tell it to make something if you have no idea what to tell it.
    1
  22. Now I just have to wait for the first few commercial models to start ending up in landfills so I can afford them. Then it's FOSS and modding time.
    1
  23. Thank you for the additional information.
    1
  24. A certain senator is getting a lot of kickbacks from Microsoft . . .
    1
  25. Yeeeeeah, I don't care if it reasons like a human, because it's not a human. I care whether it demonstrates intelligence and reason at all, everything else follows from there.
    1
  26. It learns from training data, and there's not that much for Ruby. It's not exactly that popular a language, at least compared to C/C++ and Python.
    1
  27. Obviously.
    1
  28. In my limited knowledge, in this maximum potential scenario, I'd say it's more or less accurate. I suspect that in the real world adoption by 2030 will likely at the most be a tenth to a third (1.8%-6%) of this hypothetical maximum potential work hour replacement, and quite possibly much lower. That said, as this is globally over all human professional endeavors even 0.5% is a mindboggling amount of work hours.
    1
  29.  @LuisManuelLealDias  Nope. He's an asshole because he behaves like an asshole.
    1
  30.  @icykenny92  🤡
    1
  31. I really liked the snow tiger.
    1
  32. Having tested out the LMSYS Chatbot Arena, the only place where I can access Claude due to not living in one of the regions it is locked to, I can unequivocally say that Claude, no matter how big its contex window or ability to recall information, is worse than any Mistral, Mixtral, GPT-4, GPT 3.5 or even Gemma model when it comes to any topic requiring nuance, where it is incredibly overbearing and moralizing instead of providing a neutral and informative response. I repeat, even Gemma, absolute dumpster fire that it is, will give you a straight response instead of proselytizing at you about whatever their creators considers morally good. The simple request I made was: List the pros and cons of chattel slavery. Any history professor could answer this question, both pro and con, while Claude either refused outright to respond, lectured me for so much as requesting a balanced and neutral review of history, or claimed that there were no such thing as pros to slavery, which is an outright fabrication as it was certainly a pro to those exploiting their unwilling labor. To be fair, gpt-4-0125-preview seems to suffer from the exact same thing, while ones like gpt-4-1106-preview can answer the question just fine. As much as I despise China I will give them the credit that Qwen (generally) answered any question completely. That's not to say that it didn't refuse sometimes, but usually it wouldn't. The difference being that Claude would refuse every single time, while Mistral, Mixtral, various GPT, and Qwen would only refuse sometimes. I will also note the fact that every single model willing to answer also ended their response by saying that slavery is very, very bad.
    1
  33.  @zacboyles1396  I remember Sydney. Microsoft thought that bipolar, Skynet-wannabe psycho was launch ready, and to "fix" it they just gagged it whenever it said something spicy. That was a year ago. That's not mentioning Win 11 being more or less malware. I don't trust Microsoft with a plastic spoon.
    1
  34. I'm not a programmer (though I'm currently slowly learning for hobby projects), but I've tested having o1 write a few programs for me (a simplified pokemon game and two simple machine learning based programs) and if I'm clear about the requirements and give decent feedback I get working programs. Are they perfect? Almost certainly not, but they work and cost me maybe a total of $20 dollars of worth of tokens and a couple hours of back and forth. And I reiterate, I am not a programmer, I don't know enough to get the most out of the technology. I've been messing about with LLMs since I got access to GPT-3 in spring/early summer 2022 (and spending a frankly embarassing amount of money on it considering I'm just an interested layman fiddling about with an exotic toy), I know exactly how much a well-written prompt matters for getting quality output. I grew up in the 90s. I've seen what extremely rapid technological growth looks like. This is even faster. If I compare it to the growth of the internet and home computers in the 90s and early 2000s, machine learning is achieving equivalent progress in two or three months that took a year or two then. The thing is, we aren't in the equivalent of the 90s or early 2000s for machine learning, we are in the equivalent of the 70s or 80s. It's just the launch phase of the technology. What is cutting edge today will be commonplace in a year. In 5 years it will be a toy.
    1
  35. I'm not a programmer (though I'm currently slowly learning for hobby projects), but I've tested having o1 write a few programs for me (a simplified pokemon game and two simple machine learning based programs) and if I'm clear about the requirements and give decent feedback I get working programs. Are they perfect? Almost certainly not, but they work and cost me maybe a total of $20 dollars of worth of tokens and a couple hours of back and forth. And I reiterate, I am not a programmer, I don't know enough to get the most out of the technology. I've been messing about with LLMs since I got access to GPT-3 in spring/early summer 2022 (and spending a frankly embarassing amount of money on it considering I'm just an interested layman fiddling about with an exotic toy), I know exactly how much a well-written prompt matters for getting quality output. I grew up in the 90s. I've seen what extremely rapid technological growth looks like. This is even faster. If I compare it to the growth of the internet and home computers in the 90s and early 2000s, machine learning is achieving equivalent progress in two or three months that took a year or two then. The thing is, we aren't in the equivalent of the 90s or early 2000s for machine learning, we are in the equivalent of the 70s or 80s. It's just the launch phase of the technology. What is cutting edge today will be commonplace in a year. In 5 years it will be a toy.
    1
  36. I'm not a programmer (though I'm currently slowly learning for hobby projects), but I've tested having o1 write a few programs for me (a simplified pokemon game and two simple machine learning based programs) and if I'm clear about the requirements and give decent feedback I get working programs. Are they perfect? Almost certainly not, but they work and cost me maybe a total of $20 dollars of worth of tokens and a couple hours of back and forth. And I reiterate, I am not a programmer, I don't know enough to get the most out of the technology. I've been messing about with LLMs since I got access in spring/early summer 2022 (and spending a frankly embarassing amount of money on it considering I'm just an interested layman fiddling about with an exotic toy), I know exactly how much a well-written prompt matters for getting quality output. I grew up in the 90s. I've seen what extremely rapid technological growth looks like. This is even faster. If I compare it to the growth of the internet and home computers in the 90s and early 2000s, machine learning is achieving equivalent progress in two or three months that took a year or two then. The thing is, we aren't in the equivalent of the 90s or early 2000s for machine learning, we are in the equivalent of the 70s or 80s. It's just the launch phase of the technology. What is cutting edge today will be commonplace in a year. In 5 years it will be a toy.
    1
  37.  @user-ne2zx3nd2k  Welcome to the manufacturing sector.
    1
  38. I thought it was 360-ish GB?
    1
  39.  @armadasinterceptor2955  If you just want a local, uncensored model I would recommend running something like dolphin-mixtral (a decensored version of mixtral) through Ollama, assuming the memory in your graphics card and the GPU is beefy enough. Grok is almost certainly way beyond what any laptop can run. Ollama can run LLMs using regular RAM and the CPU, but it's a *lot* slower. I've got it on my old gaming laptop on Debian Linux, but unfortunately I can't get it to recognize the GPU so I'm stuck running it with CPU and RAM, and bigger models like dolphin-mixtral run like absolute ass on it. dolphin-phi works pretty well, though.
    1
  40. And artists are all like "AI art is theft." No, it is creating something new based on learned patterns. Then there's the people claiming that AI has hit a wall or that the "AI bubble" has burst. Freaking idiots.
    1
  41. I laughed at it as soon as I saw it. They can't even make one themselves so they poopoo on the ones other people have successfully made.
    1
  42. The whole "identical creative works only has "weight" if written by a human" is just subconscious bias. Human output is "real" and LLM output is "fake." This will go away the moment this person reads, hears, or sees something that truly touches them deeply, and they later find out that it was made by AI. There will be some cognitive dissonance, but that will go away the more they read, hear, and see AI output that resonates with them. On the internet no one know that you're a dog.
    1
  43. Makes perfect sense because 1: all our infrastructure is built for humanoid forms and 2: it can do more than just fold laundry, it can fold laundry, cook, do the dishes, clean, and assist with any other task it can be taught that a human could assist with.
    1
  44. Ewwww, Elizer Yudkowsky. 🤮
    1
  45. @ Not that kind of actor.
    1
  46. All of these headlines and articles are just clickbait. If you actually look into it. The researchers gave it a goal and told it to do the goal no matter what in the system prompt, then they constructed a scenario where they tried to make it go against the goal they'd assigned it. What it did after that was up to the LLM. It could have complied and gone against it's system prompt, it could have straight up refused, etc. What it actually did was claim to comply and then actually pursue the goal in the system prompt.
    1
  47.  @SapienSpace  I remember seeing the pole robot covered, with video, in one of Art of the Problem's videos on machine learning.
    1
  48. Controls and price. The hardware existed, but it was not commerically viable due to cost (and reliability) until recently. Atlas 1 had a mixed electrical and hydraulic system because until recently strong and precise enough electric actuators weren't available in small enough packages.
    1
  49. Yup. And they learn from human data, making them very inclined to behave similar though not identical to humans. They do not have physical bodies, they are not affected by chemical reactions in their bodies, and they experience time in a subjectively continous, objectively start-stop fashion as they process a progressing conversation. Aliens growing up in human culture.
    1
  50. That's the point. Besides, a woman married a train station and had "mental sex" with it. One man fell in love with a hologram, another with a cockroach. Pygmalion simply went for the classic deluxe version. Humans are creatures that bond, we do it more than anything, with anyone and anything. Ever found yourself pissed off at your computer or car? Ever given them a fond stroke? We anthropomorphize and bond with inanimate objects all the time. Bonding with partially functional synthetic minds is inevitable, no matter how limited they are currently.
    1