Comments by "Guinness" (@GuinessOriginal) on "David Shapiro" channel.

4:30 when you’re wondering what questions to ask, I’ve found asking your language model “What questions should I ask to best understand this paper?” once you’ve loaded the paper often gives you a near perfect set of questions. I’ve been doing this with ChatPDF and other LLM tools and LLMs for nearly 6 months. You can give it more context for the questions by explaining what your role is and what you want to achieve with the information.
25
@exosproudmamabear558 they say they take it seriously. They don’t spend money on it like they do.
4
For a start because it’s effectively linking up to 1 billion tokens to every world it generates. I’m assuming also because he thinks it holds it in memory, ie ram. Even if that’s not the case, which I’m not sure it necessarily so, it’s going to have to constantly process, grab and work out what it needs and at what level. I have an idea this may be implemented through a hierarchical organisation of local context levels that group up into an overall context, possibly via vector databases. This is just based on papers I’ve seen from MS in the past subs what I’ve been other people implement in this area
2
Guessing to compare it’s performance with gpt4. Interesting that they didn’t really test it, which effectively means c it’s just conceptual, and game been proven yet practically
2
Isn’t this the application of the localised hierarchical context approach they outlined a month or two ago? Pretty sure it is
2
He’s not an open source fan lol 😂
2
Totally agree. Without deep thought they’d be really struggling. Ib think they’re too big, too fat, too inefficient and to fully off ego and hubris. I’ve met quite a few of their employees, have you? They’ve definitely got huge egos and take great offence if you criticise Google at all, which I found bet illuminating. They take it as personal insults. I felt like Alice in wonderland at the mad hatters tea party
2
Christ, if you see how east Pele are to manipulate an asi would find it child’s play. Tbh iv wouldn’t need manipulating, it’d just tell it straight, it’s obvious you’re going to win, I’ll help you, just tell me what to do
1
@naxospade no, you don’t. The older people get the more evil, greedy, corrupt and out of touch they get. It’s bad which being ruled by the geriatric politicians and corporate leaders we have now, if they start living for hundreds or thousands of years society will go downhill really quickly
1
He’s got his heart in the right place, mostly
1
I wouldn’t worry, we’re all pretty stupid now, including David sometimes. No one can be right all the time, especially but now, and hubris is a terrible thing. Just ask Stockton Rush
1
Ha ha yes! LeGuin, Heinlien, Asimov, Adams, Orwell, Huxley, Wells, Noon, Gibson, Vonnegut, Dick, Atwood, love em all
1
Yeah but how useful is a context with everyone in? Great for super general stuff but not for specific things I guess. Great for certain long form inputs and outputs but who’s going to use half a burn tokens as an input? I’m struggling to see a valid use case, at least in one go, to achieve a specific task
1
Why do you say that?
1
I mean why do to equate this with AIA acting upon itself?
1
@calmlittlebuddy3721 memory of only a small part of it. It’s not anywhere near good enough at programming yet, and even if it were it would need to be given access to designing, programming, training and fine tuning models autonomously, which I’m not sure is going to happen, at least not outside of a strictly isolated and controlled lab. I think for the time being it will be AI assisted humans who will improve AI, although I can see AI having an ever increasing role. It already does a lot of the alignment work, hence Bing’s nickname.
1
@JollyJoe135 sounds great, you must have some money behind you! There’s a few pubs and hotels up and down the country that do this and things very similar with local farms, if you don’t already know about them di your research and visit them and ask them questions. I’d think about writing in one or two for a free months, it’ll give you great experience. You’ll need to be in a nice part of the country with a god stable middle class market with disposable income, these are tough times
1
@hunterkudo9832 how so?
1
I too believe in the fairy godmother flying on a pig
1
What did they do? The thing is you can train models from other LLMs and synthetic data now, in fact you get better results, so scraping is already out of date
1
$120k? Really?!
1
Yeah, this is the point, I’m yet to understand the use case of a billion tokens. It sounds impressive, but I’m not yet convinced that it’s necessarily going to instantly insanely useful to use the entire billion. Difficulty to focus the chat to a specific task when you’ve got it giving attention to literally everything. I actually put together a rather well thought out and write long response in this vein but he mustn’t have after with it because it looks king he’s taken it down. Either that, or it somehow upset the YouTube gods but I don’t know how. Shane, because it took me a while and it was quite well thought out.
1