The LLMs do have "latent knowledge," indisputably, the latent knowledge is beyond reproach. Because what we do know about the "black box" is that inside it, is a database of not just facts, but understanding, and we know the model "understands" nearly every topic better than any human. Where the doubt-worthy part happens is the generative step, since it is tasked with producing a new "understanding" that didn't already exist, the mathematical domain of the generative function exceeds the domain of reality. And, second of all, because the reasoning faculties are far less proven than the understanding faculties, and many queries require reasoning about existing understandings to derive a good, new one.
*or any digitized proprietary works, just as long as they can be parsed correctly. don't worry, the means of how to optain these works doesn't seem to matter[0]
I claim it's normal to hate public transport. Online, there are some loudmouthed public transport enthusiasts. To them, everyone who isn't doing public transport is a racist, boomer, redneck, luddite, and whatever aspersion you've got.
The real reason America has so many cars is people like cars better, and America developed in a time where people were rich enough to make it happen. People don't like public transport. I asked someone who grew up in another country, in a huge city with only public transport--and reputedly good, clean public transport at that--what they think of public transport, and they said it's gross and for poor people. (It wasn't a code for racism, their country was ethnically monotone.)
People like that don't visit threads like this though. You just get this echo chamber of young, childless, cosmopolitans who only care about a certain kind of efficiency in transport.
I think it's because software engineers are the only group that can unanimously operate LLMs effectively and build them into larger systems. They'll automate their own jobs first and move on to building the toolkits to automate the others.
It's the other way around. The model is impeccable at "understanding text." It's a gigantic mathematical spreadsheet that quantifies meaning. The model probably "understands" better than any human ever could. Running that backwards into producing new text is where it gets hand-wavy & it becomes unclear if the generative algorithms are really progressing on the same track that humans are on, or just some parallel track that diverges or even terminates early.
Only if you wildly oversimply to the level of being misleading.
The precise mechanism LLMs use for reaching their probability distributions is why they are able to pass most undergraduate level exams, whereas the Markov chain projects I made 15-20 years ago were not.
Even as an intermediary, word2vec had to build a space in which the concept of "gender" exists such that "man" -> "woman" ~= "king" -> "queen".
3 lines? That's still going to be oversimplifed to the point of being wrong, but OK.
Make a bunch of neural nets to recognise every concept, the same way you would make them to recognise numbers or letters in handwiting recognition. Glue them together with more neural nets. Put another on the end to turn concepts back into words.
... Oh interesting. And those concepts are hand picked or generated automatically somehow?
> For a less wrong but still introductory summary that still glosses over stuff, about 1.5 hours of 3blue1brown videos
Sorry, my religion forbids me from watching talking heads. I'll have to live with your summary for now. Until I run into someone who condensed those 1.5 hours in text that takes at most 30 min to read...
> Oh interesting. And those concepts are hand picked or generated automatically somehow?
Fully automated.
> Sorry, my religion forbids me from watching talking heads.
What about professional maths communicators who created their own open sourced python library for creating video content and doesn't even show their face on most videos?
You're unlikely to get a better time-quality trade-off on any maths topic than a 3blue1brown video.
He's the kind of presenter that others try to mimic because he's so good at what he does — you may recognise the visuals from elsewhere because of the library he created[0] in order to visualise the topics he was discussing.
Simplifying to that point is more of what a Markov chain is. LLMs are able to generalize a lot more than that, and it's sufficient to "understand text" on a decent level. Even a relatively small model can take, e.g. even this poorly prompted request:
"The user has requested 'remind me to pay my bills 8 PM tomorrow'. The current date is 2025-02-24. Your available commands are 'set_reminder' (time, description), 'set_alarm' (time), 'send_email' (to, subject, content). Respond with the command and its inputs."
And the most likely response will be what the user wanted.
A Markov chain (only using the probabilities of word orders from sentences in its training set) could never output a command that wasn't stitched together from existing ones (i.e. it would always output a valid command name, but if no one had requested a reminder for a date in 2026 before it was trained, it would never output that year). No amount of documents saying "2026 is the year after 2025" would make a Markov chain understand that fact, but LLMs are able to "understand" that.
Wrong, wrong. Opposite of everything he said. All his examples are backwards. The article is basically inversing the Single Responsibility Principle.
First of all, consistency does not matter at all, ever. THat's his main thesis so it's already wrong. Furthermore, all his examples are backwards. If you didn't know the existence of "bot" users, you probably don't want your new auth mechanism to support them. Otherwise, the "nasty surprise" is the inverse of what he said: not that you find you don't support bot users, but you find out that you do.
Build stuff that does exactly what you want it to do, nothing more. This means doing the opposite of what he said. Do not re-use legacy code with overloaded meanings.
> Build stuff that does exactly what you want it to do, nothing more
This is also confusing to me. In a multi-million line codebase, it's extremely difficult to find an actual place where you have zero side effects with ANYTHING you write.
Wrong. If code is written consistently everywhere, that allows any dev to dive in anywhere to get work done. Which is what you often have to do in large code bases to make cross functional updates.
Code bases where devs pick a different framework or library for every little thing are a nightmare to maintain. Agreed on standards is what gets your team out of the weeds to work on a higher and more productive level.
The phrase 'mental illness' has expanded as a category a thousand-fold this millennium. I see no evidence he has any mental illness, other than a particular kind that is fully voluntarily and reversible.
Same, had to call around a lot to find primary care, and was being given multi-month waitlist estimates for seeing an ENT specialist. I've had more luck recently as I was able to get into see an ENT in less than 30 days. It's also crazy how much everyone tries to upsell you. It' hard to tell what tests or procedures I really need.
reply