4 ms·

> “funny magic parrot box” can write some mid-tier high school essays That's your take on LLMs? Ask it how it is possible for a photon to travel across the un

> “funny magic parrot box” can write some mid-tier high school essays That's your take on LLMs? Ask it how it is possible for a photon to travel across the universe, arriving at the same time it departed, resulting in the journey taking zero time (in its reference frame). Ask what implications are if certain viral amino sequences result in messenger RNA translocating to the host cell nucleus, potentially with the entire genome. Ask if aircraft fly due to Bernoulli's Principle or Newton's Third Law and physical impact. This is "crypto-bro territory"? No, not quite.

wussboy 3y ago
The answers to all of those questions would be parroted from other research, and to the extent they were novel they would be incorrect.
sensanaty 3y ago
And it parroting answers back at you from textbooks and papers that are most definitely in its training data, probably with the identical wording you're using, is proof to the contrary of it being a "magic parrot box" as the other person put it? Or do you genuinely believe ChatGPT, a LLM, actually "came up" with these answers on its own?
jcgrillo 3y ago
The "crypto-bro" behavior I see is a whole bunch of people burning a ton of calories wildly casting about for industrial applications of what amounts to nothing more than a neat (albeit eye-wateringly expensive) toy. These LLMs seem like a solution in search of a problem in just the same way that blockchains are. Please prove me wrong, I'd really love to be wrong about this!
- msp26 3y ago
  Language models have completely overhauled the NLP space. If you have a problem involving natural language data, you can prototype working pipeline in an afternoon. Often this prototype is very close in performance to a 'proper' solution.
  - jcgrillo 3y ago
    > If you have a problem involving natural language data That's a big "if", isn't it? We're seeing claims like "The future is an LLM at the front of just about everything: “Human” is the new programming language"[1] but so far that's not panning out, and it seems really dubious. Natural language seems like an absolutely atrocious user interface. As a machine operator, I'm going to use levers, wheels, and buttons to control the machine. As a computer programmer I'm going to use programming languages to control the machine. I'm not going to speak English to it. So, ok, this marks an advance in NLP. How do we get from there to "omg it's gonna change everything!!!1111oneeleven" [1] https://techcrunch.com/2023/08/08/nvidia-ceo-we-bet-the-farm-on-ai-and-no-one-knew-it/ https://techcrunch.com/2023/08/08/nvidia-ceo-we-bet-the-farm...
  - FridgeSeal 3y ago
    I’ve been thinking a lot about this recently. It seems like they’ve accelerated our capabilities- previously tiresome and difficult-to-automate things are easier- but have done very little for our fundamental understanding. We have a tool, but cannot dissect it and explain how it fits together. LLM’a themselves don’t appear (happy to be wrong here) to actually have improved our understanding our NLP and associated theory. Yeah, it can parse a sentence and bang out some JSON/sql/mid-tier-essay, but these models (so far) aren’t helping us figure out how and why, and I think that understanding is critical to progress further. Anthropic seems to be trying to push a bit further on that front at least, but for all we know, they might just turn into another scummy OpenAI on us.
    - jcgrillo 3y ago
      I think in order for something to properly be a tool it needs to behave deterministically. I don't need to understand every particular of how it works internally, but as the user I need to be able to rely on consistent, predictable results. Otherwise it's worse than useless. Hand tools, machine tools, programming languages, vehicles, CAD/CAM/CAE tools are all like this. You may have to do some learning to become proficient in the tool, but once you're proficient in its use it's very unlikely to ever truly surprise you. Generally those "surprising" experiences are pretty traumatic--hopefully only emotionally (if you've ever experienced a chainsaw kick back you know what I mean). So I'm not sure how I could use an LLM as a tool, but maybe I'm just not a sufficiently proficient user? It seems like they're just too full of "surprises".

wussboy 3y ago

The answers to all of those questions would be parroted from other research, and to the extent they were novel they would be incorrect.

sensanaty 3y ago

And it parroting answers back at you from textbooks and papers that are most definitely in its training data, probably with the identical wording you're using, is proof to the contrary of it being a "magic parrot box" as the other person put it? Or do you genuinely believe ChatGPT, a LLM, actually "came up" with these answers on its own?

jcgrillo 3y ago

The "crypto-bro" behavior I see is a whole bunch of people burning a ton of calories wildly casting about for industrial applications of what amounts to nothing more than a neat (albeit eye-wateringly expensive) toy. These LLMs seem like a solution in search of a problem in just the same way that blockchains are. Please prove me wrong, I'd really love to be wrong about this!

msp26 3y ago
Language models have completely overhauled the NLP space. If you have a problem involving natural language data, you can prototype working pipeline in an afternoon. Often this prototype is very close in performance to a 'proper' solution.
- jcgrillo 3y ago
  > If you have a problem involving natural language data That's a big "if", isn't it? We're seeing claims like "The future is an LLM at the front of just about everything: “Human” is the new programming language"[1] but so far that's not panning out, and it seems really dubious. Natural language seems like an absolutely atrocious user interface. As a machine operator, I'm going to use levers, wheels, and buttons to control the machine. As a computer programmer I'm going to use programming languages to control the machine. I'm not going to speak English to it. So, ok, this marks an advance in NLP. How do we get from there to "omg it's gonna change everything!!!1111oneeleven" [1] https://techcrunch.com/2023/08/08/nvidia-ceo-we-bet-the-farm-on-ai-and-no-one-knew-it/ https://techcrunch.com/2023/08/08/nvidia-ceo-we-bet-the-farm...
- FridgeSeal 3y ago
  I’ve been thinking a lot about this recently. It seems like they’ve accelerated our capabilities- previously tiresome and difficult-to-automate things are easier- but have done very little for our fundamental understanding. We have a tool, but cannot dissect it and explain how it fits together. LLM’a themselves don’t appear (happy to be wrong here) to actually have improved our understanding our NLP and associated theory. Yeah, it can parse a sentence and bang out some JSON/sql/mid-tier-essay, but these models (so far) aren’t helping us figure out how and why, and I think that understanding is critical to progress further. Anthropic seems to be trying to push a bit further on that front at least, but for all we know, they might just turn into another scummy OpenAI on us.
  - jcgrillo 3y ago
    I think in order for something to properly be a tool it needs to behave deterministically. I don't need to understand every particular of how it works internally, but as the user I need to be able to rely on consistent, predictable results. Otherwise it's worse than useless. Hand tools, machine tools, programming languages, vehicles, CAD/CAM/CAE tools are all like this. You may have to do some learning to become proficient in the tool, but once you're proficient in its use it's very unlikely to ever truly surprise you. Generally those "surprising" experiences are pretty traumatic--hopefully only emotionally (if you've ever experienced a chainsaw kick back you know what I mean). So I'm not sure how I could use an LLM as a tool, but maybe I'm just not a sufficiently proficient user? It seems like they're just too full of "surprises".