I respectfully disagree with the dead end part of your argument. A dead end would be if they provided no value.
While the environmental and social downsides are massive negatives on the tech, it is actually doing something.
Past iterations are completely useless, but more recent iterations show us a more polished side to LLMs that actually do enhance how we do some things.
Is it worth it? My gut says no, but its both too late and too early to call it. (late in the environmental and societal impact, too early in the tech iteration)
As far as the “dead end” argument goes, I have to say that’s a hard disagree. Humanity is filled with technological advances that “stand on the shoulders of giants” and improve on previous techs. Even if LLMs themselves don’t prove to be the thing that we’ve been promised by the people driving it, it is taking us one step closer to AGI (whether that’s a good goal or not, that’s still up for debate)
From here on, I think there’s still quite a bit these models can improve, and I hope a lot of that improvement goes into making it more energy efficient, more water efficient in turn.
If by a dead end you mean that we can’t reach an AGI from an LLM, I think that’s correct, however an LLM might help us figure out what is needed for an AGI.
If it was used in a research as a step? Perchance.
Pouring everything we have into it? Dumbest fucking decision of our lives.
We could have put all that effort into previous versions and could tweak them enough to gather perhaps slightly worse results, maybe even better, we will never know.
Making this shit more efficient is to me also dumb.
What in the fuck are we doing that requires this shit? It helps with coding? We can make better frameworks. Translations? We had those before, even TTS. Emails? Just use a template. The other side is not reading that slop anyway. So what exactly are we doing here?
You didn’t actually say what you think LLM’s are enhancing. Just that you feel that they are. Honestly I think that’s the biggest part, they’re big shiny things that look like they’re doing a lot. But they actually aren’t. LLMs are chatbots and they will never be anything more than just chatbots.
So much so, that it’s pretty much 100% necessary in software engineering now. And I hate it that I’m forced to use something that I know is so detrimental in other aspects.
I used it to make a dialog system in a video game. It made it, but it was needlessly complex and ten times as long as the code needed to be. No thanks, i don’t need a buggy mess that’s unmaintainable.
Yes, I’ve made a dialog system before. The context I found myself in was a game jam with a short amount of time using an engine I hadn’t used in years.
Thought it would help instead of following a tutorial. But honestly, by the end of the jam, I really didn’t feel like rewriting the dialog system bcz it was so messy.
LLMs are not chat bots, they do natural language generation AKA: they can produce human readble text, they can also parse text;
As of now, they take an input and follow patterns to guess what the output should be, it is really useful to be fair, they help in translation (see Deepl, a very good translator), they can take data and make it more readble to humans, summarize text*, parse text and data structures ex: i can give a JSON file to an LLM so i can get back a TOML file, document hard to read code etc etc
*but i’d argue that it’s rarely useful, you will hardly have to summarize a text for yourself because you usually need to know any detail in it but i can see someone needing a summary once
The fact that you think it’s bad at one thing in your list but adequate at the others is part of the problem. It’s bad at all of those things, because it’s a chatbot. Admittedly a very advanced chatbot, but still just a chatbot.
The most important take away here is what of your list was impossible before LLMs? Because the reality is that absolutely everything that you mentioned was possible before LLMs. All that LLMs have added is the chat interface part.
Granted, the technology that allowed LLM’s is likely to be very useful and already has been in places like protein folding, but that happened before LLMs.
I respectfully disagree with the dead end part of your argument. A dead end would be if they provided no value.
While the environmental and social downsides are massive negatives on the tech, it is actually doing something.
Past iterations are completely useless, but more recent iterations show us a more polished side to LLMs that actually do enhance how we do some things.
Is it worth it? My gut says no, but its both too late and too early to call it. (late in the environmental and societal impact, too early in the tech iteration)
As far as the “dead end” argument goes, I have to say that’s a hard disagree. Humanity is filled with technological advances that “stand on the shoulders of giants” and improve on previous techs. Even if LLMs themselves don’t prove to be the thing that we’ve been promised by the people driving it, it is taking us one step closer to AGI (whether that’s a good goal or not, that’s still up for debate)
From here on, I think there’s still quite a bit these models can improve, and I hope a lot of that improvement goes into making it more energy efficient, more water efficient in turn.
If by a dead end you mean that we can’t reach an AGI from an LLM, I think that’s correct, however an LLM might help us figure out what is needed for an AGI.
If it was used in a research as a step? Perchance.
Pouring everything we have into it? Dumbest fucking decision of our lives.
We could have put all that effort into previous versions and could tweak them enough to gather perhaps slightly worse results, maybe even better, we will never know.
Making this shit more efficient is to me also dumb.
What in the fuck are we doing that requires this shit? It helps with coding? We can make better frameworks. Translations? We had those before, even TTS. Emails? Just use a template. The other side is not reading that slop anyway. So what exactly are we doing here?
You didn’t actually say what you think LLM’s are enhancing. Just that you feel that they are. Honestly I think that’s the biggest part, they’re big shiny things that look like they’re doing a lot. But they actually aren’t. LLMs are chatbots and they will never be anything more than just chatbots.
Summarizing and finding codeblocks. Fucking A+.
So much so, that it’s pretty much 100% necessary in software engineering now. And I hate it that I’m forced to use something that I know is so detrimental in other aspects.
I’ve been a software developer for over 15 years, I’ve never used one. It’s not necessary at all.
And your 15 years are superior to mine because?..
I used it to make a dialog system in a video game. It made it, but it was needlessly complex and ten times as long as the code needed to be. No thanks, i don’t need a buggy mess that’s unmaintainable.
Note: I didn’t say - use it to code.
But real question for you. Is the alternative you wouldn’t have done that at all?
Yes, I’ve made a dialog system before. The context I found myself in was a game jam with a short amount of time using an engine I hadn’t used in years.
Thought it would help instead of following a tutorial. But honestly, by the end of the jam, I really didn’t feel like rewriting the dialog system bcz it was so messy.
LLMs are not chat bots, they do natural language generation AKA: they can produce human readble text, they can also parse text; As of now, they take an input and follow patterns to guess what the output should be, it is really useful to be fair, they help in translation (see Deepl, a very good translator), they can take data and make it more readble to humans, summarize text*, parse text and data structures ex: i can give a JSON file to an LLM so i can get back a TOML file, document hard to read code etc etc
*but i’d argue that it’s rarely useful, you will hardly have to summarize a text for yourself because you usually need to know any detail in it but i can see someone needing a summary once
The fact that you think it’s bad at one thing in your list but adequate at the others is part of the problem. It’s bad at all of those things, because it’s a chatbot. Admittedly a very advanced chatbot, but still just a chatbot.
The most important take away here is what of your list was impossible before LLMs? Because the reality is that absolutely everything that you mentioned was possible before LLMs. All that LLMs have added is the chat interface part.
Granted, the technology that allowed LLM’s is likely to be very useful and already has been in places like protein folding, but that happened before LLMs.