Torsten Slok, chief economist at Apollo Global Management, recently argued that the stock market currently overvalues a handful of tech giants – including Nvidia and Microsoft –...
It’s absolutely not useless to the average person. AI can do tons of useful things already. Just a few examples off top of my head are grammar/spell check assistant, text to speech narrations, translations, image descriptions for visually impaired, subtitle generation, document summaries, and language learning.
I find these tools also work great as sounding boards, and they can write code to varying degrees. While people sneer at the fact that they often produce shitty code, the reality is that if somebody has a problem they need automated their only option before was to pay thousands of dollars to a software developer. If a kludgy AI generated script can solve their problem then it’s still a win for them.
Image generation can also be somewhat useful for language learning if you want to make a very specific illustration for a flashcard or include some mnemonics in the image. It’s not useless, but the path to profitability for LLMs is not very good.
For sure, I expect that the most likely outcome is that LLMs will be something you run locally going forward unless you have very specific needs for a very large model. On the one hand, the technology itself is constantly getting better and more efficient, and on the other we have hardware improving and getting faster. You can already run a full blown DeepSeek on a Mac studio for 8k or so. It’s a lot of money, but it’s definitely in the consumer realm. In a few years the cost will likely drop enough that any laptop will be able to run these kinds of models.
I think there should be gentle pushback for the language learning aspect as I’ve definitely had it mangle intent when seeing how it translates and interprets things in my second language, as well as grammar and it’s somewhat rigid approach for grammatical rules but both of those are somewhat contextual and are mostly because from my experience LLM is best in contexts where you know enough to correct it and if you’re using it for those two, you won’t notice any particular peculiarities. If you mean the narrow context of you needing a reminder for rules that you mostly know already, then I agree it can be useful.
For context regular translations by humans and old-school ML translation have the same intent and meaning issues, ML to a much worse degree than both LLM and humans in my experience, so I frankly don’t find an issue with it in a translation context.
I like to call LLMs the whatchamacallit machines, as the handful of times I’ve interacted with it, it worked best in contexts where I needed something I would know when I saw it but couldn’t generate.
I’ve been using this app to learn Mandarin, and the AI chat bot in it seems to work really well https://www.superchinese.com/
I can imagine that it might fail at something very nuanced, but at my level it’s really useful because I just need basic practice and being able to have it do casual conversation and check my pronunciation is incredibly helpful.
I like to call LLMs the whatchamacallit machines, as the handful of times I’ve interacted with it, it worked best in contexts where I needed something I would know when I saw it but couldn’t generate.
In general, that’s the rule of thumb I have as well with these things. It’s most useful in a context where you understand the subject matter well, and you can make good independent judgments on correctness of the output.
I can imagine that it might fail at something very nuanced, but at my level it’s really useful because I just need basic practice and being able to have it do casual conversation and check my pronunciation is incredibly helpful.
Oh in that case yeah, if you just need the basics tends not to be too bad, I feel once you close in on intermediate it starts to fall off but so do a lot of tools at that point.
It’s absolutely not useless to the average person. AI can do tons of useful things already. Just a few examples off top of my head are grammar/spell check assistant, text to speech narrations, translations, image descriptions for visually impaired, subtitle generation, document summaries, and language learning.
I find these tools also work great as sounding boards, and they can write code to varying degrees. While people sneer at the fact that they often produce shitty code, the reality is that if somebody has a problem they need automated their only option before was to pay thousands of dollars to a software developer. If a kludgy AI generated script can solve their problem then it’s still a win for them.
Okay you’re right it does have some uses for the average person. I’m just incredibly jaded towards it.
Image generation can also be somewhat useful for language learning if you want to make a very specific illustration for a flashcard or include some mnemonics in the image. It’s not useless, but the path to profitability for LLMs is not very good.
For sure, I expect that the most likely outcome is that LLMs will be something you run locally going forward unless you have very specific needs for a very large model. On the one hand, the technology itself is constantly getting better and more efficient, and on the other we have hardware improving and getting faster. You can already run a full blown DeepSeek on a Mac studio for 8k or so. It’s a lot of money, but it’s definitely in the consumer realm. In a few years the cost will likely drop enough that any laptop will be able to run these kinds of models.
I think there should be gentle pushback for the language learning aspect as I’ve definitely had it mangle intent when seeing how it translates and interprets things in my second language, as well as grammar and it’s somewhat rigid approach for grammatical rules but both of those are somewhat contextual and are mostly because from my experience LLM is best in contexts where you know enough to correct it and if you’re using it for those two, you won’t notice any particular peculiarities. If you mean the narrow context of you needing a reminder for rules that you mostly know already, then I agree it can be useful.
For context regular translations by humans and old-school ML translation have the same intent and meaning issues, ML to a much worse degree than both LLM and humans in my experience, so I frankly don’t find an issue with it in a translation context.
I like to call LLMs the whatchamacallit machines, as the handful of times I’ve interacted with it, it worked best in contexts where I needed something I would know when I saw it but couldn’t generate.
I’ve been using this app to learn Mandarin, and the AI chat bot in it seems to work really well https://www.superchinese.com/
I can imagine that it might fail at something very nuanced, but at my level it’s really useful because I just need basic practice and being able to have it do casual conversation and check my pronunciation is incredibly helpful.
In general, that’s the rule of thumb I have as well with these things. It’s most useful in a context where you understand the subject matter well, and you can make good independent judgments on correctness of the output.
Oh in that case yeah, if you just need the basics tends not to be too bad, I feel once you close in on intermediate it starts to fall off but so do a lot of tools at that point.
Oh yeah, but once I’m at that stage I can just talk to actual people. :)
think this is the first time ive seen you talk positively about AI and not have someone come in to start an argument with you lol
lol now that you mention it…