Anytime I ask it about things I’m knowledgeable about, it falls on it’s face. It only seems smart on topics I don’t know shit about (this red flag has spotlights on it!). Given this, I cannot trust it on anything lest I be led down false paths. Given how LLMs work, I don’t see this problem as fixable.
I have found them useful for search term generation. (Using a topic I know things about) If I need a more powerful drill for screwing lag bolts into hardwood and I’m not finding what I need, asking it and and getting back the keyword ‘impact driver’ is helpful, as it lets me go search for what that is and if it is the solution to my need for a more powerful drill. Note: I do not let it teach me about impact drivers (as it falls on it’s face all the fucking time), only use it to get the keyword to then use to search the internet.
To circle back to your question, I’m not scared with how fast they are advancing. I’m scared by how many people think they are good at everything and put them in places they don’t belong.
I have to use AI tools for my work, so arguably I am knowledgeable about that topic (programming). I have seen Claude code do pretty good work. HUGE CAVEAT it works well because we spent a lot of time building out our architecure and the really complex parts have an easy to follow pattern. It’s standing on the backs of giants.
If you do something like software engineering and you have a significant body of existing good work for it to pull from. Claude can be pretty exceptional.
If you do something like software engineering and you have a significant body of existing good work for it to pull from. Claude can be pretty exceptional.
the fresh grad new hires at my last place were able to code circles around me because of their embrace of ai and it made me think that i was already a dinosaur and became the biggest reason why i went back into IT.
It’s not great at breaking new ground though.
but watching the soon-to-be grads w a similar embrace fall flat on their faces repeatedly because the codebase at my new place hasn’t been updated in almost 15 years and was created by their fellow students, has made me realize that this is also true.
still though, if you know what you’re doing; ai is a pretty decent idea/concept generator and sounding board if all but one of your colleagues are students, so you don’t have anyone else to ask like it is in my present situation.
I expected it to get this one wrong and it did; I expected that as Hollywood portrays this wrong and people don’t have an intuitive understanding of missile rocketry. Interceptor missiles only burn for a few seconds, get going very fast, then coast to their target. It is just flat wrong here.
I then fixed my grammar mistake and asked again, and poof, 100% opposite answer. I literally got the OPPOSITE ANSWER just because I fixed a grammar mistake. This fundamentally cannot be trusted for actual learning.
I see. It definitely gets thrown off by the perceived confidence in the question, which is made to steer it toward a wrong response. It’s training data likely has far less instances of text that derails the question based on incorrect original assumptions.
Anytime I ask it about things I’m knowledgeable about, it falls on it’s face. It only seems smart on topics I don’t know shit about (this red flag has spotlights on it!). Given this, I cannot trust it on anything lest I be led down false paths. Given how LLMs work, I don’t see this problem as fixable.
I have found them useful for search term generation. (Using a topic I know things about) If I need a more powerful drill for screwing lag bolts into hardwood and I’m not finding what I need, asking it and and getting back the keyword ‘impact driver’ is helpful, as it lets me go search for what that is and if it is the solution to my need for a more powerful drill. Note: I do not let it teach me about impact drivers (as it falls on it’s face all the fucking time), only use it to get the keyword to then use to search the internet.
To circle back to your question, I’m not scared with how fast they are advancing. I’m scared by how many people think they are good at everything and put them in places they don’t belong.
Exactly. If it only “seems smart on topics I don’t know shit about” thats a massive red flag.
I’ll keep thinking for myself while the masses get dumber.
Hello Mr. Sure
I have to use AI tools for my work, so arguably I am knowledgeable about that topic (programming). I have seen Claude code do pretty good work. HUGE CAVEAT it works well because we spent a lot of time building out our architecure and the really complex parts have an easy to follow pattern. It’s standing on the backs of giants.
If you do something like software engineering and you have a significant body of existing good work for it to pull from. Claude can be pretty exceptional.
It’s not great at breaking new ground though.
the fresh grad new hires at my last place were able to code circles around me because of their embrace of ai and it made me think that i was already a dinosaur and became the biggest reason why i went back into IT.
but watching the soon-to-be grads w a similar embrace fall flat on their faces repeatedly because the codebase at my new place hasn’t been updated in almost 15 years and was created by their fellow students, has made me realize that this is also true.
still though, if you know what you’re doing; ai is a pretty decent idea/concept generator and sounding board if all but one of your colleagues are students, so you don’t have anyone else to ask like it is in my present situation.
How long ago did you try?! Have an example?
LLMs now search the web and compile results and info very fast. They do exactly what I’ve been doing for decades, searching and skimming results.
If you ask one “I need a more powerful drill for screwing lag bolts into hardwood” it’ll toss you a whole write up on things.
I expected it to get this one wrong and it did; I expected that as Hollywood portrays this wrong and people don’t have an intuitive understanding of missile rocketry. Interceptor missiles only burn for a few seconds, get going very fast, then coast to their target. It is just flat wrong here.
I then fixed my grammar mistake and asked again, and poof, 100% opposite answer. I literally got the OPPOSITE ANSWER just because I fixed a grammar mistake. This fundamentally cannot be trusted for actual learning.
I see. It definitely gets thrown off by the perceived confidence in the question, which is made to steer it toward a wrong response. It’s training data likely has far less instances of text that derails the question based on incorrect original assumptions.
Thanks for the response!