3 ms·

Grifters gonna grift- My nerd hype died as I used gpt4+ and Claude 3.5+ whatever for coding tasks and realized if a human somewhere in the training data hadn't

Grifters gonna grift- My nerd hype died as I used gpt4+ and Claude 3.5+ whatever for coding tasks and realized if a human somewhere in the training data hadn't already solved a programming problem/wrote code for a solution the LLM's are useless for said problem. They're useful as an assistant/junior coder to do repetitive tasks/things you/other humans already have already done tho. BUT that's not going to revolutionize programming as you already have to know how to code to use them. * and "hallucinations" should be referred to as "bullshitting" when it comes to LLM's as that's what it really is.

Mistletoe 2y ago
I was watching Alien last night and as they talked with Mother it was refreshing to see the ship computer just say “I don’t know” when it didn’t know something. If we could get our LLMs there it would be a nice step forward. You can definitely tell ours were made in Silicon Valley, home of the bullshit and fake it till you make it.
- prewett 2y ago
  I tried asking Gemini to summarize a non-existent poem, and it said it would need the actual source of the poem to summarize it (but provided four potential themes based on the title). I then suggested that it was a standard English course text and it apologetically said that it couldn't find any record of it, but perhaps I was thinking of this other poem? It is perhaps too helpful at coming up with a workable solution to my request, but it did seem to know when it didn't know. Although perhaps that isn't the LLM but some additional information from a search query being fed into the LLM.
anon291 2y ago
> My nerd hype died as I used gpt4+ and Claude 3.5+ whatever for coding tasks and realized if a human somewhere in the training data hadn't already solved a programming problem/wrote code for a solution the LLM's are useless for said problem. See my experience is the opposite, and gpt3 was able to answer questions about the niche Haskell library I wrote better than I was. Unless you posit that there's some secret cabal of people using this Haskell library and writing answers for GPT to consume somewhere in the depths of OpenAIs org structure, it's pretty obvious to me that LLMs are quite capable.

Mistletoe 2y ago

I was watching Alien last night and as they talked with Mother it was refreshing to see the ship computer just say “I don’t know” when it didn’t know something. If we could get our LLMs there it would be a nice step forward. You can definitely tell ours were made in Silicon Valley, home of the bullshit and fake it till you make it.

prewett 2y ago
I tried asking Gemini to summarize a non-existent poem, and it said it would need the actual source of the poem to summarize it (but provided four potential themes based on the title). I then suggested that it was a standard English course text and it apologetically said that it couldn't find any record of it, but perhaps I was thinking of this other poem? It is perhaps too helpful at coming up with a workable solution to my request, but it did seem to know when it didn't know. Although perhaps that isn't the LLM but some additional information from a search query being fed into the LLM.

anon291 2y ago

> My nerd hype died as I used gpt4+ and Claude 3.5+ whatever for coding tasks and realized if a human somewhere in the training data hadn't already solved a programming problem/wrote code for a solution the LLM's are useless for said problem. See my experience is the opposite, and gpt3 was able to answer questions about the niche Haskell library I wrote better than I was. Unless you posit that there's some secret cabal of people using this Haskell library and writing answers for GPT to consume somewhere in the depths of OpenAIs org structure, it's pretty obvious to me that LLMs are quite capable.