Comment on The Generative AI Con.
Greg@lemmy.ca 1 week agoIt is the best option for certain use cases. OpenAI, Anthropic, etc sell tokens, so they have a clear incentive to promote LLM reasoning as an everything solution. LLM read is normally an inefficient use of processor cycles for most use cases. However, because LLM reasoning is so flexible, even though it’s inefficient from a cycle perspective, it is still the best option in many cases because the current alternatives are even more inefficient (from a cycle or human time perspective).
Identifying typos in a project update is a task that LLMs can efficiently solve.
sem@lemmy.blahaj.zone 1 week ago
Yes I think it’s a good option for spell check, or for detecting when the word it sees seems unlikely given the context.
For things where it’s generating text, or categorizing things, It might be the easiest option. Or currently the cheapest option. But I don’t think it’s the best option if you consider everyone involved.
Greg@lemmy.ca 1 week ago
Can you expand on this? Do you mean from an environmental perspective because of the resource usage, social perspective because of jobs losses, and / or other groups being disadvantaged because of limited access to these tools?
sem@lemmy.blahaj.zone 6 days ago
Basically the LLM may make people’s jobs easier, for instance someone can get a meeting summary with less effort, but they produce worse results if you consider everyone affected by the work product, like considering whose views are underrepresented in the summary. Or, if you’re using it to categorize text, you can’t find out why it is producing incorrect results and improve it the way you could with other machine learning techniques. I think Emily Bender can do a better job explaining it than I can:
m.youtube.com/watch?v=3Ul_bGiUH4M&t=36m35s
check out the part where she talks about the problems with relying on LLMs to generate meeting summaries and with using it to clarify customer support calls as “resolved” or “not resolved”. I tried to get close to that second part since the video is long.
Greg@lemmy.ca 6 days ago
I agree and I think this comes back to execution of the technology as opposed to the technology itself. For context, I work as an ML engineer and I’ve been concerned with bias in AI long before ChatGPT. I’m interested in other folks perspectives on this technology. The hype and spin from tech companies is a frustrating distraction from the real benefits and risks of AI.