> [email protected] wrote: >> > [email protected] wrote: >> > 2. If they were perfect for what they are useful for, would they be >> > worth the fully-loaded cost including externalities? >> >> I'm still not sold on what they are "useful" for. Sure they can do some >> things, but their output is not deterministic. Well, not in any >> predictable way. It is technically "possible" to analyze an output based >> on the input and figure out how it got there, but the sheer amount of >> data >> and a lack of data and tools make it impossible on a practical level. >> This >> leads to the conclusion that it can not be trusted. >> >> What good is a computational or analytical tool that can not be trusted >> to >> produce accurate data? > > That's not a computational tool. It might be an analytical tool, > in the sense that we frequently see data that we want to impute > meaning to, rather than extract meaning from. LLMs are bad at > highlighting meaning but good at suggesting patterns that look > like meaning.
That's the problem, it is not seen as an analytical tool. It is seen as something that can write code, documents, and answer questions. You see it everywhere, people quoting ChatGPT as if that has *any* sort of authority. > > >> > It appears to me that the answer to the second question is >> > "rarely, if ever", so the answer to the first question is moot. >> > >> > "We should generally dismiss solutions which cost far more than we are >> > willing to pay." >> >> That is an interesting statement. I like it with one caveat, "willing to >> pay" is very subjective. > > Necessarily so. Extreme circumstances -- life and death -- make us > willing to do things that we would not countenance in a less exciting > situation. We may be willing to pay with a significant immediate risk > to our lives to protect our family or community, but not willing to pay > in mere inconvenience for the benefit of a stranger we deem unworthy. > > >> > You might argue that a subscription to Claude costs $17 per >> > month, but this is much like saying that the cost of your first >> > hit of meth is $5. >> >> Is it $5? lol, I have no idea. > > I looked it up. > https://www.addictionresource.net/cost-of-drugs/illicit/ seemed > a reasonable source to me. > >> > And people actually like meth. >> >> >> I imagine it is like alcohol. Night time drinker is hated by his morning >> self. :-) > > I'm just going by the economics: people pay for things that they like > when the price is worth it to them. > > The market in general rejects LLMs outside of software development. I'm not sure I'm seeing that. The LLMs are useful for a lot of things that are not "original." They are superb at doing "format" type of stuff. Think of it this way: I have a new compression algorithm that uses domain specific knowledge to compress 3D printing STL files. To create an application to use that algorithm is PITA code. That's perfect for LLMs and using an LLM to write the stuff that is really just typing is a good idea. The best way to explain it is in legal filings. Lawyers are starting to use LLMs to write briefs. An LLM can pass the bar because there are so many bar exams examples to use for training. When an LLM writes a brief, it can save a lawyer the cost of a paralegal or the time he would do it himself, BUT!! and this has real world examples, you should never trust the LLM to cite or reference actual legal precedents. > > > -dsr- > _______________________________________________ Discuss mailing list [email protected] https://lists.blu.org/mailman/listinfo/discuss
