LLMs are complete trash with no viability in the short or long term.
Some of the traditional ML stuff is useful (classifiers etc).
That's where it is.
LLMs are a tool. Use them correctly, and they can be highly beneficial. Some people are content to use a shovel and dig a large hole, but maybe it would be faster to use a backhoe loader. That's what LLMs can be like in some situations -- upgrading from a shovel to a tractor.
You can ignore them all you want to, but they literally save me months of work for each project I work on. Granted, I mostly use them for scripting/coding (I'm not a computer scientist, just a normal Ph.D.-holding scientist), but they help me code my scientific work and statistical analyses. What I do is highly study- and analysis-specific. I need to create my own workflows using multiple validated external applications. These pipelines/workflows will likely only be used one time ever (by me) because they are specific to my data. This means that while I incorporate what others do and have produced, there is a lot of unique code to produce to use my data.
When produced, all final code is publicly available on my GitHub account for independent verification during the peer review process and after publication. All statistical analyses are also independently run by a statistician in SAS. To date, with my LLM-helped R code and her 'traditionally' written SAS code, we've had the same results 100% of the time. We've even pulled in a 3rd person a few times to also run independently using other software, and they get the same results. That's not necessary to do because I'm using standard R packages and statistical approaches, but I'm a stickler for validation.
My other generated code is primarily Bash and Python and uses independently developed and validated research tools (as I mentioned previously) within the scripts. Before LLMs I figured the scripting all out on my own, which was fine, but it took days to weeks to months to do web searches, dig through Stack Exchange, and contact the developers of code (rarely needed to do, but it happened). Now I can produce much more elegant code without having to do most of that in a fraction of the time. Why don't I hire someone to code for me? I wouldn't mind doing that if NIH or my university gave me money to do that. They don't, so I get to do it myself (which I enjoy anyway but I also enjoy being much more efficient).
I was talking with a colleague who has also been impressed with LLMs. He mentioned his multi-year postdoc was almost entirely coding some specific analyses. Now, with LLMs, he could do the same thing in a couple weeks.
Is what we do with LLMs "complete trash"? Not remotely. It's better and even more reproducible than what we were able to do before because the LLMs help make the code more generalizable.
Broadly gesticulating and saying, "LLMs are complete trash" is simply wrong. Maybe there's a more tactful way of saying that, maybe they are worthless to you, but to me and many other people, they are life-changing for the better. If you don't like them, don't use them. I'm getting more and better work done now than I ever was. And no, the LLMs are not "hallucinating" data. My use for them is to accelerate my coding. I know the input and output. I'm not generating new data with LLMs, I'm scripting more efficiently.
