Ideally, the whole point of an LLM is to create a problem that can interpret human language so that we can interact with software in a "human" way. It shouldn't matter if the data is garbage, as long as the result is a program that understands english.
Once you have that, you can point it to a random pile (list of websites, source code repository) of information that you know is mostly decent, and it will go through the long and painful process of dissecting it for you. This only works if the LLM is capable of doing this without mixing in all of its childhood memories. You know, like we do - we all spent a lot of time engaging in really stupid conversations that were only intended to practice listening and speaking. As adults we throw the subject matter from that phase away, retaining the grammar and sentence structure. We don't really care about that time a fox jumped over a lazy dog. I've actually been fiddling with package management for ROCm in Slackware so at some point I'm going to try and fire up some sort of ML or LLM app to verify that it all works. At that point I'll be taking a look at the various options for building your own model since I'm working on a project that would be an interesting use case for LLMs. -Ben On Friday, March 22nd, 2024 at 5:57 PM, American Citizen <website.read...@gmail.com> wrote: > Okay, I get it, that AI is as only good as the input and if it is fed > garbage, it can only spout garbage > > (wish that there was someway to clue the investors into this fact) > > > On 3/22/24 17:51, Russell Senior wrote: > > > On 3/22/24 17:39, Ben Koenig wrote: > > > > > On Friday, March 22nd, 2024 at 5:04 PM, American Citizen > > > website.read...@gmail.com wrote: > > > > > > > A few years ago, I took my Linux OS which is openSuse Leap v15.3 or so > > > > and ran a check on the documentation such as the man1 through man9 > > > > pages > > > > (run the %man man command to pull all this up) versus the actual > > > > executables on the system. > > > > > > > > I was surprised to find < 15% of the command executables were > > > > documented. Naturally I was hoping for something like 50% to 75%. > > > > > > > > If I am going to talk to an AI program, such as ChatBot or one of the > > > > newer popular AI program and ask it to generate the documentation for > > > > the complete OS, what AI chatbot would you choose? > > > > > > > > My idea is to clue the AI program into the actual OS, then ask it to > > > > finish documenting 100% of all the executables, or report to me all > > > > executables which have no available documentation at all, period. > > > > > > > > This means the AI program would scour the internet for any and all > > > > documentation for each command, and there are 10,000's of > > > > executables to > > > > examine. (which is why I believe this is an AI task) > > > > > > > > Your thoughts? > > > > > > > > - Randall > > > > That would be an interesting experiment to see what it comes up with. > > > > I would question the results simply due to the quality of current LLM > > > > implementations. > > > > > > From recent anecdotal experience, I recently bought an expensive > > > Logitech keyboard and it was behaving strangely so I tried to look up > > > how to perform a "factory reset" for this model. The search results I > > > found via DDG were interesting, there were multiple duplicate hits > > > for what appeared to be a tech blog with generic instruction pages > > > for my device. However there were multiple iterations of this page, > > > for this keyboard model, each of which had instructions referencing > > > physical features that do not exist on this actual keyboard. These > > > appeared to be AI generated help pages that were clogging up actual > > > search results. They were very well written, If I hadn't had the > > > actual device in front of my I might have actually believed that > > > there was a pinhole reset button next to the USB port. > > > > > > If you do this, you may need to find a way to define a "web of trust" > > > that allows the AI to differentiate between human written articles, > > > and AI written summaries. As it is right now, you might find yourself > > > telling an AI to summarize help pages that are AI written summaries of > > > AI written summaries of ( > > > AI written summaries of ( > > > AI written summaries of ( > > > AI written summaries of (actual manuals) > > > ) > > > ) > > > ) > > > > > > Recursion FTW! :) > > > > It seems inevitable that the AI serpent will stupidly eat its tail and > > devolve into even more of a stochastic septic tank than it is now. If > > I was an investor, I would be shorting hard into the AI bubble. To me, > > the only open question is whether humans get stupider faster than the > > machines.
