I know how LLMs work - and you missed the context of "keywords". The current versions (in most free and even paid forms) can NOT read a "thread" like you posted unless you feed them the full text yourself. Instead they are aware of ReefCentral content that was pulled in on the last data pull, the thread title (that you posted) and whatever else related it knows about Reefing and/or the words in the title -- so it makes up a summary based on that, not what was actually written in the thread.
Ummm... it's very clear you don't know how this works lol. I think you're the one hallucinating with your explanations, or at least just trying to guess how it works .
ChatGPT can read any publicly available content, such as the posts in this thread, in their entirety. Thats why we can get full summaries like the ones above. However, the model is prohibited from reproducing that text verbatim for legal reasons. Another limitation is that grabbing all the data in the correct format can be tricky and you may get varying results from different web layouts. Also, I believe chatGPT-4o has a token limit for web plugins at 128,000 (for inputs and outputs) which equates to roughly 96,000 words.
You don't need to walk me through anything.
I thought this forum was all about learning lol. Is there nothing that can be taught to BeanAnimal?
Pick the right metrics and a banana can appear smarter than "human level intelligence". There is plenty of data out there showing how often it fails and to what extent and based on the intelligence of the user. You are aware of this though, if you are an LLM researcher. Yes, the breadth of information that it can pull in and reference is astonishing... but ignoring that it jumbles it all together and can't, itself tell you when or where it started hallucinating is scary.
With all due respect, your data is anecdotal. Thats why analysis on large sets of testing is more informative. Re you rejecting this data, its pretty convenient to find a way throw out any data you dont agree with, without providing a real critique or data of your own lol.
Earlier in this thread, there was an example of a Mn dosing calculation that was incorrect from ChatGPT. There was a typo in the input, i used the fixed prompt with a newer chatGPT and got the right answer.
The point I have made is that everyone needs due diligence. You're a respected member of this community, however, you are saying inaccurate statements with respect to AI (because you're clearly not an expert in this field). I might argue that's worse than an LLM outputting "StRawRRBeRRy" lol
I get it, you think AI LLM is the greatest thing in the universe, I don't. I think it is an extremely flawed and misunderstood piece of software that people are lulled into thinking is smart, sentient and can be conversed with intelligently -- to the point of the latest commercials begging people to befriend it. It is as, or more dangerous than it is good -- because it is almost universally and fully trusted even though it is capable and often blurts out insanely wrong or misleading crap and those driving it have no clue.
I think you're being very dramatic again and trying to misquote me to build a strawman argument haha. Im sure its easier to discredit me if you paint me as as a crazy fanatic; the reality is that Im interested in the potential of a new tool.
Any serious researcher in ML has to be honest with the limitations of models. I am more than willing to laugh at failures as well as marvel in successes. The main point is that this is a tool we can use (or misuse).
I think your resistance to AI is largely uninformed. You haven't discussed the effect of different models, prompts, areas of success or failures, even the distinction between an LLM and an agent. When was the last time you even tried to ask an LLM anything? These models progress so quickly it's impossible to make a blanket statement on how they perform, especially when you're misusing a model that's already 4 generations old.
Your rhetoric makes it clear your don't "want" to like AI. I don't know why, but you're entitled to your opinion. Maybe you are able to summarize this entire thread as an olde English tale better than I can?
