Re: [TramsDownUnder] My first test of AI

On 3/4/25 12:15, 'David Critchley' via TramsDownUnder wrote:

Hey group,  

AI is certainly an interesting beast.  To say it is 'intelligent' is probably more fantasy than fact, but that being said it is certainly a helpful tool.  As John Cowper alluded to; the response is very similar to some of his work in his book on the new SLR lines.   This is because AI, in very simple terms, scours the internet for 'answers' that best fit the question or activity requested.

Remember that 'LLM' 'AI' is a statistical beast - it's just hovered up every bit of information it can find and just generates an answer statistically. A wildly incorrect piece of information that has been quoted multiple times with out any real checking will be weighted better than the sole properly researched correct information - as the wrong answer was seen multiple times and the correct one fewer or even only once.

And the citation problem is sort of baked in, the LLMs basically work by sucking down enormous amounts of information, actually citing it all would result in a list of citations many times larger than the actual information part of the response.

Any decent human researcher keeps copious notes of sources and should be able to cite sources and often have to make value judgements on the accuracy of those sources.

This has happened even locally in our hobby - SPER's freight car 24s - the number of people who said 'will never run, doesn't have motors' was considerable. Multiple people reported 'but it doesn't have motors' when discussion of operating it came up. Few actually went down and looked under the car - which if they had, they would have seen the rather unique bogies did in fact have motors.....