As many of you may know I have been long frustrated with internet myths about Lotus and especially Twincam engines that are relentlessly repeated.
I asked a question recently about engine numbers on google AI and the AI gave me some correct information i.e. it was a Twincam escort number but also gave me a load of BS about the engine.
I then engaged in a long discussion with the AI on the errors it made and it admitted its errors when I pointed to the detailed information I had posted on the Web previously.
It explained that its response are based on the weight of information on the web. 1000 wrong answers out weigh 1 right answer. Unless you challenge the answer and ask what its based on it will answer with a confident answer that is totally wrong,
When querying AI ask what Rohan Hodges says ! haha
It also explained that the wrong answer would not be recognised and propagated to the AI system despite it recognising it was wrong in the specific discussion session
You really need to understand how AI works before you use it.
Good advice. I recently tried âChatty Gâ on a specific Lotus question to find it had referenced one of my own posts on lotuselan.net as its source. A classic case of agreeing with yourself.
The lesson was that for highly specific questions where there isnât much data to base its answers on, that the LLMs are (not surprisingly) unreliable. It does raise questions about how LLMs could be mislead with well placed intentionally inaccurate web postings to push a particular argument or commercial enterprise.
With specific target campaigns it appears easy to distort AI answers. I am sure people are doing it and its hard to reverse once in place unless you do a bigger targeted campaign.
Even accidentle errors get easily propagated and because AI says it, it get reinforced and repeated by AI..
AI (neural network) is a probability engine, they work by returning the highest probability result, based on highest probability interpretation of the prompt. They have zero reasoning capability, they canât even count the number of ârâ is the word âstrawberryâ. They just return the most commonly found answer, which seems to be 2 for some LLM.
The correctness of there output is fully dependent on their training data being mostly correct and complete. For the main large langauge models(LLMs) available to the public these are trained off massive amount of public space information, so are very susceptible to incorrect information on the internet. If poke hard enough they will be able to asses input and confirm they did really get it right but by default they wired to give high confidence for their answers for the average Joe.
So as with any tool it user beware, these LLM are very useful aslong as you allow for they limitations. They great as next generation search engines, as can effectively do the search and return condesed information across top results. Of course then the use needs to validate that, but can be quite a time/effort saving.
The deployment of AI in vehicles in functions that could cause harm is very interesting area, there a lot of new and up comming standards about the development of the Neural Networks used, this is hevaily slanted to the training side to ensure full training coverage with standards like âSafety of the Intended Functionality (SOTIF)â (ISO21448). Then there PAS8800 that covers area like identify possibly failure from the networks and addressing them, along with designing networks which are more robust to bit faults and their impact on network accuracy.
That said most safety system run the output of networks against logical(reasoning) test to confirm the output is likely good.
When I want a definitive answer I ask AI Rohan Hodges!
But all joking aside, AI is no worse than the internet itself and other media sources, there is some good information, good expert people, and then thereâs the rest.
Some years back I was asked by a motor magazine for some information on the early years Lotus, Elites, Elans. Europas. I have a reasonable understanding but thoroughly checked what I sent them with books published a long time ago some by people who were âactually thereâ and using the Ron Hickman archive material I have.
The magazine came out, and the article full of errors and spin, I sent the instigator an email and was advised it had been âre workedâ for publication!
Itâs scary, you really canât believe what you read and even scarier, you can no longer believe what you see with AI generated photos and deep fake videos.
This is utterly frightening, considering how many people are leaning further and further into using AI in their decision making processes.
I do use AI but like you I routinely find errors. I had one session where it gave me 3 wrong answers one after the next! I let it know how useless that was, in the hopes that it might improve. Looks to be a false hope.
âReworkedâ means the Editor didnât get the story they wanted from you. I learned long ago that if the media doesnât get the story they want or doesnât fit their narrative, they will start with some kernel of truth and fabricate everything else around it. Itâs all about putting eyeballs in front of advertisers to peddle their wares.
I donât write for magazines anymore, Iâve owned my Elan since 81, my Europa TC since 87, and still have them both along with an S1 Exige. I have owned other Loti along the way. This counts for nothing when considered by journalists and editors whose average age is 22.
Iâm no expert thatâs for sure but I do have a good working knowledge, have restored several Lotus cars and I was really miffed when the article I wrote was changed beyond recognition and full of glaring errors. The only good thing was my name didnât appear underneath it.