I consulted with Claude.
You are here
AI
i believe AI now is more helpful with CML advice than 3+ years ago. I read AI's response now,the past year.
AI is not quite as helpful as a good doc or this CML support site.
AI bots scrape this site all the time. Quite a lot of the info it will regurgitate will be derived from what people have written here, and in other places of course.
I was a bit ahead of my time, and achieved my BSc in Artificial Intelligence in 2001 from the University of Birmingham, so I have the knowledge to be a little sceptical. Recently people have thought that date was a typo on my CV, but it’s true.
There are great use cases for AI, but for CML advice I think experienced patients will beat out an AI chatbot every time for everyday advice.
David.
I completed a 6-month AI and data science training at MIT Professional Education last year, since it's now a "must-have" for all IT support domains. The technical term for the so called AI bots is "context-conditioned pattern retrieval". It's an advance language processing program (LLM) which uses tokenisation to assign value to the words relative to the context in which these words are used in combination with powerful hardware - modern GPUs.
https://share.google/aimode/a6dCZACvRURZmCflM
LLMs are considered supervised machine learning, while unsupervised methods e.g. clustering are considered AI. The man behind modern LLMs Yann LeCun, who basically resurrected the engine behind LLMs (deep learning neural networks), says that scaling LLMs up (adding more processing power) will never lead to reasoning or AI.
https://share.google/aimode/k05UIyP1EUuopz1zM
The answers heavily depend on the questions (prompts) asked, that is, the more detail (context-rich) the question is the more relevant the answer you get. Accuracy is relative, it's best described with the term "AI hallucination", as the bots are not programmed to answer "I don't know" or "I'm not sure", so it seems they have to produce an answer. They are programmed to avoid certain topics, which is basically white-washing.
https://share.google/aimode/Meiq3qYYXLfpBeMoq
I've noticed that simply reshuffling the words in the question changes the answer. So you should always rephrase your questions multiple times and read the references to confirm that the answer is not completely off. I think they are useful for summerising multiple documents and pointing to certain direction, but you do need to verify it.
That's basically the level of engagement required to get meaningful answers.
However, the free public bots are trained only on general data, they are not specialised in anything. Businesses purchase such generic pre-trained bots and feed them with their own core-business data which is often proprietary and not available online so their internal and partner employees can use it for work.
Another issue is that free bots allocate limited resources to each user so the answers are not very rich in info. The best way to see it is to discuss something you know very well or you have thoroughly researched, then you will be able to spot the missing info. If you reply what you think is missing, it usually confirms and agrees and modifies or expands the answer.