Interesting if you are following AI evolution

[Maxone73]

Well, Google seems to have introduced what is considered the most advanced AI today (and it's open source), Gemini 2.5

I have always suspected that they would have made such a move sooner or later, they have the money, the knowledge (they theorised most of the concepts of modern models, including tensors on which all LLMs are based) and a nobel prize to prove it (with DeepMind), and, most important, the data.

It's rapidly evolving, in every direction.

[Justfor_]

It failed my trap question, as all other previously tested AI chats (ChatGPT, Deepseek, Perplexity). I ask them how many kgr/hr of diesel oil a burner nozzle rated in Europe is cross-specified in USgal/hr for the US market. The trap lies in that the nominal operating pressures in the European spec vs the US one are different (1000kPa vs 100 psi). All tested chatbots take for granted that the operating pressures are the same and perform a simple volume to weight rate conversion. A sharp engineer wouldn't fall into such a trap. At minimum, they would classified the question as ill defined because the pressure element was lacking.

[Maxone73 in reply to Justfor_]

That's not a failure in logic but the lack of a specific information. Ask it what are the operating pressures to see if it knows them first. In a, let's say, more democratic field like pure math, some AIs are outperforming all the human researchers in solving very difficult problem at a very high rate.

[Justfor_ in reply to Maxone73]

The basic rule is that if one doesn't have the information needed to provide a valid answer, either asks for it or declines answering. Making arbitrary assumptions is a no-no.

[Maxone73 in reply to Justfor_]

Yes, that a known limit of LLMs that they are trying to address

[KocoPr]

lol! Give them an answer even if it’s wrong. Sounds like a human