I think the best way to understand AI is that it wasn't created for absolute honesty nor the premise of accuracy.
If I may quote TARS from Interstellar : Absolute honesty isn't always the most diplomatic nor the safest form of communication with emotional beings.
Also most people don't ask questions seeking honesty (looks over at the Eww Ass Ay and the stark and widening gulf between Democrat and Republican), they're seeking validation. However for the sake of argument lets pretend someone is asking a technical question (from a positoin of ignorance) and is seeking the truth. My experience is that you'll get that 90% honesty that TARS cited.
So I asked a very specific question about EV charging (a subject I happen to know a bit about)
Claude went on to say how it doesn't cause any harm ... however I know differently (from decades of working with batteries), and I knew this was partially wrong
you see it only works in the case where you have not avoided the most critcal balance window listed in point 1. above.
If you go on charging at 80% (which all the EV makers seem to suggest in their advertising you should do) you'll hit problems.
So after I was assured by Claude that it was all OK, I pushed back with the quesion below. The answer is worth understanding and is why Claude should have said "it can't prevent balancing issues" when it first answered my question.
But it didn't and if I didn't know enough to push back it would have misled me. Aside from the "acceptance rate cliff" he mentions there's also the very likely outcome of "inaccurate range estimation because that only comes to light as you discharge (on the highway or taking off from the lights) and one "cell" in the pack buckles down more and you go from having enough range to "we need to charge".
As to how we'll get around this constraint I'm not sure, because all models seem to focus on making you happy (not informing you).
