It seems to me that the main point is not the best term according to a mathematician … who is likely to want a paragraph or two rather than a single term … but whether or not a particular term encourages an accurate and meaningful concept of the tool … and, since formula seems to encourage this idea of direct adjustment of weights, then formula must not be the best concept.
Overall, I agree with you. Prior discussion I thought was very confusing. This may be a naive view of mine, but it seems that how a neural net is built and the math used to populate/update it is much more relevant than that associated with any subsequent processing. Wading through layers of a neural net when processing seems to be very similar to the way my former company’s software, managed distribution of production and revenue transactions among what at times were thousands of miles of gathering systems with connections to hundreds or even thousands of wells.
That may be a bad analogy, but still it seems conceiving the design and training a neural net is a way way harder problem than any later usage of it. And I find usage is still a big problem until someone figures out a clear way to eliminate at least most of errors that are due to its probabilistic nature.
This is probably naive of me - especially considering the cost of what I’m about to say - but airliners have triple redundant (or more) computers which cross check each other and “vote” on the most important actions before allowing something to proceed; maybe there should be multiple “answers” which a third system checks (if the first two are substantially different) and then produces for the end users.
I understand that creating three AI agents (with different training, data sets, etc) would be vastly expensive, but I’ll bet the company that comes out with one that doesn’t hallucinate or give really bad answers will be the one that wins. That’s a big part of what made Google triumph over lesser search sites.
Depends on how much more expensive it is - and what it means to “win.”
We know that you can create some level of autonomous passenger vehicle as TaaS - Waymo’s doing it in ten cities - but we don’t yet know at what scale and cost that system will settle down at. If that type of TaaS is cheaper than human drivers (Ubers and taxis) it can scale to take over that market. But if it’s not materially cheaper than human drivers, it’s not going to grow much beyond that.
The dream is to have a Level 5 vehicle, or a Level 4 vehicle that’s robust enough to function outside of a TaaS network. But we don’t even yet know if that’s possible with anything close to existing tech, or what such a car would cost.