An interview with Kazuto Yamaki of Sigma that appeared in LensVid raised an interesting point. I wonder how many people caught it.
"Sigma doesn’t use AI to design lenses, for now. Mr. Yamaki said the knowledge required to make a high-quality lens is not available publicly; it is proprietary and kept inside the company…”
As I’ve written a number of times, AI as being practiced at the moment is backwards looking, and the big AI companies have been scraping pretty much any information they can find anywhere. Public data, even Copyrighted or otherwise protected data (e.g. license agreements) is considered fair game to the Silicon Valley AI Vacuums.
What Yamaki-san is referring to here is a fear that if they used one of the existing large language models to help with optical designs, in so doing they’d also expose the designs that they’ve been doing and probably the reasons for their decisions, too. But I don’t understand why they haven’t RAGed their own data in a private model (RAG = Retrieval-augmented generation). We’re about to discover who the truly tech-savvy companies are and which ones are meandering around just trying to make sense of technology changes.
There’s nothing stopping an organization from running their own local AI that is trained on their own proprietary data, and trained in a way that doesn’t get communicated back to the Silicon Valley Overlords. We’re going to see companies that get this (Apple is one), and companies that don’t (the Sigma comment would indicate that they currently don’t).
So the question that comes up here is this: what would having a RAGed local AI do for Sigma? Well, it would help their designers dealing with optical issues. How? Such a system would have memory of how certain problems were tackled in the past and what the tradeoffs and outcomes were. In other words, institutional knowledge that is passed down to new designers, but more quickly and precisely.
I should point out that one reason why I can answer a lot of Nikon questions quickly via email is that I have begun RAGing much of my 40 years of writing about Nikon in a way so that I don’t always have to pull it out of my aging brain neuron by neuron. The local AI tends to confabulate a bit less often than my wetware does.
Which brings me to a slightly off-topic point. Mainstream media has been writing a lot lately about how OpenAI is harming the established math world by solving (or attempting to solve, the results are still a bit unclear) as many of the outstanding math questions as possible. The media seems to think the AI engine in question, ChatGPT, did those proofs all on its own. Nope. AI is directed. By someone. At something. It’s unclear why OpenAI felt that they needed to tackle outstanding math proofs, though it was probably about generating bragging rights for their reasoning engine. The implication would be that if their model can solve things that haven’t been solved to date, then their AI must be better than any other at structuring conclusions from existing data.
But that’s exactly the reason why a savvy tech company would be developing their own internal AI, whether by RAG or by directed training. There are unsolved optical problems, just like there are unsolved problems everywhere you look. So getting back to why I’m covering this one comment: if you don’t use all the tools available to you to move your knowledge base forward, what happens when one of your competitors does?
That’s actually the real danger of AI at the moment: who’s directing it at what? I’ve found the currently available AI—even the LLM that I run on the Mac I’m typing this on—to be an accelerant to what I know and what I’m doing. Almost too much of an accelerant, as I’ve now started way more projects than usual, which is slowing down some of my usual output (e.g. lens reviews; sorry, I’ll fix that in the not too distant future). But it’s also accelerating my knowledge of the things I write about. For instance, I was curious about something that seemed strange to me about a recent camera, so my AI and I set out to test the USB communications between camera and devices. After several different pokes, I found what was causing my initial curiosity: the engineers on that project made a mistake. Actually two that I can find, and both involve PTP (Picture Transfer Protocol, a defined standard). It’s actually the intersection between USB Power Delivery and PTP that this camera fumbles, as it prioritizes USB PD over PTP, when it should probably be the other way around (as most devices do).
That little AI sleuthing allowed me to answer a user question I previously didn’t know the answer to (why the camera worked properly with a computer but not an iPad). The answer: the computer was supplying more power than the iPad, so USB PD and PTP could both run. On the iPad the camera’s USB PD asked for more power than could be delivered, so the PTP request was terminated when the PD shut down.
Okay, I’ve rambled on enough about a tangential subject. But it does touch the boundary of this site’s topics, and Tokyo needs to careful here. The Chinese are, at least at the moment, not taking the same proprietary approach as they try to advance. What I’m seeing there is a lot of sharing of knowledge as they progress in optics. Propreitary only works as long as it stays ahead of the non-proprietary advancement.