Could AI help firms like Boeing perform more robust and automated safety checks? I'm curious about how much an issue like this can be chalked up to human error vs. poor, semi-automated QA.
This is hilarious, but I think we might be being a bit unfair.
Why couldn't you use something a camera on eyeglasses while doing the work correctly to fine tune a multimodal model, and then infer to a user wearing the same glasses? Audio reply saying "nope, try again."
You would need multiple frames per second, so not today, but not that far into the future, right?
edit: Zero-shot, I just took photos of me "fixing things" around the house, and here is what ChatGPT told me. It does not suck. Do this with fine tuning, many frames per second, and what am I missing?
The reason why it's bad is that this is an overcomplicated process that introduces potentially unreliable bells and whistles for little upside. At the end of the day, there aren't fundamental issues with a trained employee following a checklist. Machine learning is extremely powerful but tons of issues can be solved in a straightforward algorithmic fashion, so we should really be using it where it could make a big difference.
Thanks. I get it, and generally agree with your point here. Paying for good people and giving them enough time to get the job done is generally enough. I was just trying to be fair to op's question as devil's advocate. I don't think it should just be dismissed as a joke. I have always been a huge cryptocurrency skeptic, and I just don't see the same trajectory for ML.
So back in the day, Expert Systems were a big thing at Boeing. Searching Boeing's job openings today, it's a still a word used in hiring, but I don't understand what they mean by it in terms of manufacturing. Do you have any idea what they mean by that today?
Yeah, I understand your sentiment. I hope I didn't appear to technologically conservative in my comment, it is kind of a minor pushback to people promoting generative AI as a panacea that makes every problem ever easier.
I can't tell you about what Boeing does with expert systems. I've never been employed by them and I never really looked into it or where they use them. It's especially unclear because Boeing does a lot more than just pure manufacturing.
If a missing some major required component aside from finetuning and frames per sec, which will require a few years for everything to be fast enough... please let me know what it is.
Wait, oh it will be understanding of time... the sequencing of the frames. That is missing for now, right?
I think you're giving it too much credit for solving examples that are obviously wrong. Not only a non-expert but even most children know those pairings don't go together.
Given that the discussion in this post is around torquing to the right specifications, I don't know if just fine tuning is enough. It might need more serious training on videos of assembly. Even then, can it distinguish between the right torque or not from video?
I think it's interesting, but you should deliberately try to make it fail to see where the edges are. Like hold a wrench of the obviously wrong size next to a nut and ask if that will work. Force perspective so they look the same size. See if it will prevent you from mixing potentially dangerous chemical combinations. Will it warn you to wear protective eyewear when using a circular saw?
Oh definitely, I was giving it some softballs. Maybe I am also giving too much credit to what fine tuning can achieve.
But in a manufacturing environment, if one labelled wrench sizes with different colors, then things like noticing wrench size would be easier. Also, I imagined that the camera would of known (same) geometry during training and inference in a Boeing implementation.
I also wonder if multimodal LLMs are blowing my mind unnecessarily. It really feels like a huge leap to me though.
They are a pretty huge leap! I just think for manufacturing and assembly where it is critical to get specific values right we cannot rely on neural nets in their current form.
I think there are digital wrenches that will record the torque they applied, you don't need to mess around with colors. But at that point, it's sort of like, why bother teaching the AI to notice if people are doing it right, stick a camera on the wrench and QR sticker on the bolt. Make a more formal verification process that can be guaranteed to match specifications.
Neural networks for quality assessment aren't necessarily unheard of. I'm not sure how they'd be using them in this case, though?
Maybe something like a camera monitoring that they are using the correct tool at various parts of assembly? But I'm not sure how feasible this would be at airplane levels of volume.
I bet it'll be common in car assembly at some point, though.
