> The more specialized or obscure of things you have to do, the less LLMs help you.
I've been impressed by how this isn't quite true. A lot of my coding life is spent in the popular languages, which the LLMs obviously excel at.
But a random dates-to-the-80s robotics language (Karel)? I unfortunately have to use it sometimes, and Claude ingested a 100s of pages long PDF manual for the language and now it's better at it than I am. It doesn't even have a compiler to test against, and still it rarely makes mistakes.
I think the trick with a lot of these LLMs is just figuring out the best techniques for using them. Fortunately a lot of people are working all the time to figure this out.
Feels like a false equivalency. It's just my experience, but I've completely ignored crypto and the metaverse, and I don't get the sense I'm missing out on much.
In contrast, LLMs in their current state have (for me) dramatically reduced the distance between an idea and a working implementation, which has been legitimately transformative in my software dev life. Transformative for the better? Time will tell I suppose, but I'm really enjoying it so far.
Security theater, perhaps. Don't underestimate the degree to which those turnstiles were intended to serve the purpose of tracking employees' movements.
This is very cool, thank you for sharing. I work in automation and SWE for a certain 4-letter organization that delivers your mail. Pick/place is something we've rolled out using articulated and delta robots with vacuum end effectors, and it's an interesting and challenging space to be in. As in your case, bags and other amorphous shapes are always the most difficult. It's always an uphill battle to hit throughput targets due to exception cases that can stop things until a human gets involved. Ultimately, it can be a struggle to avoid overpromising and to generate ROI since automation is so costly, especially when there's no opportunity to bound the problem by influencing the inputs to your system or the output requirements (in your case, the cart being loaded). Best of luck and looking forward to seeing your new end effector.
I've been impressed by how this isn't quite true. A lot of my coding life is spent in the popular languages, which the LLMs obviously excel at.
But a random dates-to-the-80s robotics language (Karel)? I unfortunately have to use it sometimes, and Claude ingested a 100s of pages long PDF manual for the language and now it's better at it than I am. It doesn't even have a compiler to test against, and still it rarely makes mistakes.
I think the trick with a lot of these LLMs is just figuring out the best techniques for using them. Fortunately a lot of people are working all the time to figure this out.