Before AI the focus was on making software that "just works" for other developers. Designing a system and it's interfaces where people from other teams could independently hook into. One example is this autonomous bulldozer where the control system was a bunch of spagetti. How do you manage multiple events, commands, sensor data from various locations? With some kind of ROS-like RTOS. But that takes a learning curve to learn all the ROS bloat. (I met one of the original middleware developers when I interviewed for a robotic tractor company and he's like "lol, we never use that".) Anyways, I applied some design patterns to the ROS spageti and ended up with a synchronized, multithreaded interface that the ML image people could understand, the controls people could use.
Today, that would be done in 5 minutes.
So, now I'm happily making real stuff "just work". Before it would have taken months to build up a sliding random linear network code error correction system. Especially because we were using an antiquated TI dsp chip. Now, in two weeks plus some debug time, Boom: Deployed, and our radio audio quality is nicely improved.
Need to do parameter optimization for a product? Code up some CLI hooks, test harness, instruct the LLM to have at it. Autonomous over night work and Boom: multi-variable plots showing the tradeoffs as parameters are tweaked.
thanks for sharing! I enjoyed reading the stories you've shared. By the way what does "sliding random linear network code error correction system" mean?
Its a very cool application of linear algebra! You can recover lost packets by sending "repair" packets with redundant information.
Its acronym is RLNC and is part of a larger family of fountain codes and in general erasure correction codes.
The sliding refers to the redundant repair packets are "encoded" over a sliding window.
Essentially for encoding repair packets youre creating a matrix, then taking a random linear combination of the last N packets. Then you send a repair packets with the seed of the random coefficients. You send as many repair packets as airtime allows!
The receiving you need to compute a matrix inversion through gauss Jordan elimination. That has to happens super fast, so luckily there are hardware accelerated operations over cyclical integers (Galios Fields).
A huge undertaking without AI unless youre a wizard at mathematical programming. Something one would only do if you're tail end of the IQ spectrum or you're a 40 year vet.
Theres a lot of blah blah about the gravy train stopping. Idk. I feel like this new train is fire.
What im excited about next is trying some hardware design with this XLS Google HDL which generates verilog from a rust like language. Freaking get the LLM in the loop with the debugger and logic analyzer and you're off to the races.
wow, thank you for the detailed explanation. it's good to know.
I find that it's frequently enough to just know something exists, then you can ask AI to do it and it does so. The example you've mentioned would be something like that.
I don't think it would do it by itself unless you asked it to, right?
Delivering software that "just works".
Before AI the focus was on making software that "just works" for other developers. Designing a system and it's interfaces where people from other teams could independently hook into. One example is this autonomous bulldozer where the control system was a bunch of spagetti. How do you manage multiple events, commands, sensor data from various locations? With some kind of ROS-like RTOS. But that takes a learning curve to learn all the ROS bloat. (I met one of the original middleware developers when I interviewed for a robotic tractor company and he's like "lol, we never use that".) Anyways, I applied some design patterns to the ROS spageti and ended up with a synchronized, multithreaded interface that the ML image people could understand, the controls people could use.
Today, that would be done in 5 minutes.
So, now I'm happily making real stuff "just work". Before it would have taken months to build up a sliding random linear network code error correction system. Especially because we were using an antiquated TI dsp chip. Now, in two weeks plus some debug time, Boom: Deployed, and our radio audio quality is nicely improved.
Need to do parameter optimization for a product? Code up some CLI hooks, test harness, instruct the LLM to have at it. Autonomous over night work and Boom: multi-variable plots showing the tradeoffs as parameters are tweaked.
It's different for sure, but still fun.
thanks for sharing! I enjoyed reading the stories you've shared. By the way what does "sliding random linear network code error correction system" mean?
Its a very cool application of linear algebra! You can recover lost packets by sending "repair" packets with redundant information.
Its acronym is RLNC and is part of a larger family of fountain codes and in general erasure correction codes.
The sliding refers to the redundant repair packets are "encoded" over a sliding window.
Essentially for encoding repair packets youre creating a matrix, then taking a random linear combination of the last N packets. Then you send a repair packets with the seed of the random coefficients. You send as many repair packets as airtime allows!
The receiving you need to compute a matrix inversion through gauss Jordan elimination. That has to happens super fast, so luckily there are hardware accelerated operations over cyclical integers (Galios Fields).
A huge undertaking without AI unless youre a wizard at mathematical programming. Something one would only do if you're tail end of the IQ spectrum or you're a 40 year vet.
Theres a lot of blah blah about the gravy train stopping. Idk. I feel like this new train is fire.
What im excited about next is trying some hardware design with this XLS Google HDL which generates verilog from a rust like language. Freaking get the LLM in the loop with the debugger and logic analyzer and you're off to the races.
wow, thank you for the detailed explanation. it's good to know.
I find that it's frequently enough to just know something exists, then you can ask AI to do it and it does so. The example you've mentioned would be something like that.
I don't think it would do it by itself unless you asked it to, right?
Not saying yes to whatever AI suggests and always guiding it as per my instructions.
These days, I'm proud of the live I have built outside of tech. It makes the AI-related identity crisis a lot more manageable.
[flagged]
Having not succumbed to ai psychosis
[dead]