The paper refers to an ARM Cortex M4 with around a MB of flash and and few hundreds of KB of RAM; well that's perhaps a MCU, but to be honest that's quite powerful.
Micropython[1] targets that type of MCU, and even runs in a PIC16, so I'm not impressed.
When I was working on my AVR based 8-bit microcomputer (32KB of flash, 2KB of RAM) I did a lot of research regarding Lisp, and I gave up because it can't be done.
> It is still the most favored programming language for artificial intelligence research.
Correction: It is still the most favored programming language for GOFAI research.
It's pretty clear at this point (to me and to most active AI researchers) that some massively parallel statistical system (like Deep Learning based neural network, or Hawkins HTM, or something similar) will be a key ingredient, probably the dominant paradigm, in achieving electronic general intelligence. And lisp is pretty much absent in communities involved with that paradigm.
It's clear that 'statistical systems' are a hundred years or more away from 'achieving electronic general intelligence'.
'statistical systems', 'deep learning neural networks', ... are totally dumb.
Even commercial systems with massive money invested like Google Translate are only showing how far away from 'intelligent' translations they are.
It does not matter if Lisp is not used in those communities for AI research. The whole approach is limited and does not matter much in 'general intelligence research'.
Don't get me wrong. I'm a big fan of lisp family and learning racket myself these days.
I'm just saying that a statement along the lines of 'lisp is an artificial intelligence language' is really not valid anymore. You could do metaprogramming, metalinguistics, etc, all you want, and have top-notch productivity as a human programmer, but as it stands, these features have not panned out in the realization of an electronic intelligence.
Massivel parallel statistical (or stochastic, pick your favorite term) systems, despite all their limitations (at the moment), appear to me to be headed in the right direction, simply because they attempt to mimic the massively parallel and stochastic nature of brain circuitry. Is that the only way? not necessarily, but the problem has stumped us for so many decades, and has manifested itself to be so complex, that the uphill task of reverse engineering the computational model of the only working example seems to be "the lesser of all evils".
Back when I did more embedded systems as a profession I used various Forths and Schemes and it was far and away my preferred environment.
Those didn't run on a VM, though; while even the cheapest controllers are an order of magnitude (or two) faster than what I had back then I do wonder about the performance implications. For those types of systems portability wasn't really a concern, the worst case was that we'd upgrade the processor to a later family member, making the port almost always trivial.
Normal PicoLisp has a lot built-in: database, Prolog, web server, and lots more at less than 600K for the OpenWRT distribution. The MiniPicoLisp distribution, sans all the libraries, is only 73kB.
It is a very well-thought Lisp geared for decades for practical web and database applications.
Check the rosetta code entries for it to compare to other languages, simply amazing!
Micropython[1] targets that type of MCU, and even runs in a PIC16, so I'm not impressed.
When I was working on my AVR based 8-bit microcomputer (32KB of flash, 2KB of RAM) I did a lot of research regarding Lisp, and I gave up because it can't be done.
1: https://github.com/micropython/micropython
Edit: yes, I evaluated PicoLisp and Picobit (and many others).