Florian CΓ€sar 2021-02-24 08:29:42 On explainability of statistical methods (and their applicability to the FoC, perhaps). "AI" explainability is somewhat hot right now as we've realized that it's hard to make systems these work reliably (and trust their output) if we don't really understand why and how they do what they do. On a larger scale, better explainability could lead to faster iteration, better models and more widespread adoption in areas where uncertainty calibrated output is important.
I'm definitely not an expert and mostly focused on NLP, where the majority of research output is measured against various accuracy-based benchmarks with less investigation into the why's & potential biases (though there's some interesting progress like https://github.com/marcotcr/checklist).
- Do any of you have experiences with getting explanations out of statistical systems? Jack Rusher?
- Taking a larger view, any thoughts on how AI might help the Future of Coding (thinking of Codota, Kite, etc.)?
Tak Tran 2021-02-24 09:06:07 Hi, Iβm doing some research into making chatbots, and wondering if anyone knows of existing ways to model the algorithm of a chatbot? Iβm also interested in end user programming, so a model that enables that would be ideal. Also, thinking about how to store that algorithm in a declarative, language agnostic way. Any tips/links would be much appreciated π
Francois Laberge 2021-02-25 15:42:22 Anyone here aware of or experimenting with work focused on making coding/designing on a phone work well? I think about this all the time, I get these coding ideas I wish I could just quickly make or just do doodles when I have my phone but not computer out
Kartik Agaram 2021-02-26 19:04:09 Question for Tudor Girba of https://gtoolkit.com: how often do you end up finding problems in the code for viewing a specific type of object? Like, where you're debugging a specific app and you spend a long time befuddled before realizing that your senses have been lying to you because of a bug in some app-specific view that you wrote. It was hiding some crucial bit of state, or it gets stuck in an infinite loop that hangs the whole system, and now you have to debug the debug system without being able to use the debug system.
I wonder if this might be the reason people dislike debug-by-print as well. It is more powerful, but it forces everyone to wrestle with something typically considered the domain of systems programmers: the act of observing getting perturbed by the thing doing the observing. Or as James Mickens eloquently put it: I HAVE NO TOOLS BECAUSE I HAVE DESTROYED MY TOOLS WITH MY TOOLS. (https://www.usenix.org/system/files/1311_05-08_mickens.pdf)
Daniel Garcia 2021-02-27 12:49:16
There are, of course, the explicit computer processes which we use, and which our philosophy requires the augmented man to be able to design and build for himself. A number of people, outside our research group here, maintain stoutly that a practical augmentation system should not require the human to have to do any computer programming -- they feel that this is too specialized a capability to burden people with. Well, what that means in our eyes, if translated to a home workshop, would be like saying that you can't require the operating human to know how to adjust his tools, or set up jigs, or change drill sizes, and the like. You can see there that these skills are easy to learn in the context of what the human has to learn anyway about using the tools, and that they provide for much greater flexibility in finding convenient ways to use the tools to help shape materials.
I wonder how similar was the term computer programming
for D. Engelbart than it is for us today π€.
From: http://www.invisiblerevolution.net/engelbart/full_62_paper_augm_hum_int.html
Mariano Guerra 2021-02-28 15:17:37 Which dsls to define schemas you like? I'm looking for something flexible/extensible/as-declarative-as-possible. Looking for inspiration, I don't even care if it doesn't run as long as it can be made to run π