You can add location information to your Tweets, such as your city or precise location, from the web and via third-party applications. You always have the option to delete your Tweet location history. Learn more
3) But w/o guarantees, will it introduce bugs? It might. In fact, it *will*. Copilot's current suggestions are mind-blowingly precise (30%), totally out of line (30%), or plausible yet buggy (40%). [Numbers pulled out of my head.] But it doesn't matter as much as you think: 4/
Boilerplate, API calls, test setup, data flow between structures, adapting SO patterns to your codebase… this is 80% of dev work. We've tried automating it for years but it's always intertwined with "smart" work. The ML/HCI model of "contextual code generation" solves that. 5/
Let me be precise. If your context/problem are unique, you will struggle with bugs in the current Copilot tech. And you will spend mental energy reviewing AI suggestions. But this is an *opportunity* for neuro-symbolic program synthesis research, not a reason to dismiss it. 6/
Neural code generation models learn a *contextualized language space* for us, in which we now apply PL analysis/synthesis methods more efficiently – with nice prelim results. Of course the goal is to get top-1 right. (And we will, eventually.) But I'm excited about the space. 7/
4) Does the code generation model truly understand program semantics? It's a philosophical question (define "understand" for neural nets, please) but right now – probably not. Program semantics is in execution, and neural language models don't observe execution. However: 8/
Program semantics is *also* in how people write, use, and communicate with/about code. People routinely "reason" about code semantics when reviewing PRs or reading Github. So do language models. Like in natural language, semantics partly arises from usage. 9/
It's incomplete, so new research on encoding actual execution and augmenting models with it is crucial. There's a huge (actively explored) space in deeply integrating static analysis, tracing, and other symbolic insight – *with* the language model, not to fight it. 10/
Maybe I'll outline more concrete research directions in a blog post, Twitter threads make me scream. I'll add to the thread as new questions arise, though. 11/
Hmm, on this point though, people have a tendency to trust machines more than themselves sometimes (consider the field of human-robot interaction). Are you worried about people using buggy generated code, but assuming it is correct since it came from a tool?
I'd really love to see a user study of people trusting Copilot suggestions to not be buggy when told 1) they came from Copilot, or 2) they came from a human
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.
Are there any formal guarantees on the generated code? No, not at the moment.
But it does synthesize code from scratch, given implicit intent. It does not "search" for it, conventionally speaking. 2/