Interview with Christian Szegedy, Discoverer of AI Adversarial Examples [video](scale.ai)
scale.ai
Interview with Christian Szegedy, Discoverer of AI Adversarial Examples [video]
https://scale.ai/interviews/christian-szegedy
6 comments
The first generation of automated software engineers will write one bug per 100 lines of code. This will allow them to blend in while working remotely and provide job security for existing human developers.
The second gen senior automated software engineers will write one bug per 1,000 lines but will become too expensive to maintain as they age.
All efficiency gains will therefore be rolled back to the naive first gen's level in order to free up capital for the third gen automated angel investor.
The second gen senior automated software engineers will write one bug per 1,000 lines but will become too expensive to maintain as they age.
All efficiency gains will therefore be rolled back to the naive first gen's level in order to free up capital for the third gen automated angel investor.
It does not look very worrying. When he described the idea I wondered how you would actually do this given the amount of domain knowledge that is actually necessary to implement an underspecified program specification.
According to the interview Szegredy's current research is going after theorem proving in mathematics because there is very limited domain knowledge.
According to the interview Szegredy's current research is going after theorem proving in mathematics because there is very limited domain knowledge.
Looking at his recent publications [1] it seems more like he wants to create an "automated mathematician"
[1] https://arxiv.org/search/cs?searchtype=author&query=Szegedy%...
[1] https://arxiv.org/search/cs?searchtype=author&query=Szegedy%...
The interview goes into how this ties into formal reasoning towards the latter half.
This is a good interview. I was a little disappointed that it did not dig more deeply into adversarial examples. The compensation was that Szegedy argued that human-like reasoning may be based on relatively simply mechanisms that we just need to find. He cited AlphaGo as a solution that turned out to be much simpler than expected.
How does that work and should we be worried? The article doesn’t touch on this topic beyond the introduction.
(Note: I am aware of using genetic algorithms to evolve code to produce a solution to some specific problem. "Automated software engineer" seems to contemplate something more sophisticated than that.)