At some point, Veselin — former colleague and founder of DeepCode — asked: "Why don’t you work for me?" That was intriguing for many reasons. Working with him, applying graph theory to solve software security problems, doing so with modern C++, a triple data store, and applied AI techniques made the decision a no-brainer.
While we were deep in optimizing the engine, the idea came up to present some of it. I’d also just got a call from Yann at Confoo, a conference where I hadn’t spoken in a few years — I’d been happily hiding inside Google. So it was settled: condense everything down to a point Snyk was happy with, and out came this:
Symbolic techniques perform explainable predictions and are essentially immune to hallucinations — but they require significant manual effort to scale across languages, frameworks, and rulesets. Neural techniques are a promising new direction, yet their ability to discover bugs in code is limited by hallucinations, and in the end they’re not significantly better than guessing. Let’s combine them to overcome the shortcomings of each.
The conference was great, as all the others have been. The conference was great, as all the others have been. I made a bunch of new friends — and all of us speak at conferences.