Omar Khattab

🔗 Thoughts on Research Impact in AI. Grad students often ask: how do I do research that makes a difference in the current, crowded AI space? This is a blogpost that summarizes my perspective in six guidelines for making research impact via open-source artifacts. Link below.

You cannot separate reasoning and knowledge as cleanly as you think. If you'd asked me what I care about in 2020/2021, I'd have said it was “decoupling the capacity that language models have for understanding text from how they store knowledge” (quote from link below this tweet). I was reminded of that on seeing an insightful account I follow say: “The only thing your LLM really needs to know is stuff like vocabulary, logic, and grammar. Everything else is mostly compute waste that we don't yet know how to get rid of.” This perspective, which I carried for years, is definitely more correct than then-mainstream ways of thinking of all-knowing monolithic LLMs as the point we'll converge to. But it's still, well, too naive. I no longer think that knowledge and reasoning can be productively decoupled. Yes, reasoning as symbol manipulation doesn’t need any world knowledge. But solving any real problem starts by knowing how to model it, which symbols to use, and what manipulations are likely to be useful. Knowing what to search for, even or perhaps especially for factual questions, is entirely about having enough context on questions even mean and what topics or sub-questions might be relevant or related. In other words, any productive reasoning needs to be grounded in knowledge—of facts and of analogies, as much as it is of techniques. You cannot decouple reasoning and knowledge as cleanly as you think. Now, systems must obviously be able to use retrieval as well as other tools. But this is not a replacement of intuitively knowing a lot of things and having a good “gut feeling” or vibe-like knowledge about things. The fact that large language models are large is fundamental. Modularity and composition are absolutely key, but they never let you get away with weak components.

Introducing Devin Security Swarm A more cost effective and accurate way to find security vulnerabilities in complex codebases, based on a new architecture: Agentic MapReduce.