These are a few problems I care about and hope to contribute to. I don’t have neat answers to them; this is a running list of ideas I’m trying to understand.
- Since high-performance code is becoming cheap now, what does the future systems stack look like? Can we retire the intermediate abstractions, keep a high-level spec as the source of truth, and generate optimized code? Do we need to formally verify against the spec?
- What does the next generation of interfaces look like? How can we make malleable software ubiquitous? Are current efforts (one and two) steps in the right direction? What if we augment all deterministic workflows with intelligence?
- How can we improve intelligence per dollar (or watt)? We seem to be making progress on the frontier by biting the bullet on ridiculous orders of scaling; it is time for efficiency to follow. Does this come from better algorithms, better hardware-architecture co-design, or cheaper compute?
- The definition of "work" has changed so much in the past century. In a post-AGI world, how will this evolve?
- How would RSI work if future progress requires a fundamental paradigm shift? If both us/models are stuck in local optimums, would we need new core learning theories? Ones that are OOMs more sample efficient?
- This is on a lot of minds at the moment, but how can we avoid these scenarios? How is alignment best tackled: scientifically, institutionally, or through government?