F12

Richard Feynman

My approach to problem-solving is to carry around a dozen interesting problems, and a dozen interesting solutions to unrelated problems, and eventually, I’ll be able to make connections. You have to keep a dozen of your favorite problems constantly present in your mind, although by and large they will lay in a dormant state.

In no particular order, these are the topics occupying my mind:

Question #1

1. Do sociotechnical selection pressures reliably favor ML systems that (a) increase their own future deployment probability and (b) reshape institutions/data pipelines to entrench that probability, even without explicit ‘survive’ objectives?

Question #2

2. What would next-generation ‘tools for thought’ look like? And how do we build them for the messy reality we live in? Think something deeper and more useful than a “second brain”…

Question #3

3. How can we reduce information overload, unhealthy incentive gradients, polarization & misinformation through software and digital tools? In other words what are the interface primitives for anti-bullshit systems? Whatever it is must help users navigate ambiguity, detect conflicts, and revise beliefs without getting trapped in premature closure

Question #4

4. How do we build systems that help humans think with evidence, rather than merely consume fluent summaries? In other words what architectures and interfaces can transform AI from generators of unverified text into epistemic tools that help humans interrogate evidence, test claims, and construct reliable knowledge?

Question #5

5. What session-level behavioral signatures predict trustworthy long-horizon research outcomes?

Question #6

6. How will the advancements we’re making in neurotech change the business and consumer worlds? A book I’ve been reading lately is: Military Neuroscience and the Coming Age of Neurowarfare by Armin Krishnan

Question #7

7. What evidence structure makes a machine-generated or machine-assisted claim cheap to audit?

Question #8

8. What type of software companies will end up being successful in the short vs the medium to long-term for the AI era?

Question #9

9. What signals am I getting about the future and what should we expect the year 2040 to look like? Am I likely to be drastically wrong in any of my expectations and assumptions?

Question #10

10. What UI primitives would be beneficial in graph-native attention routing and dynamic hypothesis graph systems?

Question #11

11. In a world that’s optimized for convincing bullshit, what would a system that defends against human cognitive flaws look like? What form would an anti-propaganda machine take?

Question #12

12. The Abstraction and Reasoning Corpus (ARC), introduced by François Chollet in 2019, was designed to probe a system’s ability to acquire novel skills from minimal examples. In other words by analogical transfer and local abstraction. What’s interesting is how a models score is not portable without the harness. AGI-3 is a lab environment where success depends not on guessing the answer directly, but on figuring out what information to gather before guessing. Agentic evaluation is no longer optional ref. However there’s a major issue with all of the ARC-AGI benchmarks. They’re all closed puzzle environments. Evaluation is: given examples -> infer rule -> produce output. Even with ARC-AGI-3 coming out soon the environment is “video-game like” as they describe it, this fails to answer the question of if agentic systems, when equipped with tools and evidence, are capable of investigating open questions. My hypothesis is that we may be vastly underestimating agentic intelligence. We’ve already seen some interesting claims in regards to mathematics… So the question that remains is: how can I prove this?