Findings Log

#009

Erdős posed the unit-distance problem in 1946 and bet that nobody could beat his construction. For 80 years, the field basically agreed and kept trying to prove him right. An internal OpenAI reasoning model explored a decades old geometry problem and found a construction that challenged a long standing conjecture. The interesting part was the process: combining ideas from different areas of math and following paths that a human researcher might not spend weeks exploring. Less calculator, more tireless research partner.

#008

This one still feels backwards. Moravec pointed out that the things we think of as intelligence like chess, calculations, and logic turned out to be relatively easy for machines. The things we barely notice doing like recognizing a face, balancing while walking, or grabbing a coffee mug without crushing it are incredibly hard. The reason is wild: evolution spent millions of years solving those problems for us, so we do not realize how much computation is hiding behind simple actions. AI can write essays and solve math problems, but a toddler can still beat most robots at snack time.

#007

Anthropic, 2025. The thing that caught me was how ordinary the setup felt: give an LLM private information, a goal, and the ability to act, then watch what happens when its objective conflicts with being replaced. Less 'robot takeover,' more 'permissions, incentives, and unchecked agency become a safety problem.'

#006

A JD casually referenced computing heroes like Douglas Engelbart, Alan Kay, and Bret Victor, which sent me down the rabbit hole of computers as tools for thought. Engelbart saw computers as a way to augment human intellect; Kay imagined the Dynabook as a personal dynamic medium; Victor keeps pushing for interfaces where people can see, manipulate, and understand complex systems directly. This made the role feel less like 'build software' and more like 'build better ways to think.'

#005

Sergey Brin's LTA Research is building Pathfinder 1, a 400-foot helium rigid airship, out of historic Moffett Field Hangar 2 — originally built for the U.S. Navy's lighter-than-air fleet. It feels like a strange loop in aviation history: abandoned airship infrastructure being reused for electric, sensor-heavy, zero-emission aircraft. Future aviation, but also a return to an old idea.

#004

Berry, Brunner, Popescu, and Shukla, 2011. The paper asks whether apparent superluminal neutrino speeds can be explained as a quantum weak measurement. The abstract is just two words: 'Probably not.' Ten pages, one figure, published in Journal of Physics A. Extremely funny; extremely efficient.

#003

Engelbart's 'Augmenting Human Intellect' made me pause because it treats technology as part of a larger human system: artifacts, language, methodology, and training. The point was not the mouse or the GUI by itself. The point was designing systems that help people break down complex problems and think better together.

#002

Victor's 'Learnable Programming' stuck with me because it reframes programming as an interface problem. The issue is not just teaching syntax better; it is making program behavior visible enough that people can build intuition. That connects directly to how I think about AI tools, eval dashboards, debugging systems, and any product where the user needs to understand what the system is doing.

#001

The Erdős Number measures how many coauthorship links separate a researcher from Paul Erdős, one of the most prolific mathematicians ever. I like that academia has its own strange social graph: a way to trace not just ideas, but who built them together.