Weekly Links: The LLM User Panel, Mathematics Proofs as Tree Search, and Security Flaws

AI & Humans seem to have proven the Riemmann Hypothesis, Robots start building in space, and security flaws continue to be flushed out.

Weekly Links: The LLM User Panel,  Mathematics Proofs as Tree Search, and Security Flaws

This week in AI: someone already completed an Ironman in GTA6 (hopefully they were fuelling with energy gels all the way...), Anthropic IPOs, and America's AI safety rules might be about to change.

On to this week's main stories. After a few weeks of company and business trend-type news, I'm going a little more technical. Hope you enjoy these:

  • LLMs Reproduce Human Purchase Intent via Semantic Similarity Elicitation of Likert Ratings. Using LLMs directly to get opinions on products gives you very middle-of-the-range average responses. Asking LLMs to act in personas, generate unfiltered reactions and aggregate them seems to do much better. This is an interesting technique for generating user feedback. Tuning the number of personas is likely to be hard, but this might be valuable for discovering spikey reactions (and your most valuable fans).
  • Stripe evals. I find this super interesting: payment company Stripe released an eval benchmark for how well different models integrate with its own APIs. I.e. it's defining the hill that models/agents should climb to interact with it. Feels like this could be the future for quite a few high-profile APIs. Maybe it's even something a service like apis.io could do if you threw enough tokens at it.
  • Lean and Mathlib Communities Certify Claude-Assisted Riemann Proof. Friday's announcement by Anthropic Levent Alpöge of a 17M-line lean proof of the Riemann hypothesis seems to be gaining credibility as a legitimate solution. This is another of the 7 Millennium Prize problems to potentially fall to an AI+Human team. What's more fascinating, though, is some of the commentary on how to prove it reaches its target. In a nutshell, the proof doesn't really build out much of the mathematical thought needed to get to a proof; instead, it takes a crazy complex hike through many other fields of mathematics and the proofs within them to build the argument. Something that would have been nearly impossible for any human mathematician to attempt. Assuming the proof is ratified, what does this mean? I'd say: 1) now that we have the hypothesis proven, we can probably work back through some of those other interesting areas; 2) we're probably seeing mathematics as a giant tree of knowledge that can now be brute-force searched.
  • Security flaw found in Java, world's third-largest programming language. The bug, which has been present for 12 years, is quite serious but has now been patched. Java is used in many of the world's transaction systems and especially in banking. While Java systems are often not the first line of interaction with the outside world these days, this is still very serious. It's interesting that much of the cybersecurity narrative about AI is negative. Logically, AI has accelerated the discovery and exploitation of flaws, so there is a negative side. On the other hand, these flaws have been present for years without detection by good actors. That doesn't mean they hadn't been discovered by bad actors.
  • A Robot Just Built a Structure in Space All on Its Own – For The First Time Ever. Quite cool that we can now construct things in space with a robot, one step closer to Dyson sphere constructions.

Wishing you a great weekend.