Wey Xie's Blog
This is quick jot-down of experiment results. I didn't put effort into writing this nicely for the reader, sorry. I just wanted to log my own thoughts. I tested LLMs performance on a poker river spot for EV loss against Nash EV loss vs best response (adaptive counterplay) Best response vs fixed strategy I wanted to find out what their current game-theoretic reasoning capabilities are. Namely, I…
I started this essay to investigate where reasoning came from, and I ended it caring about model welfare. Let me take you through this ride. Quick run through the training pipeline: We start with pretraining on large amounts of data, which anneals towards the end with higher quality data (sometimes called mid-training), as the later stages have a more direct effect on the final weights. Then we…
Reasoning is dual-use: It’s that which allows safety researchers to accelerate alignment; it’s also that which allows misaligned AI to take control. In May 2026, an OpenAI model has disproved a central conjecture in discrete geometry . Improvements in reasoning are useful and will continue to happen. But given their dual-use nature, we must steer them to selectively favor safe over unsafe…
I learned that I don’t want to build another business in an area where I lack domain expertise. I want to work on what pulls me intrinsically without the need for rational justification. Game theory is what I’ve been doing my entire life, and it’s game theory I want to continue doing. I could come up with reasons why this matters or makes economic sense, but no matter how many I’d come up with,…
Frontier models will continue scaling, and combined with training innovation they’ll become significantly more useful even than they are today. But they won’t reach general intelligence. A few providers will sell cloud access much like today, with expensive tiers out of reach for most people. A base model subscription will be a normal part of everyday life; poorer households will struggle to keep…
[This is a quick un-edited jot-down.] Since February: ran LLM experiments on deception, realized that LLMs still have some pretty big limitations. Also prompt sensitivity running these experiments requires very little technical skill and honestly feels quite superficial. Lots of prompt engineering/design required. And then you'd really need mech interp to extract anything useful. joined bluedot…
A couple days ago I was watching my non computer scientist wife train a machine learning model on her astrophysics work, and I felt the acceleration. I think about AI, its rate of development, and its societal consequences a lot. But for the first time I felt visceral fear. AGI could happen. And it could be dangerous. Dario and Demis had a chat at the World Economic Forum a few days ago, and it…
[Edit 2026-06-14: When writing this essay I was still unfamiliar with academic computational game theory and its terms. What I'm really after is opponent modeling. I'll write more on this in the future.] [Edit: My core intuition here is correct, but I'm not expressing well how my approach to Nash differs from Nash computed by CFR. In fact, because of that, the argument is a bit messy and may sound…
Epistemic confidence: 60%. Writing this essay allowed me to see that models could have a spectrum of real intelligence that scales. As of now, I still err on the side of the argument made here. [Edit: Since writing this I dove deeper into scaling, and now find some of my views here too naive. The training data quality problem however seems to be a real problem.] A couple days ago Google DeepMind…