Earlier this year, I sat in on a fascinating war game in Paris. Four teams, two representing the U.S. and Chinese governments; two representing OpenAI and DeepSeek. Over several hours, under the watchful eye of a moderator, the teams gamed out how these actors, and the world at large, might react to accelerating AI progress. It was sort of like watching a game of Dungeons and Dragons, except the players were former government officials and AI researchers—and the game board was planet earth.
I won’t spoil the outcome of that wargame, which is the culmination of my latest piece for TIME. I use the exercise as a jumping-off point to explore how Artificial General Intelligence—once a pie-in-the-sky idea only seriously entertained by Silicon Valley weirdos—has become an increasingly salient dimension of great power competition between the U.S. and China.
As part of my reporting, I interviewed Dean Ball, who recently helped architect President Donald Trump’s new AI policy; and Jake Sullivan, President Biden’s national security adviser, who became increasingly concerned with the potential risks of AGI as his term went on. There’s a minor scoop in there, too, about classified AGI scenario planning exercises that Sullivan organized in the White House Situation Room back in 2022.
I hope you’ll give it a read.
A (new) newsletter
Earlier this summer, my colleague Andrew Chow and I launched In the Loop — an official TIME newsletter about the world of artificial intelligence.
Twice a week, we give you the run down on the most important storylines and people in the field. It’s tailored specifically to be a bite-sized briefing that you can read alongside your morning coffee (if you’re US-based) or your lunch (if you’re in Europe) or your dinner (if you’re in Asia).
I think you should subscribe! Mainly because it has been really good so far. But also because my bosses will like me if you do.
I promise I’ll still publish on Substack when I have big things to share! But for a weekly fix, In the Loop is the place to go.
Thanks for reading!
Subject: Story idea — Alignment before meaning in LLMs (structural, not values)
Hi Billy,
I’m sharing a short story idea that frames LLM alignment failures as a structural problem, not primarily a values/dataset issue.
Many failures (hallucination, inconsistency, misalignment) appear before meaning stabilizes — when the model’s perception locks into fixed “identity forms” (self/other/intent/agent).
I’m developing a framework called NCAF (Non-Conceptual Alignment Framework) with a companion method PMD (Perceptual Model Decomposition). Core claim: alignment must happen *before meaning*, at the level of perceptual structure.
If this is interesting, I can share a 1–2 page summary with diagrams + examples.
My post: https://northstarai.substack.com
Best,
Lee Chungsam