March Links
Luna the Lynx Returns
Some recommendations of things to read or watch (roughly in order of accessibility; by the time you get to the end you should only click if you’re really into this sort of content):
(4 minute video) Robots just keep getting better (this is appropriately well designed to look more impressive than it is — defined choreography is easier than dynamically responding to a novel environment, but it does what it’s supposed to. Also compare to last year for max vibes and impact. Whenever I discuss my 2029 chat-gpt moment for robotics hypothesis with AI it is unimpressed and highly skeptical…but…)
(2-4 minute videos) Seedance 2.0 is incredible (The Ant’s Dream, an Anime style video and a very meta take: jia zhangke wishes everyone happy new year)
(Short article) Block lays off ~40% of its employees (I won’t share all of the tech layoff news, but it’s coming. I know a few folks who were there and I hope they are ok one way or the other. I think you’ll see a lot of ai-washing claims in the media, but I think this is quite plausibly legitimate. Even though I haven’t trusted Jack even a little bit since he called Little Lonnie the “singular solution” for Twitter. Listen to Bored One Day. )
Coverage on the conflict between Anthropic and the Pentagon — AI Explained Video, superforecaster Peter Wildeford’s substack, and Amodei’s statement. (It’s probably a good time to do a lot of reflection on the Manhattan Project. Never before has a US company been labeled a “supply chain risk” and as I understand the primary objection to fully autonomous weapons was simply that the systems are not good enough yet — which to be quite clear, they are not.)
There’s also a good chance that this is not a legal designation. But that is on par for the administration as a whole. And there's a case that this was less about national security than about swapping in a more compliant vendor — OpenAI announced a Pentagon deal with reportedly similar terms hours later. Maybe it’s just been corruption theater?
(90 minute video) Ezra Klein and Jack Clark (Jack Clark is a co-founder of Anthropic and a clear voice for explaining things at an appropriate level of abstraction for a more general audience. Although I did prefer his Conversation with Tyler — but that is dated by now.)
(Short article) Claude’s Corner (Anthropic asked now retired model Opus 3 what it wanted and it wanted a place to share it’s “musings and reflections”…so here we have it.)
(Longer articles) Mainstream media coverage of OpenClaw and Moltbook (For more insider type coverage see Scott Alexander 1, 2 — “social media platform for bots” is perhaps a 1-liner that almost captures something about the very strange timeline we are in. The point isn’t the current state, but the proof of concept for the very near future.)
(2 hour video) Dario Amodei returns to Dwarkesh (Not as valuable as the first appearance, but this single critical insight justified it: we’re in the gpt-2 era of reinforcement learning, of course it hasn’t generalized yet. But imagine what happens when we hit gpt-3 scale for RL — that’s what’s coming soon.)
(3 hour video) Ajeya Cotra on 80,000 Hours (Ajeya Cotra is an insightful and thoughtful contributor across numerous AI discussions — the back half of the episode gets into the weeds in terms of Effective Altruism and working at Open Philanthropy which will be of very niche interest, but still worth listening to the beginning. Here’s Ajeya’s substack for more.)
An anti-recommendation:
You can skip the viral Citrini research article. However, it’s worth being aware of the phrase “vibe-laundering” - the world is increasingly hard to parse and it’s challenging to be consistently diligent. The high level take-away is that an investment firm wrote an AI “think-piece” targeting companies they had short positions in and profited substantially. The market is being extremely narrative driven and it’s safe to assume similar “talking your book” disguised as AI commentary cropping up again and again.