This pen will kill you

Today is September 13, 2026, and we have five interesting things that happened in the last ~72 hours. I am gonna draw some connections between them because I think they are connected, deeply, and part of the same historical thrust. I think that, in hindsight, this post will be prescient.

Thing #1: Ex-Anthropic Employee Jacob Coxon does a big press run with this headline: “We really do earnestly believe AI could kill all humans”. It gets a lot of attention from a broader audience.

Thing #2: OpenAI shares a solution to the Navier-Stokes Problem, a mathematical problem that has to do with fluid and is really complicated to solve. It’s one of the Millennium Prize problems, and it’s a big deal. A lot of the world’s smartest mathematicians have been working on solving it for the last 100+ years. There was drama about the release which isn’t relevant to this post, but still going to mention it. Word on the street is, internally, labs have solved a lot of other open math problems too.

Thing #3: Deepseek launches DeepSeek-V4.1-Flash, an open source model. It benchmarks exceptionally well on agentic coding benchmarks like DeepSWE v1.1 and CyberGym (on-par/better than GPT 5.6-Sol).

Thing #4: Tibo (Codex’s de-facto point person on Twitter) announces OpenAI is temporarily pausing new Codex Pro sign-ups due to an inability to serve all their current demand post-Astra release.

Thing #5: Dario posts “We Must Pace the Frontier” blog post. Elon, Sam, and Demis all chime in to agree. So we have a situation where all the frontier labs are saying that they need to slow down because AI is too powerful.

Three points

I have three points to make on this overall tapestry of stuff going on:

1. The frontier labs have crazy stuff proven internally. Anyone saying they think they are bluffing crazy stuff and it’s all hot air is gonna be sorely mistaken when news comes out, and they will probably continue to move goalposts. We just solved Navier-Stokes. It’s a big deal and much, much more is coming.

2. The frontier labs are doing a coordinated effort to do regulatory capture, pull up the ladder behind them, and enact American protectionism (especially against China), and, ultimately, boost their company shareholder value and future IPO values. They are doing right by their shareholders to maximize shareholder value. After all, no better marketing than “It will kill you”. Religion has been doing this since the beginning of time and it’s incredibly effective.

3. The frontier lab’s ability to charge people $200/mo for ChatGPT subscriptions is dwindling and their business models will drastically change over our lifetimes. I don’t think this means that they are screwed. They have an insane amount of value and assets – compute power, know-how, software, hardware, infra, braintrust, everything. But their P&L will change entirely in composition over the coming decades. People will be able to get good models running locally, very fast, very soon.

Where are we headed?

Intelligence will be something that costs nothing and happens instantly. In ten years or so, we will be able to run models smarter than these on edge devices. What does life look like then? Remember when AOL used to send those CDs in the mail? Do you see what I’m saying?

Game-theory-wise, something that is potentially troublesome, if I wanted to be anxious about something, is that certain organizations will have access to the best models, and then there will be the rest of us who have to access the publicly available models. Who will decide who gets access to the best models and who gets the snicklefritz? Probably the frontier labs and the government. This will put a big weight on the open source community to release crazy models and make it so there are publicly available SOTA models. Already, if you think about it, Open Source catches up to the frontier labs in 6 months-ish after they release a model, but that’s their publicly available release. What they have internally is way ahead.