More or Less · Topics · Episode #167
Sam's Hot Take: The Labs Only Get Valuable as Defense Companies
Clip from the full recording, 20:07–23:43 · download mp4 · watch at this point on YouTube
Sam's thesis for how OpenAI and Anthropic become durably valuable: stop competing in the token market and become defense companies — build zero-day models you never release, then charge nations a fortune when their banks get subverted. Dave: 'I think we're already there.'
Sam unveils his big structural take: on a 5-to-10-year horizon, the only way OpenAI and Anthropic justify trillion-dollar valuations is as defense companies. You can’t make money competing in a commoditizing token marketplace — but 10,000 PhDs developing zero-day capabilities against other LLMs and NOT releasing them is a real product. The business model is nuclear deterrence: keep unreleased ‘zero-day models’ in reserve, and when a country or a bank gets subverted by a rogue AI, there are only one or two phone numbers to call — and no price you won’t pay. The model has to stay unreleased, because the second an intelligence is exposed, it gets built on top of and around. Dave’s response isn’t skepticism — it’s that this is already happening: the national-security conversation pops up every week, up to the president. Brit notes non-lab companies are already circling the space; Dave counters nobody else can afford to train at that level. Sam’s image for the future: ‘instead of 10,000 nuclear weapons, you’re gonna have 10,000 super intelligences in silos.’
Key points
- Sam: the token marketplace is unwinnable — the durable business is developing zero-day capabilities against other LLMs and holding them in reserve.
- The deterrence economics: once your country/bank is subverted by a rogue AI, there's no price you won't pay, and only one or two vendors who can fix it.
- Why the models must stay unreleased: any exposed intelligence immediately gets built on top of and around — only a never-seen model works as the antidote.
- Dave: this isn't hypothetical — it's the weekly national-security conversation up to the president.
- Brit: non-lab startups are already trying; Dave: nobody else can afford the training runs.
Where they landed
Quotes
“What you can do at the edge with 10,000 PhDs is develop a bunch of zero days against other LLMs and not release them.”— Sam
“Instead of having 10,000 nuclear weapons, you're gonna have 10,000 super intelligences in silos.”— Sam
“Once your country is subverted by an evil AI, there's no price you won't pay to fix it.”— Sam
“I think we're already there. I think this is not hypothetical and this is actually just what's happening.”— Dave
Suggested tweets
Sam Lessin's endgame for OpenAI and Anthropic: defense companies. Build zero-day models you never release, wait for a nation's banks to get subverted, name your price.
Tweet this →'Instead of 10,000 nuclear weapons, you're gonna have 10,000 super intelligences in silos.' — Sam Lessin on the real lab business model
Tweet this →The wildest part of Sam's AI-deterrence thesis isn't the thesis. It's Dave Morin replying: 'I think we're already there.'
Tweet this →All topics · Full episode #167 + transcript · Subscribe on YouTube · Spotify