MORE OR LESS — Meta Muse vs Instinct vs OpenAI Dots: Do AI Agents Have Any Moat? — I switch between muse, instinct, and dots at basically zero cost — one github repo, a few databases, feed the new thing my context, done… if the cost of writing software is going to zero, what exactly are you paying for? we also get into the anthropic S-1 leak — over half a trillion dollars in mostly non-cancelable contracts booked over the next 10 years (who's making the money? not anthropic — the cloud providers), aura's pulled IPO, and why flat revenue is terrifying. All episodes + transcripts →
THE BIG IDEA
AI LABS SHOULD PUT $1T IN ESCROW, MINIMUM for for pratical disaster recovery from their digital ‘Lab Leaks’ (covid coded)
My big issue with the Doomers is basically that on purpose or not, they are asking everyone to look at the crazy tail risk (and take their word for it) and skipping the CRAZY real high-likelihood fat middle risk they are causing in the process. Like ‘hey of course all the banks are going to get hacked because of our work and distillations of it’… but don’t worry about those TRILLIONS in dollars of much more practical damage, because what you should really care about is end of humanity low-likelihood-infinite-risk cases (don’t worry about astroids or aliens btw)… It feels like real slight of hand on purpose or not / with explicit regulatory capture intention or not (some I believe come by their religion honestly, others understand the game and are playing it on purpose & have been for a long time).
This week, I came up with what I think is the RIGHT proposal… which is for the labs to basically have insurance, so that when things go terribly wrong (and they will) at their hand, THEY pay for them (not dumping the liability on the PEOPLE / US government).
When EXXON has an oil spill (huge balance sheet, limited scoped damage) they have the balance sheet to pay, we should have NO confidence that AI companies will have the balance sheet to support the risks they are creating (they might already have negative value when you look at ability to generate profit vs. liability they case) … so we need as a society to MAKE SURE, that THE PEOPLE don’t pay while they profit / that they don’t dump all the practical risk on us…
How? Well we have several easy and sane models for this.
Post ’08 we forced banks to hold a lot of cash… can we learn from that and force the AI labs to hold a TON of cash for paying out damages when they inevitably cause theM?
We can force them easily to have a lot of insurance… insurance companies would LOVE this / if they have to carry $1T insurance policies each (realistic for practical non-existential threats forget all humans die!) … then Wall Street would be happy (except those who took big economic positions already on a risk-externalization thesis… and good news, insurance rates will GO DOWN when they do things / prove they are being safe! That is capitalism driving safety working!
We could also just have a government fund for them…Call it a national AI disaster-recovery fund … if the private markets won’t finance A and OpenAI insurance, the US government CAN backstop it and just say (look we will give you each $1T of insurance at $200B a year in cash payments to US) — btw, this is the Opposite of taking equity in them AND has the benefit of getting the government some much needed cash … maybe :)
The basic point? You can have whatever risk on tails you want … this is actually the whole EA problem… in general / anything times infinity is infinity… so you get no signal in HOW to invest relatively / trade (the west and 0 / infinity have had a terrible relationship since the time of the pythagoreans for this very reason… infinity breaks planning and western useful math) —
But we live in the REAL world, and we can’t let these infinity conversations break our normal math and sane decision making… if we have every reason to believe the labs are setting us up for HUGE national liability, they need to at a minimum be setup / we need to be setup to make sure they can afford to pay the damages from their work (not we the people)…
Oh and open source? Here is the thing… actually this is kinda easy (and to the lab’s benefit) — the model is simple. When you use a closed model like OpenAI or Anthropic that is a SERVICE… it is THEIR liability for bad things they do (or they do on behalf of their paid users)… they absorb the risk BUT… if you want to run an open source model and do-by-self, you inherit not just the technical cost but the liability footprint… people will choose to use service-based-AI to protect themselves from liability as much as anything else, and that makes SECURITY a good business… which is what it SHOULD BE… that is capitalism WORKING when financially people are incentivized to be safe.
Stop getting snowed by infinity cases and dazed / missing the trick and make sure the labs have the solvency to pay for the risks they are creating (and not risk dumping on the American People)
HOT TAKES
MY PERSONAL AI INFRASTRUCTURE, AS DESCRIBED BY MY PERSONAL AI — for those that are curious, I had my personal AI describe its own setup… love muse, but practically, for me — this slaps… at the cost of about $1500 in CPU / database per month … which even meta can’t afford.
FIN — TEN YEARS EARLY — personal assistant + context + credential vault + voice & photo interface + task multi-threading… we nailed the story a decade ago :) — these videos we made to advertise fin a decade ago still perfectly work today / are exactly the use case and interface of the dream… they age very very well. :)
Everyone is just scared, and everyone is acting as a result short term and transactional — everyone is terrified… (or if you aren’t then you are just rich and insulated and fascinated) … but either way, because there is no trust in the future to the positive or negative… timelines come in to months not years, everyone gets super transactional / trust is low, stress is high, and most people are not feeling so good… rebranding AI to ‘SI’ isn’t gonna help…
NY, SF, DC triangle trade of 2026… — Our greatest defense is no one has an incentive for the AI narrative to break… DC needs the GDP growth to service debt, SF party of all parties… and NYC loves it 2 / because for the first time SF really needs MONEY (prior tech was too capital efficient)… in the railroad era we were more of a real free market, things could fail — now does our managed political economy mean we will all just pretend it works!? Or let’s talk sub-prime, orsaid differently bath salts
Spending on US Human Education vs. US Machine Education — we still spend more training the people than the robots… for now… but might not be the case in a few years.
ARI EMANUEL’S BOOK — jess has been raving for days, I started it last night and stayed up too late reading it… as a jewish highly Dyslexic kid this hits hard…. I wish I had more RAGE :)
PHOTOS ARE THE CONTEXT GOLDMINE — I am REALLY into photos as a source of context for AI agents to do more for you… who you are, where you live, who matters to you — it's all already in your camera roll… I have a hand rolled version pulling all my camera context, ripping it through google places API, image identification, etc… but Marissa’s Dazzle is now out with a consumer version of it.
THE MULTI-AGENT CONSUMER FUTURE — it is so obvious… weak 'top of wallet,' you are gonna use all of them, and they are all going to use each other… great consumer surplus, impossible to see how any startup wins this game (and I include the labs in that as ‘startups’).
KALSHI FOR UNITED BUSINESS CLASS — how is there no polymarket for the price of the 8am SFO–EWR business-class seat on future dates? perfect way to express / hedge the cycle (and whooo dog right now)… economy is K-shaped, but not everyone has a G650!
FIRST OJAI IN A WAYMO — as a self-styled unpaid waymo influencer… very excited to get my first ojai — excellent!
USE OF (AI) TOOLS
Live Fact Checker — I now run my own version of Granola as Mac software — transcribe and summarize, but also hooked into my CRM and knowledge base for fact checking and suggesting questions to ask… it is awesome, way better than granola AND it is catching ME making mistakes on calls :) … love it
Event Roster — can’t figure out how to share this one in a screenshot… but find it great to snapshot who is at an event / the RSVP list and get auto-built pages of every person, headshot, who is already in your CRM, etc. with easy ‘add notes’ and ‘send email’ functions — again… so much great software in the world to want, with zero cost to build.
AI Task List — Build yourself a iOS / Mac / Web Note Pad — I basically built my own apple notes, with the KEY improvement that I can feed an always updating version to my bots as I go… every time I add or modify muse, my personal bot, etc. all snapshot it and propose how to fix things for me/ take things off the list. Money.
The Game Mafia, but for 2026 AI — you might remember the classic party game… someone is the mafia and kills villagers / people overnight, the detective tries to identify it — the villagers try to find mafia and kill it… well, I amuse myself — 2026 version… now we tried playing where the ‘mafia’ is AGI killing people, and the ‘detective’ is Greg Brockman who is hunting the AGI … and then there are just the sheeple…. Killing each other in an attempt to figure out who the AGI is!
HT to Muse on my Covid Shot — and an honorable mention for muse… I wasn’t going to get a covid shot, but apparently it is very bad this year (good marketing CVS!) … and muse booked it for me in one shot. Well done.
All the best,
Sam.
P.S. I have a new fun idea… 50/50 raffles for very fancy people at high end conferences… which is hilarious… and you will see me implement.