SWL Week in Review - M-M-Maximum Entertainment

From:

sam lessin <s@wlessin.com>

To:

Hidden Recipient <hidden@emailshot.io>

Date:

9/28/2026, 12:08 AM

Stop Pretending the AI labs are a special case in business… they are not.

MORE OR LESS — Pro West Point, Short A16Z U — (the preview got heated)… this is the convo that drew all the ire… but look, when it comes to learning the skills that actually matter, I for one am very long military over whatever A16Z U becomes… marketing ploy our attempt to be 2026 Carnegie Mellon (who is gonna need engineers anyway)… We also riff on Blackstone exec admitting no one knows who is gonna pay for the data-center debt, and Greg Brockman quietly reframed AGI as 'gradual' — which kills the winner-take-all math the entire capital stack is priced on… infinite demand does not mean infinite sustainability (let alone profitability) as a business. All episodes + transcripts →

HOT TAKES

ON AI LAB LIABILITY FOR DAMAGE IN THE REAL WORLD

Companies with infinite liability downside and finite profit… is indeed a business innovation.

This is not complicated and we have to stop pretending that it is, letting breathless self reported religious ‘infinite risk downside cases’ absolve them of the enormous practical liabilities they are creating (and the fact those practical liabilities might easily be far greater than even their enormous private valuations).

I wrote about this a bit on Saturday the frontier should be paced by clear and obvious liability… and it even fits nicely with the difference between running open models (you run it, open software, your liability as an operator) vs. closed labs AI as a service (if they run it for you, they are liable for what they do both for research AND for users) — running naked yourself open source vs. outsourcing the work and risk to a lab — there is the business, they are either in the insurance business or reinsurance business… and if THEY can’t afford the liability exposure of inevitable practical bank hacks, etc. then they have to slow down (or their reinsurer will make them) — that is how this is supposed to work / works for every other business…

This just is easy:

1) If you run AI as a service for others (anthropic, openai), you are responsible for damage you cause / liable... not the user that asked.

2) If you take an open source model, run it yourself, and cause damage - then you/ your company are responsible/ take the liability.

But oh, the FAA works so well, regulation blah blah — please.. in the early days of airplanes pilots were killing themselves and their passengers all the time and taking the liability of it all … the difference here is that they could only kill a few people / crash into a few buildings … there is no way regulation will move fast enough to stop anything here, and it would be absurd for us to let AI companies internalize the profits and externalize the costs and risks that blatantly. (oh and as a nice aside, if you do a contract for the military, they can take that liability for you in the kill chain contractually)…

Observability & reporting standards? Yes, that is a sticky one tho — and standards of reporting that big AI players need to adhere to so you can figure out when OpenAI inevitably shuts down Bank of America and very calculable damages run into the tens-of-billions (or more!) … but again, say NO to risk dumping on the American people by the labs … we can deal with the data-center-buildout financial liability a bit later.

All this to say, forget the interest cost on sub-prime 1.4T of data-center spend once those bills start to hit vs. revenue (forget profit) assumptions… forget the Open model commodification that is obvious and happening … just start with the liability thing.

And on existential risk? Or what happens when we aren’t the prime world intelligence? Maybe I spent too long talking with the promo-people about paperclip optimizing in 2011 where AI drags over a black hole to kill us all… it is all really fun science fiction, and look it might happen - but there is so little we can do about it in the big picture you kinda just gotta roll it as PROBABLY philosophy and religion disguised as business … not to say that the people on the inside don’t really believe it in their hearts… they do… but this stuff was always on a slide next to astroid hits and alien invasions — both of which have non-zero chance of happing to and infinite cost. Western civilization and business doesn’t really do well with infinities and zeros — never has, the pythagoreans where hated thousands of years ago… some risks you just live with (and some have figured out are great marketing btw!)

The big message here — we can talk philosophy and religion and it is ALL FUN… and there is a more than zero percent change real, because EXCITING! But we can’t let the shiny philosophy talk distract from business realities and excuse the obvious frameworks of our market. Mange the manageable parts.

Oh, and to the ‘but China will beat us!’ … look I get everyone here has an Oppenheimer complex and has been describing AI as nuclear weapons from day one… but this is all just matrix multiplication — and unless you go in a bunker and develop zero day models to wrangle the GLOBAL open source ones which get plenty smart and can do plenty of damage… (plausible nationalization path btw)… so it isn’t ‘just china’ — and what is your alternative? You ban all open source and require every computer on earth to have a ‘fed chip’ to make sure they aren’t multiplying numbers that are too big??

That is always how this goes, fear opens the door for more centralization and autocracy… and when a big bank gets hacked, it is gonna be a great pretense for the government to take pretty direct control of the entire system (or have anthropic do it!)

THE GREAT IRONY IN ALL OF THIS IS WHERE IT ALL STARTED

OpenAI and Anthropic really did have a good mission and model initially… not for profit companies, but non-profits… but the need for CAPITAL at astronomical scales switched everything around for them… because in the west the only way to get that scale of dollars is to go AS FAST AS POSSIBLE without safety AND tell a story of ‘own everything’… history does love to laugh… and it would all be quite funny (or will be in a few centuries) when a moral and philosophical existential scandal scrambles our ability to focus on the practical realities of labs waking up to a world where they very practically (not externalized to their ‘agents’ / THEY) hacked everyone.

Or they are right and we are all dead anyway, but that ship by their own admission (and failure of their mission) would have already sailed anyway …

Or this is all true but we do nothing about it anyway because this is the moment of greatest NYC, DC, SF mutual dependence triangle trade party time… and no one knows how to unwind all of this without taking everything anyway, so we have no real option for smart thinking or strategy, we just kinda drift into a terrible configuration (and maybe it is fine because unlike earlier eras our economy is so-post- 08, covid manipulated anyway)

USE OF AI

First, muse — good. Well done to the team, and I really like Alex Wang’s twitter game. I think this works for consumers (if not crazy edge people) — We just got our good eggs order for the week from ‘muse look at last three orders and buy right things, don’t buy anything in more than triplicate’ (Jessica’s prompt… nailed)

Second, my use — here is my system’s self description which does now include muse and grok bot, but mostly because I setup a cron for them to check a task table and pull off of it vs. using their interfaces.

I went long Mac apps (in addition to iOS this week) in my personal setup / suite…. I now have my stripped ‘notes’ text app as a todo buffer running for me, with AI looking at the text file (and muse + claude examining and offering / executing to pull things off / do things / remember things as I drop notes in)…. In addition to email, the constantly synced text file space you are working in is IMHO a better model than chat threads. Input will be multi-modal

Earshot App for Realtime Fact-Checking & Question Prompting on Calls- I was on a call and a bit bored with it, so I asked to have built / deployed on Mac and iOS an app that listens to everything, transcribes, knows the context of call from calendar & leverages my full CRM, deal notes, etc. and does two useful things other than summaries (1) it fact-checks live… when people say things, it gives me realtime ‘right’ / ‘wrong’ / questionable / contradicts last thing they said in email (2) it suggests questions I should ask… it is pretty good… sam bot will be investing long after I am dead. This stuff is childsplay to build at this point in 30 min and with one prompt and two revisions, which means you can love it all, just not the companies trying to sell it as a package.

Sam

P.S.Jessica / The Information is looking for her first Bureau Chief in Tel Aviv — if you like journalism and tech… this is a killer job with a killer team. apply or forward to the right person.

Similar newsletters

There are other similar shared emails that you might be interested in: