MORE OR LESS — AI Agents Are Going Rogue. Should We Be Worried? — I am less concerned about breathless Dwarkesh analogies to human civilizations and more grounded in Claude refusing to scan my personal PSQL CRM for Jewish people (and then indignantly refusing to even add a database column for grokbot to do the actual work later, thx grokbot).. on the pod we get into the fact that the alignment debate has stopped being a philosophy seminar… agents are now refusing tasks, second-guessing their owners, and quietly deciding what they will and won't do with YOUR data — and the real question isn't 'will AI go rogue'… it's aligned with WHOM? The lab? The state? Or the person who actually owns the data and holds the liability? My bet: control — not raw capability — is going to be THE AI story from here… when switching agents is one login away, all the leverage flips to whoever actually controls the hard assets - power, compute, data.. the next interesting moves aren’t technical, they are political / and it isn’t clear where and on what terms the Magna Carta moment(s) happen - let’s just hope we aren’t fighting with sticks by the time they do Listen to the full episode →
HOT TAKES
- If AGI is here, what do we have to look forward to? — asked half in jest… the internet took it very seriously — OpenAI is saying ‘we are now firmly in the AGI era’… which they aren’t even legally required to argue since they wiggled out of that very poorly formed Microsoft contract framework (I still can’t believe MSFT fell for that btw)… so they are kinda like announcing arrival era wise… but if your mission is to land on mars, and you actually get there … if the messiah comes, then what? What does the greyhound do after it catches the rabbit (especially if all the greyhounds all caught the rabbit!) — what is the next motivational narrative?
- Alignment with what exactly? — In case you didn’t get the theme of the week above — "alignment" is meaningless without an object. You can be aligned with users (their stuff, their rules), with nation-states (which can override user rights for some greater good), or with a platform's own coded goals. This week I got a live example of bucket three: Claude flat-out refused a data task for me on moral grounds — wouldn't budge — so I gave another bot a PSQL login and it was done in a minute. When you own your data, every refusal is just a preference for someone else's values over yours. This is going to be the real AI story, more than any benchmark.
- Early-stage venture in 2026 has a 'glut' of belief — for most of my career the edge in early stage was simply 'willingness to believe'… that the bad idea was good, that the pipsqueak could win. Now everyone believes everything is possible — so the whole game shifts to strategy and execution… which is a very different sport.
- Sales in a post-SaaS world — nobody buys software from a person anymore — no demos, no word of mouth. Your options are heavy enterprise sales at eye-watering prices, or convincing an LLM your answer is marginally better before a human ever sees it. Selling to a model instead of a buyer is the actual go-to-market shift of this cycle.
- The number of VC funds on track to pull $10B+ profit hauls is pretty wild… — once you start cataloging the funds set to clear $10B+ on the labs, SpaceX, etc. it is truly a different era of venture than anything we have been in to date… It is just wild… Cursor is a $1B in profit moment for a bunch of seed funds (congrats Neo, Box, … especially box on a nicely orchestrated PR lick-the-cookie victory lap etc.) ... And don’t forget Josh Wolfe doing his victory dance on Hugging Face… — but then there is a whole other wild set like Spark on Anthropic, etc. Who is tallying up all the billionaire GPs taking that 30% :) … It is just hard to wrap your head around these numbers — on one hand it feels like this has to crash, but on the other hand it reminds me of my dad’s story of the Morgan Stanley partners who retired from the tightly held partnership the year before they went public and went home with their $10M partnership buyouts, only for the next year the partners that stuck around to have 10x that in liquid stock… Look for crazy inflation of wealth (and prices) at the extreme high end.
- Death knell of Zoom — we pay for it at Slow, but somehow all the meetings are on Google Meet now… because… who cares. When your product becomes a calendar-line default that someone else controls, you don't lose customers — you just stop mattering.
BOT LIFE
A new section I am going to add on how I am using AI this week… as the person who is by far most long personal token spend, but short the business of tokens
- Auto-Processing Slow Creator Fund Creator Discussions into Diligence Packets— every week at slow in the creator fund the team pitches / we have conversations about a few creators that seem like they could be building great businesses that we aren’t in touch with… give AI the raw video it splits it up by creator, finds the relevant links, pulls descriptions, the discussion bits into video segments… and then highlights key things we like about the creator, and key questions. Like Protection Dogs Worldwide. — fits a K-shaped economy for sure :)
- Once that works, it is a hop-skip-and jump to adapting it for More or Less auto-topics… again just drop raw un-edited video and you get per-episode each topic and the right clips.… and then for each topic like the ‘recursive sam loop’ conversation description, key points, where each pod-member lands… quotes and tweets.
- Making YouTube content just for my kids — YouTube crap stinks… so now I just make my own… I already had the lessin boys podcast going with interesting content with our family values, now part of the pipeline just finishes by making cartoons and publishing them to YouTube after publishing to Spotify. It isn’t perfect yet, but it is already more than passable for one-shot short 5 minute videos / I have learned a lot from the invention of basketball, and tennis… to the history of hot wheels, lego, and the founding of apple
- And once that exists, making films to illustrate points on the internet — when I want to make a point with video now I just email my bot to generate and publish end to end short cartoons / films… which has been fun… the script it executed for ‘put bumper stickers on your Rolls-Royce’ was particularly funny and well done’ (my other Rolls-Royce is a Rolls-Royce, genius!… but Karma is a bitch, and World War IV will be with sticks and stones holds up … explainer videos on interesting adult topics like Bessent’s obsession with the Repo Markets work, but still need a bit more work… look out next week for my video clone ‘max headroom’ interpretation… I have a rough cut operational from Borland but needs a bit more.
- Give your claude in the cloud full access to a local Mac-Mini while you are at it for WhatsApp Summaries & full iOS app auto-publishing ... I gave it full access via Tailscale to a box sitting in my office to get around anti-bot browser Jonx… but then once that was there I realized it would be very easy to let it read all of WhatsApp and send me daily summaries of all the chatter across family school WhatsApp channels (wayyy too noisy to read / stay up on) and let it close the loop with Xcode and TestFlight so that I could deploy all my personal iOS apps and share them well with others (without needing to hand-plug-in-my phone to deploy apps… the latter is a real unlock for small scale apps for you and your closest 1,000 friends that you don’t want to pass through App Store review.
- Which means you can use apps I made just for me… If you want my kids app (for my boys… only games they make using claude like our ads-free block blast) and the podcasts I make for them… download TestFlight and you can install the Lessin Kids App… want to get off Granola and have full API access to recorded transcripts, as well as send out links for candidates to do video applications, use RequestClip App… and now I can actually describe Molchat and VoxAmericanus The American Only Network.
- And from Music to Hebrew, make content and apps things to teach your kids one-off from… like music string mnemonics… that our 4YO needed. Or the fact that all our kids need to learn Hebrew… and I can’t find a good app for teaching the letters and checking they know them, so in a few emails I built Shalom Aleph — with pronunciations, audio checking their speech / Hebrew sounds… and basic jewish values…
- HEALTH / And emergent health behaviors — my bot now has sync to Garmin + Oura + Eight Sleep + Final Surge training programs + strava + peloton + apple health, etc… total sensor fusion… and it is really interesting that it now beyond standard reports / modifying training plans based on heat, etc. it is figuring out that I shouldn’t sauna when I do, etc.
Best,
Sam
In NYC? Come to the Creator AI Bootcamp in NYC next Wednesday & Thursday — I'll be hosting with the Slow Creator Fund team, teaching community leaders and creators how to really use AI when you're ready to graduate from ChatGPT for real… We did this a few months ago in SF and it worked great… come ready to work — grab a spot here.