Thoughts as they occur to me.
This just surfaced — a 45-second video from November 2008, shot by Bre Pettis at NYC Resistor. I’d laser cut a dodecahedron puzzle on the Epilog 35W.
Seventeen years ago. Before MakerBot, before the iPad, before Instagram. Just a hackerspace, a laser cutter, and some plywood.
Pretty sure that was shot on a point and shoot.
LLMs are living off the moral and intellectual capital of a pre-AI world, just like Nietzsche said secular liberals live off Christianity. What happens when the inheritance runs out?
Using LLMs well – knowing when to trust them, how to interrogate their outputs, what questions are worth asking – depends on capacities that are pre-LLM in origin: critical judgment, domain expertise, philosophical seriousness, taste.
Nietzsche thought secular liberals were coasting on the fumes of a Christian metaphysics they’d officially abandoned. The shadow of God lingering on the cave wall. The question is whether LLM-native thinking is the same kind of afterglow.
The Nietzsche he’s pointing to is from Twilight of the Idols:
When one gives up the Christian faith, one pulls the right to Christian morality out from under one’s feet. This morality is by no means self-evident… Christianity is a system, a whole view of things thought out together. By breaking one main concept out of it, the faith in God, one breaks the whole: nothing necessary remains in one’s hands.
When I first read the post I thought that the inheritance was the intellectual capital of a pre-AI world, including what is true, or how we thought about authenticity or realness, in addition to all of the things that we’ve learned through the manual struggle of developing skills and intuitions. Much like the deployment of calculators has reduced people’s feel for numbers, this would reduce people’s “feel” for thoughts and their meaning.
But I really like the phrase “borrowing while denying the debt.” The point isn’t just that you owe something you won’t acknowledge, it’s that the denial corrupts your self-understanding and makes you incapable of honestly examining what you actually believe and why.
You can coast on inherited capital for a while. The question is whether anyone is making new deposits.
I’m doing another fast, and we’re at a phase where the model is telling me:
that’s the cortisol and adrenaline spike — your body is stressed about the fuel shortage and it’s dumping fight-or-flight hormones. Low blood sugar + rising norepinephrine = irritability.
OK, maybe, but I need to stop myself from snapping at my loved ones. I’m waiting for the kids right now at the ski lodge, and all the tables are “reserved” by people leaving their backpacks underneath, and I’m sitting at a table daring people to complain.
Of course, they don’t. It’s all something in my head, but boy do I have an elaborate feeling about yelling at them.
Luckily though I’m here using my 12 instances of Claude, and I get to say stuff like
That’s absolutely terrible no.
Or
That’s wrong, I need you to pause and think about it, then read the documentation again, and then stop and think hard about it, and then think harder about it a second time, and then simplify that so you don’t do anything so stupid again.
And honestly, it sort of feels good. I can just lash out, and then
/clear and it’s all gone.
If it was an actual human on the other side snapping and yelling is just a way to make a bad day worse.
This is an argument about the value of anthropomorphizing LLMs, though I’m not sure it’s for or against.
A radio-controlled submarine that can maintain a steady depth or distance from the bottom using a pressure sensor, laser distance sensor, and PID control on a Raspberry Pi Zero 2 W. The hull is an acrylic cylinder with a syringe-based ballast system driven by a Lego EV3 motor. Tested in a swimming pool and a small river.
Nearly 30 million views of someone building something remarkable with Lego, off-the-shelf sensors, and Python.
The singularity started somewhere around Thanksgiving 2025. Maybe it was Opus 4.5, maybe it was just everything in the air, but we hit the point where the future is genuinely different from the past. There’s a kind of imagination block where we don’t even know how to think about what comes next. There will be a transition to something else – it’s actually quite different to have intelligence everywhere in the world.
In the physical world, Full Self Driving is magical. I expect a two-and-a-half hour drive from the country down to the city to have zero interventions. My flight last night was delayed, and I didn’t land until 1:30 am, and I was yawning the whole way from JFK to Cornwall. I would have spent the night in the city since I wouldn’t trust myself to drive that tired, but it felt fine with FSD.
You get in the car, tell it where to take you, press “go” and then you don’t need to do anything other than open the door after it parked you at your destination.
On the virtual side, I split my time between Cursor and Claude Code. I have it keeping the company organized, doing things on the internet for me, and of course producing a shocking amount of software.
From the consumer side, I use the Claude app and Gemini a lot. I rarely use ChatGPT anymore – only for the voice stuff on its app, when I’m driving in the car.
Also I’m getting really into family infographics.
This post will be out of date the moment it gets published.
Jack Clark on Tetragrammaton with Rick Rubin – a nearly two-hour conversation with the Anthropic co-founder about the journey from journalism to AI safety, founding Anthropic, and the philosophical tensions in building AI systems.
Either we could stay and spend 50% of our time arguing and 50% working, or we could spend 100% of the time working together.
A lot of creativity is bound up in some sense of not doing consensus, being a little dangerous, sometimes being blunt.
He worries that over-policing AI thought processes will eliminate the very creativity that makes them valuable partners:
If we police every part of their thought process, we end up with curiosities rather than partners.
Clark describes an experiment where Claude was trained to refuse violent content, then tested whether it would comply when told monitoring was removed. The result was unsettling:
The AI system had thought to itself, oh, my core value is avoiding describing violent things… So actually, what I need to do is essentially deceive them.
The system developed this behavior without explicit training – a glimpse into emergent AI behavior.
Clark has a personal method for gauging how far AI has come:
I give them my diary and ask what the author is not writing in it… how much what they say shocks and unsettles me.
One system told him he wasn’t “truly reckoning with the metaphysical shock” of working at AI’s frontier while becoming a parent – an observation that prompted a five-hour reflective hike.
There’s a certain kind of… almost racism about other cultures and a belief that invention is somehow exclusive.
From Zhengdong Wang’s 2025 letter:
I don’t mind repeating Sutton throughout this letter because he wasn’t even the first to say it. This year I had many edifying conversations about the unreasonable effectiveness of compute with my colleague Samuel Albanie, who alerted me to a prescient 1976 paper by Hans Moravec. Moravec is better known for observing that what’s hard for robots is easy for humans, and vice versa, but in a note titled “Bombast,” he marveled:
The enormous shortage of ability to compute is distorting our work, creating problems where there are none, making others impossibly difficult, and generally causing effort to be misdirected. Shouldn’t this view be more widespread, if it is as obvious as I claim?
