The killer app model fails for intelligence because it assumes problems can be modeled and hardcoded into software. Intelligence enables the modeling of problems in real time. Traditional software relies on pre-modeled problems: the creator anticipates a user's need, builds it,

not to rub it in, but this is low taste poor problem selection and doing it six months after Google steelman: there’s some strategic advantage but it’s a product miss if new ai os, what can I do I couldn’t do before?

browsers are simulations

imagine hundreds of thousands of knowledge workers start delegating bits and pieces of their work to agents significant enough quantities to change the kind of projects they can take on how does this change the nature of the firm

an ai company that manages to generate 90% gross margins and unlocks new demand for intelligence will win the paradigm

prompting is compromising, hence Braid

consensus is opportunity

what is your timeline for AGI as in an intelligence that can solve ill defined out of distribution problems?

I have great affection for these fringe subcultures in computing, but as I get older, I'm becoming more demanding. I want the philosophy to be effective out in the world; otherwise, it's meaningless. Just a bunch of technologists talking to each other. Feckless.

When the multiplayer browser demo went viral, I realized it's easy to create charismatic moments. It was not even a launch. I made it in iMovie in 20 minutes. But those moments are mostly a distraction from the technological imagination and invention I have in mind. They

first principle: intelligence is sample efficiency more compute does not compensate for lack of sample efficiency you misunderstand the bitter lesson

sometimes people reveal gaps in their world model Think of software at three levels: surface, function, and value Think of the ideal distribution between novelty and stability at each level Do you want the surface to be high novelty or high stability?

I feel like language models use space more than time for representation that doesn’t necessarily have to be the case intuitively I’d want the space heavy vs. time heavy trade off to be dynamic at runtime doesn’t make sense for intelligence to have a fixed resource tradeoff

there is billions of $$ in this tweet btw

if you’re in your 20s in tech right now, a couple of years spent handing the biggest ass whooping anyone has ever seen to OpenAI, Anthropic, or any known AI startup will give you 10x the career juice of any MBA or graduate degree (you’ll build the post-llm intelligence paradigm)

Sober sincere competent judgement is rare. This must be celebrated.

grinding doesn't work when there are paradigmatic issues to be solved you must pursue moonshots there's no safety on the frontier

Adam Rackis Adam most software you and I use will be written by other people, mostly by volunteers creating public goods It’s not out of line to ask for basic consideration like not being entitled and dunking on people doing the work devs know their work is a gift & not something they owe

every time i get a glimpse of the agent runtimes outside of Braid i’m shocked you all live like this still wrapping automation frameworks huh 😏

LLMs will be disrupted.