you carry magic in your hands, your soul and mind. attention and time then, is the most valuable thing you have. avoid the mind-eater voids or you will be creating your own curse instead of worlds

this must be what the furries felt like when they took away gpt 4o

fable 5

with the launch of claude fable i want to share some thoughts that i wrote last weekend we are blindly trusting everything that AI labs claim about their models, and we know that blind trust never goes well over the past 10 years the crypto space has been designing processes & cryptography to avoid having to blindly trust companies we should use what we've learned to build trust-minimized AI models too in this article i go though each step in the lifecycle of an AI model, explain why it needs verifiability, and propose how 🫶

nothing sexier than my girl in a trustless jacket

chat new opus dropped

You choose it or not, your organization is gonna be OpSec audited either way. Choose wisely

I'm convinced that AI and crypto are perfect for each other in many ways - verifiable compute - verifiable training - AI resistent sybil defence (proof of humanity etc) - agent payments & micropayments - decentralized training & GPU markets So I went down the rabbit hole to see how I could - train a model in a decentralized way - while still keeping some weights private - and zk proving that it was trained humanely/ethically That last one ran into a whole lot of problems and led me down an even deeper rabbit hole until I arrived at mechanistic interpretability From the outside, AI weights do not appear to correspond to how they act in any meaningful way Mechanistic interpretability (mech interp) is about opening up neural networks and discovering why the weights do what they do It basically says "don't just trust the black box of AI" which reminds me a lot of the philosophy of crypto Anyway, Claude and I wrote a curriculum mainly for myself to help me understand mech interp. Thought I'd open it up for other software engineers If you can't read AI papers or PyTorch but would like to, this is for you. It walks through 6 interpretability papers and replicates them in PyTorch. You can run them directly in the browser (using my huggingface space) Check it out yourself and let me know what you think! learnmechinterp.mcgee.cat/

Apple’s 2026 patch cadence looks unusual. By May 21, Apple had already shipped 7 iOS/iPadOS 26 updates and 6 macOS 26/Tahoe updates, plus a new category: Background Security Improvements. This isn’t just “bug fixes.” It may be a sign of a structural change in delivering updates.

good night

agree

don't fucking point your camera at me without asking first

goblin heart @Teknium

high on life rn

1/ The Red Guild needs you in order to survive. This is the first time we’re asking for help, and given our circumstances, it will not be the last. Please go to qf.giveth.io/project/the-red… and donate, even 10 USDC can make a difference. Continue reading below to know why.

🫠