After a few weeks of agent coding I find myself pushing hard to get Codex to produce less code 😅 Didn't really succeed yet. It seems to go very much against the model's DNA.

pro tip: after fixing a bug / adding a feature, tell codex to try to break it by writing a failing test case

yeah I'm still figuring out how to get past that initial complexity barrier with vibe coding current thinking: - step in early, co-design the architecture - force it to be modular, easy to read - every aspect of the app must be covered by tests

it's so funny how codex will confidently one-shot 1000-line PRs that span the entire codebase, but trip over when I tell it to commit

It's cute how gpt-5-codex, when planning changes, is sometimes like "we can do this bigger refactor once we have the bandwidth", as if it doesn't know it'll rip through it within minutes

Chat Control is the equivalent of police installing a camera in everybody's home that automatically reports suspicious activity. Except that it's much easier for determined criminals to bypass -.- If you're criminal anyway, you'll just illegally use an E2EE app lol EU, stop it!

last message from GPT-4o 🥲 it's pretty unemotional, that's because of my system prompt, just shows how good 4o was at instruction following. such a nice and tasteful model, way ahead of its time!

the answer, after all, might be 2130706433 https://t.co/88nyeB8bdx nice write-up @AngusGruen

> US twitter, for years: "europe can't do start-ups" > fastest-growing start-up ever launched from europe > US twitter: what must not be cannot be 😭

Weaponizing AI Assistants: With Their Permission https://t.co/txuNniqp9Q

sounds like grok4 was explicitly brainwashed to parrot right-wing platitudes great job elon on the AI safety front we really need our AIs to value "effectiveness in short-term threat elimination" and not be "paralyzed by feelings" 🦾

It's an honor to have @suneal_eth at @zksecurityXYZ, who is not only an excellent auditor but also finds critical bugs on Solana and other projects in his free time 🔥

It's funny how with #LeanLang, math concepts mix with programming concepts. I now think of `theorem`s as the public API of a theory, and `lemma`s as the internal helper functions

zk circuits: simple spec + complex implementation => perfect match for formal verification am i rite? or am i rite?

https://t.co/d7zvSMIgNE

5) If you have any circuit formalization needs, let's talk ☺️

4) That's how we got the idea for building 𝐜𝐥𝐞𝐚𝐧 - check it out! Blog post, by my great colleague @g_dellimmagine: https://t.co/sQfvcfr6q0 Code:

3) Lean is the best language I can imagine for an embedded zkDSL. It has rich math notation and is highly extensible to add your own syntax constructs. Dependent types make for an incredible programming experience if you care about correctness 😍

2) and instead of approaching formal verification as an add-on that is done after something is built - why don't we directly build stuff in a language that let's us prove theorems about our code?

Working on zk circuits in Lean has been my favourite project in the past couple of months! The idea: 1) instead of spending loads of time & money on tests/review/audits and still worrying that there are bugs, it's better to formally prove your circuits correct!