@tszzl Remember to be extra careful. 🙏 Recursive self-improvement is what causes a fast take off and fast takeoffs are the cause of the bulk of loss of control scenarios.

@Kat__Woods i think we lost control sometime in the late 18th century

@tszzl Rapid takeoff for the average corporate employee will happen when there is a model the employee can teach to perform their job functions. It will work along side them all day and can ask questions or take direction at any time.

@tszzl Watching Codex Agent in Cursor work autonomously for 20 minutes this morning to fix a longstanding issue where other AIs failed gave me the jitters. I find myself yelling lesser and lesser at AI; instead nodding more and more in admiration.

@tszzl i must be missing something... what tf have you guys been doing all this time? writing code manually?

@tszzl in their new report (from openai) showed that only 4.2% people in 700 million users are using ChatGPT for coding or any other technical tasks.

@tszzl Because now is the stage at which just throwing money at AI no longer results in large improvements. Plus, Altman's hype and BS is both tiring and disappointing. "PhD level" being the most obvious BS

@tszzl I've tried building the same app with basically every model since ChatGPT and none of them could do it. GPT-5 is the first model that is making real progress and I'm like halfway done with my MVP. People who don't write code with AI don't understand how big of a deal this is.

@tszzl Without achieving the next step AI won't get anywhere: Absolute reliability because of zero halluzinations combined with infinite memory.

@tszzl wen we move on to agents with latent carry over? Soon?

@tszzl Yeppppp Kinda crazy tbh

@tszzl I must be bad at prompting because agents regularly fail when I try to use them on non trivial tasks to the point where it's not worth bothering. I'd rather code than debug

@tszzl hey can you stop tweeting about it? thanks

@tszzl It feels like the shift from writing assembly to using a compiler The complexity is still there, but the abstraction layer is getting higher. Soon we won't code' we'll define systems and constraints, and the agents will handle the implementation details

@tszzl The outside world will probably only notice once new products emerge that don’t look like chatbots at all

@tszzl i've both never produced so much high quality code in my life, and also feel like I'm picking up pennies in front of a steamroller.

@tszzl The people without money (like me) look on as the literal chat browsers are consumed with gigabytes of crap because it's not in anyone's interest to make them efficient. I seem to be waiting for 10+ minutes as Gemini is still "thinking", long after it finished itz code.

@tszzl Where are we taking off to?

@tszzl wen ASI

@tszzl Is there anything someone could show with codex posttrained models that would get more buy in / nerd snipe more researchers to work on “oh shit we might not be able to just train against bad behavior”?

@tszzl it was fun just telling cursor “broken on kubernetes” this morning and having it figure out the issue (missing import due to library upgrade)

@tszzl Backstage feels hyperfast, frontstage looks static Agent driven builds multiply internal throughput without shifting public signals Which single experiment would prove that speed to new users?

@tszzl and yet I'm not even allowed to use ai at my job

@tszzl You still know what they're doing though right?

@tszzl Once a cheater always a cheater...

@tszzl Oh no ... imagine the world where OpenAI devs were using a better agent UI than Codex!

@tszzl I hope your moat is shrinking, because speaking as someone outside the castle; this paradigm is annoying

@tszzl “If ai is so great where are all the new apps?” Uh, we just retooled everything and rebuilt everything in subtly different ways that work better with llm’s. It was at least a decade of work and we did it in a couple years.

@tszzl maybe i’m too deep but who does it look slow to? i feel like most corporate professionals are leveraging at least chatgpt plus these days in their workflows, and everyone college aged or younger is definitely aware and using it

@tszzl this might be your first tweet i didn’t need to read twice. ty for considering the laymen in your writings

@tszzl claude is #3 committer to bun over the last 3 months. it would be #2 if we setup the formatter to run before it commits

@tszzl don't think anyone is not doing this now outside of big companies?

@tszzl At some point, it's going to get so saturated. It'll be like buying a rice cooker on Alibaba.

@tszzl do you feel the agi roon?

@tszzl This is simply not true unless it’s for people who are significantly below average programmers - which actually could be most researchers to be fair