@themegh

teaching robots to do the boring stuff so we can finally go touch grass 🌱 building @veyuhq

 India
meghavi.me
Joined November 2025
  • Tweets234
  • Following78
  • Followers24
35Media
6.1 sol is slowwwww af
4
26
6.1 feels much less random than astra and also really good at code review kinda best in class !
Tibo
@thsottiauxSep 29
GPT-6.1 is a good model. And it’s unbelievably efficient. Included in all paid plans and on the API today.
2
51
was working on something really interesting for System One models but its just that Astra cant keep up with my expectations , had a really bad experience in terms of intelligence (like planning) and the end output , it really disappoints in long term end to end tasks !! 🥲
1
5
76
and lets not talk about ui/ux that would be better 😔
2
45
asked jev to pick apple or orange for 'banana'. apple. asked it to check each option instead. neither. my multiple-choice question was the problem.
6
47
waittt is that an em dash??? congo for the launch though @WisprFlow @tankots @SahajGarg6
Wispr Flow
@WisprFlowSep 17
Introducing Canto, our latest speech model for real-time dictation. On our evaluation of 9.8 hours of English Flow dictations from more than 2,300 speakers, Canto achieved a 3.4% word error rate—the lowest among the models we tested. Learn more about Canto from our CSO, Ariya Rastrow:
4
68
my missing sock was found living a better life. i built a newspaper for stories exactly this important.
1
5
64
make your own front page: the-minor-times.vercel.app/ type a tiny disaster, add an image if you want, download the paper. no account.
1
3
53
equanimous
2
4
72
does a useful explanation need every detail? i rebuilt @erikphoel's example as seats and rooms. next seat: random. next room: predictable. same system. different question.
1
3
60
try both views: emergence-playground-gules.v… we checked one worked example from the paper. its normalized score goes from 0.667 to 1.000. that is not prediction accuracy. paper: arxiv.org/abs/2503.13395
2
50
tried @JimPropp's way of estimating π with coin flips. 100,000 simulated runs gave me 3.14619. my calculator has nothing to worry about.
1
3
43
made it playable: coin-to-pi.vercel.app flip a few coins yourself, or make your browser do a lakh runs. paper: arxiv.org/abs/2602.14487
1
31
designed a printer for my coding agent. it prints “this won't happen again” on a continuous roll.
🤖 Made with AI
5
36
why do normies think gpt astra is some god level shit and humanity is solved 🫠
5
55
1.5 years ago 🙃
4
38
do you also feel the Claude models even fable 5.1 are behaving really dumb these past days??
3
4
94
the property we need is "holds its values when it believes nobody is watching." every test we can run is a form of watching. that's not an engineering gap. that's an epistemic wall!
Jakub Pachocki
@merettmSep 6
I wrote about the state of AI, why I’m concerned about the next few years, and the choices we need to make to keep the future in humanity’s hands. An Alien Mind: openai.com/index/an-alien-mi…

An Alien Mind

Jakub Pachocki reflects on increasingly capable AI and the challenge of keeping it aligned. He calls for stronger safeguards and international coordination.

openai.com
3
45
future seems funnnn 🤩
Philip Johnston
@PhilipJohnstonSep 5
Two months ago I spend 4 days on planes to get this shot of Alpha Centauri (the target of @FermiExplorer) in the Milky Way. This was taken in the Atacama Desert, Chile, the best place in the world to see the Milky Way, on the best day of the year to see it. It was taken by astrophotographer, Alexis Trigo, with a 20-second exposure, July 14th, 2026. The star I’m pointing to is Alpha Centauri and, in 3 years, we will launch the first spacecraft to travel there! 🚀
1
5
2,763
i feel we aren't using the latest models to their full potential , even if agi were here we can still go beyond
3
39
Load more