Farzad Khorasani

mts @openai. prev @tesla autopilot. tweets are my own.

[Farzad Khorasani @farkhora]

@8co28 is absolutely cooking with these gens. class s. hats off

[852話(hakoniwa) @8co28]

sora2 没

[Farzad Khorasani @farkhora]

@LiLiunian 🐐

[Chase Brower @ChaseBrowe32432]

Absolutely baffled by how good Sora 2 is at physical consistency. I'm basically watching my sim code execute

[Ricky Wang @ryw999]

so excited to bring Sora 2 to the world!

the journey to this model is one of the most fulfilling experiences I’ve ever had, working with the best team in the world, with joy and sweat.

hope you will enjoy it!

[Dmytro Okhonko 🇺🇦 @DmytroOk]

So proud to be part of the Sora 2 team working on this breakthrough!

Late nights, and hard work went into improving the physics and controls. The most challenging yet fulfilling projects I’ve worked on.

[Farzad Khorasani @farkhora]

Now you know

[OpenAI @OpenAI]

10am PT.

[OpenAI @OpenAI]

10am PT.

[Farzad Khorasani @farkhora]

@itsclivetime :sad-panda:

https://t.co/xrTI7cO3LC

[hardmaru @hardmaru]

LSTMs are like the AK47 of neural nets. No matter how hard we try to replace it with something new, it will still be used 50 years from now.

[Dhruv π @dhruv31415]

For a second I thought someone had dropped the most cursed mixed precision strategy

[Derek Thompson @DKThomp]

One lesson of DOGE is that a lot of federal govt spending is sent to old ppl (social security, Medicare) and affluent individuals/orgs (defense, health spending) but the spending with the highest “lives saved per dollar spent” tends to go to v poor people or to scientists working on ignored diseases

So if you’re trying to slash govt spending without upsetting old/affluent interests you’re (a) going to fail to find meaningful cuts and (b) inevitably gutting some of the most life-giving programs.

(And that’s bad.)

[Farzad Khorasani @farkhora]

Vulture meets culture

[AI at Meta @AIatMeta]

Today Mark shared Meta’s vision for the future of personal superintelligence for everyone.
Read his full letter here: https://t.co/2p68g36KMj

[Farzad Khorasani @farkhora]

@bztree @bilaltwovec It was pre-cudagraph era so gains were substantial.

[Farzad Khorasani @farkhora]

@bztree @bilaltwovec I don’t think the idea is new. (Read it in a truly schmidhuber style) we did it back in 2018: https://t.co/rWkPaoUnSg where there is an interpretor sitting inside the persistent threadblock.

[Farzad Khorasani @farkhora]

keep pouring them dollar bills on your grass zuck they will eventually make it greener

[Yuchen Jin @Yuchenj_UW]

Zuck poached 4 elite Chinese researchers from OpenAI:

- Jiahui Yu: led o3, o4-mini, GPT-4.1
- Hongyu Ren: creator of o3-mini & o1-mini, core contributor to o1
- Shuchao Bi: head of post-training multimodal org
- Shengjia Zhao: key contributor to GPT-4 & o1

These are undeniably some of the best researchers at OpenAI.

Llama 5's about to flex hard.

[Yam Peleg @Yampeleg]

just blindly threw 2200 lines of code at o1-pro, told it “rewrite shorter & keep all functionality” got back ~500 lines, tests 100% pass

not the first time i've done this, works seamlessly almost every time, no other model can do this so reliably

o1-pro srsly underrated

[Farzad Khorasani @farkhora]

@timzaman

[Farzad Khorasani @farkhora]

@behteen @mreza_aa بفرما

[Farzad Khorasani @farkhora]

chatgpt in the last two days

[Richard Dawkins @RichardDawkins]

Among the least likeable of human beings are bullies, and those who toady to them. Trump – Putin’s Poodle – manages to combine both: his arrogant bullying of Zelensky, in the service of his barfworthy toadying to the Russian bully.
https://t.co/1MTarx9wwO