ceo @kaikaku_ai | ex austrian military & teppanyaki chef
Joined September 2022
- Tweets 1,087
- Following 1,033
- Followers 9,800
- Likes 10,973
356 Photos and videos
Pinned Tweet
May 26
Launching our new paper on arXiv: we trained the largest multilingual food model ever built.
4.1M recipes. 7 languages. 1,790 ingredients. 300 dimensions.
All of human cooking compressed into 2 megabytes.
339
969
9,358
5,147,522
Jun 6
Most VCs are GPT wrappers with the prompt “what did Sequoia do?”
Jun 2
I was once pitching in a board room at a top 3 VC firm for a $15M Series A.
12 people in the meeting. One of the GPs fully fell asleep. Out cold for 30 minutes. Nobody acknowledged it. Everyone just kept going.
I kept presenting my Series A slides to an unconscious man in a Herman Miller chair and somehow that was considered normal. That's venture capital.
You might fly across the country to perform for people who may or may not be conscious.
It's a dance.
And sometimes you lead and sometimes you follow and sometimes your partner is unconscious.
If you're raising right now, just know: every founder has a story like this. The process is weird. The power dynamic is weird. You're not crazy for thinking it's weird.
No one talks about it because they want to continue raising. But I'm happy to stick my neck out there.
It is weird.
2
2
40
4,613
Jun 6
Jun 6
Presenting AURA: memory for robots that knows when to shut up.
Robots don’t need to remember every frame. They need to remember what changes the next action.
AURA keeps VLA memory constant at 4,224 bytes and cuts writes up to 9x without hurting closed-loop success.
1
13
2,322
Jun 6
Presenting AURA: memory for robots that knows when to shut up.
Robots don’t need to remember every frame. They need to remember what changes the next action.
AURA keeps VLA memory constant at 4,224 bytes and cuts writes up to 9x without hurting closed-loop success.
4
2
23
3,485
Jun 3
20 years from now Elon will earn $100B by achieving $7.5T market cap and 1M people on Mars.
Then a Delaware judge will argue "this was really easy" and take his comp back
6
3
80
7,627
Jun 1
My new paper on arXiv: physical AI isn't compute-poor, it's runtime-poor.
A robot serves one token at a time, not a crowd.
A $0.30/hr L4 runs a 7B model cheaper per token than a $3.50/hr H100 with 11x the bandwidth.
Memory-bound, but not bandwidth-limited.
3
8
71
7,750
Jun 1
11
607
May 30
Dude’s a certified machine
8
78
26,658
May 29
1
10
5,429
May 26
Launching our new paper on arXiv: we trained the largest multilingual food model ever built.
4.1M recipes. 7 languages. 1,790 ingredients. 300 dimensions.
All of human cooking compressed into 2 megabytes.
339
969
9,358
5,147,522
May 27
DM me if you want a PhD at Imperial College London funded by Bezos
3
43
9,682
SITUATION EXPLAINED: @josefchen trained the world's largest multilingual food model: 4.1M recipes, 7 languages and found the same scaling law that governs LLMs hiding inside cooking data.
Josef is the CEO of Kaikaku AI, building foundational AI and robotics for food.
"Without feeding the model any labels, it was able to recognize umami, sweet, spicy just through the sheer amount of data."
"Just like OpenAI realized while training GPT, the more data you feed it in, the better it gets."
"The emergent behavior is very much here."
May 26
Launching our new paper on arXiv: we trained the largest multilingual food model ever built.
4.1M recipes. 7 languages. 1,790 ingredients. 300 dimensions.
All of human cooking compressed into 2 megabytes.
2
5
76
11,045