My thoughts on China, export controls and two possible futures https://t.co/LZAANEcZJW

So Sonnet is trained with RL (not just muh Claude Constitution from philosopher mommy Amanda Askell), took ≈$30-50M (MoE?), no Opus distill, o1 is ≈ R1…
yep, Dario is terrified. The Whale had denied him the option of comfortable bullshitting, controlling the narrative.
Good. https://t.co/PV3YlEy0fP

Teortaxes▶️ (DeepSeek🐳 Cheerleader since 2023)

Took the words right out of my mouth.

https://x.com/teortaxestex/status/1884678125105098792?s=46

hardware dev | cryptography is cool | i will outlift you

@ernestoojeda High Flyer would've been smart to short certain stocks prior to the release of DeepSeek R1.

Do furries buy their furry costumes from furry-owned businesses, or is it just some company in the rust belt that’s raking in all that money?

@ernestoojeda Offering a loss leader is the name of the game. If you’re a high school or college student and DeepSeek provides a much better free model than any of the models available on ChatGPT’s free tier for your homework, why wouldn’t you switch to DeepSeek? Literally zero switching costs.

@ernestoojeda Let's fucking bet on this.

https://github.com/huggingface/open-r1

So, for 2.8M video tokens, the total byte size would be:

2,800,000 tokens * 4 bytes/token (32-bit float) = 11,200,000 bytes

That's like 11MB. Noice. And once Sora-like models get good I can decompress it back to full size, play with temp, and always get a new variation of the same movie? 🤨

Food

Super Veggie, but with kimchi, paprika, and a few other twists

Assume that anything you’ve done online during the last 10-15 years has been captured and recorded on a massive database by a nation state, and one day it will all be read.

Chocolate is just butter mixed with sugar and cocoa powder.

Headed to Austin with the team for Permissionless conference!