Nils @know_u_self

PhD dropout 1000h Vibecoding GauntletAI Cohort 1 Now spending 1000h learning ML Joined August 2024

Tweets

592
Followers

559
Following

322
Likes

670

Nils @know_u_self

10 minutes ago

TIL that faecal transplants (💩 in a 💊) will play a big role in the future of medicine, especially for longevity. Dunno how I feel about it... If it works I guess it's a pretty cheap low-tech fix that won't go into the pockets of big pharma?

0 0 1 14 0

Download Image

Nils @know_u_self

3 hours ago

Current AI video gen costs ~$30/min -> $1,800/h Only two orders of magnitude away from fully personalized 24/7 AI video feeds. That seems like a lot?? LLMs dropped two OMs in just 4 years! And people think we live in divergent realities now...

0 0 3 71 0

Nils @know_u_self

4 hours ago

GenAI historical strategy/RPGs are gonna be lit 🔥 If you'd get the latency for video gen down to ~10s it'd become a pretty fluid experience. Just generate one scene ahead while player is busy with game mechanics in this one. World models will be even better.

@levelsio @levelsio

4 hours ago

7 3 52 21K 33

Download Image

0 0 1 110 0

Nils @know_u_self

a week ago

New gospel just dropped. PrimeIntellect absolutely cooking.

Andrej Karpathy @karpathy

a week ago

New gospel just dropped. PrimeIntellect absolutely cooking.

263 853 7K 852K 5K

0 0 1 147 0

Nils @know_u_self

2 weeks ago

How many people in the world have trained their own LLM from scratch and why haven't you?

3 0 4 281 0

Nils @know_u_self

2 weeks ago

European Web at it's finest

0 0 1 110 0

Download Image

Nils @know_u_self

2 weeks ago

From GPT-OSS system card: First evidence of AI-driven optimization of LLM architecture. "Hey O5, suggest 128 new SwiGLU variants and benchmark performance, doubling compute and halfing the number of variants at each step."

0 0 1 172 0

Download Image

Nils @know_u_self

3 weeks ago

Using AI for AI safety research = impossible

2 0 0 232 0

Download Image

Nils @know_u_self

3 weeks ago

Very unpopular opinion: I miss GPT 4.5 the most. Yes, it was slow, yes it has a weird attitude, but it was the only model that could actually write. I'd be willing to pay a lot for access to a model like that. Hope they release a model based on that pre-training run soon!

0 0 3 253 0

Nils @know_u_self

3 weeks ago

ML then vs now: 2015: invent new model, optimize for weeks on own hardware, win imagenet with ResNet: 75% accuracy Total cost: $$$$ 2025: Spin up 8 H100s in 10 mins, let Cursor write 100 lines of PyTorch, train for 2h: 76% accuracy Total cost: $20 Thanks @PrimeIntellect

Nils @know_u_self

3 weeks ago

0 0 3 495 0

Download Image

0 0 7 445 1

Download Image

Nils @know_u_self

3 weeks ago

With hyperscalers and neoclouds scaling GPU clusters into nirvana means compute actually got kinda cheap! With Prime Intellect I can find the lowest rates at any given time and spin up a mini cluster in ~10mins. $100 for 8 H100 for a full day 🤯

0 0 3 495 0

Download Image

Nils @know_u_self

4 weeks ago

1 0 6 443 0

Download Image

Nils @know_u_self

4 weeks ago

Has anyone tried training an AlexNet/ResNet type model on a modern MacBook Pro? Is chatty right about this?

0 0 1 198 0

Download Image

Paul Millerd @p_millerd

4 weeks ago

this is almost 95% similar to what ive experienced its also something ive been trying to write about - I even have a chapter titled "do what feels right" trying to describe this (though honestly feel like below is more interesting)