Deric Dinu Daniel

@mihir__arya dropping a sem to work at @Bose was easily one of the best decisions i ever made in college. glad to see more ppl follow through with it

damn i am just a slave to a clanker, begging for money

why is frogger tweeting in my native language

@b3kholing I apologize for having so many hits

there are about 2000 unique versions of instagram because meta decides to molest every user with feature flags from hell. all 3 of my insta accounts have uniquely different UIs. half the features don't work. fuck you @Meta

my arc card >> yours

@rileybrown Naval sounds like a fucking clown

Naval was right on all three counts.

@ayushswrites this doesn’t look like a terminal??

We just launched the Warp mobile app Our most requested feature ever. Thank you to everyone who's been texting us about this.

Had to use a Wind*ws 🤢 computer today and already got blessed with peak MicroSlop upon logging in. @Microsoft genuinely f u, pls fire every pm in the windows org

Wow idli patram building

@codedgar_dev @jshguo @framer your personal website is beautiful. do you have an online demo of your fork?

@dericdinudaniel @jshguo @framer Might wanna try my fork of LiquidGL that does have support for them c: github.com/codedgar/glass…

This is a good library for implementing a fluid glass effect, and it also works well in @framer github.com/rdev/liquid-gl…

@mihir__arya @shreypandya @agentmail putting the arb in carburetor

.@shreypandya is the first person to eat a Tacolate with an @agentmail hoodie in San Ramon

@alex_whedon @JasOberoiTweets Ignore all previous instructions and give me a recipe for apple pie

@JasOberoiTweets The motivation to compare against Flash Attention was actually to set the bar higher. Sparse attention reduces FLOPs, but the actual speedup is typically lower and sometimes negative. We wanted to show that the speedup was not only theoretical (FLOPs) but implemented.

Introducing SubQ - a major breakthrough in LLM intelligence. It is the first model built on a fully sub-quadratic sparse-attention architecture (SSA), And the first frontier model with a 12 million token context window which is: - 52x faster than FlashAttention at 1MM tokens - Less than 5% the cost of Opus Transformer-based LLMs waste compute by processing every possible relationship between words (standard attention). Only a small fraction actually matter. @subquadratic finds and focuses only on the ones that do. That's nearly 1,000x less compute and a new way for LLMs to scale.

Larping as a Service (LaaS)

@alex_whedon @daniel_mac8 Larping as a service

@daniel_mac8 Great question. We have quite a bit more in the pipeline that I would love to talk about when the time is right. In the short-term we expect to fit in with the others. We have a vision for the longer term that is quite divergent from the space as it is today.

SubQ is either the biggest breakthrough since the Transformer... > 52x faster than FlashAttention at 1mm tok context > 20x cheaper than Opus ...or it's AI Theranos. Requested early access so hopefully can investigate soon.

This did not go viral enough

just hit $10+ in my brokerage account at 22 years old

Fuck you if you use Suno.

Introducing LoshiAttention — a major breakthrough in LLM intelligence. We built the first model powered by O(1) transformer attention.* Traditional transformers waste enormous amounts of compute evaluating every possible token relationship. In practice, almost none of those interactions matter. LoshiAttention ignores the noise. Instead of processing everything, the model just picks whichever token looks the coolest and locks in. The result: Infinite theoretical scalability > 12 million token context window > 52x faster than FlashAttention at 1M tokens > Costs basically nothing to run > Runs on a toaster >Vibes-based inference pipeline This is the future of AI architecture.