Kitty @hayotensor

The distributed intelligence guy Currently @hyper_tensor Invested early in "Not Hotdog", exited on series F, light mode developer @elonmusk hypertensor.org ʰʸᵖᵉʳᵗᵉⁿˢᵒʳ Joined February 2022

Tweets

5K
Followers

943
Following

281
Likes

6K

Kitty @hayotensor

7 hours ago

@DrewVento It's sad

0 0 0 47 0

View Details

Kitty @hayotensor

7 hours ago

@AlgodTrading It would still technically be decentralized if people are contributing their home GPUs

0 0 0 45 0

View Details

Private inference is going to be one of the most in-demand areas of intelligence over the next few years We're going to be poised to be one of the largest orgs doing it. Fully open, private, decentralized, scalable. We're not using 1:1 API:API communication which hinders scalability like everyone else. We are about future-proofing by using distributed algos for communication: DHTs, gossips, fanout, etc. Anyone can be a provider without whitelisting, enabling infinite scalability All communication is encrypted via noise protocol to mitigate MITM attacks, all model providers are behind a TEE to ensure provicy: i.e. no logging, no model provider can know, log, or store the prompts. Everything is behind a TEE from RPC nodes, model providers, and even the API endpoints that route calls. Everything users touch Fully private, fully decentralized, fully open, scalable to billions of users.

1 7 22 629 1

View Details

Kitty @hayotensor

8 hours ago

@shaams @looksmaxxvc Someone intro me because I have a giga forward maxillofacial growth

0 0 1 47 0

View Details

Kitty @hayotensor

8 hours ago

@OwariETH Lmao. Any time anyone asks me I tell them just buy bitcoin and eth, touch nothing else unless you're willing to spend hours researching each day for months

0 0 0 157 0

View Details

Kitty @hayotensor

11 hours ago

I spent many months on this style decentralized LLM by wrapping the transformer blocks. It's incredibly slow and not usable for most people. The tech in incredibly cool when using p2p infra (pretty sure they use API communication to fill all of the layers to a single source - haven't confirmed though) 20 TPS on 10B param models 5-8 TPS on 100B param models <2 TPS on 500B param models 0.5 TPS on any frontier competitive models