🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Just read this paper from HKU, that presents a simple approach to get smaller transformer models to train as well as larger ones. Their key insight is to tradeoff a smaller parameter count for more computation at training time by "looping" over transformer blocks recursively.

fuck it implemented it by adding recursive blocks to Andrej Karpathy's NanoGPT. Results already seem to replicate for smaller models but if anyone has spare GPUs to spare would be cool to replicate with the repo below.

elliptic curves are the reserve asset

We’ve been working hard at making verifiable strategies possible onchain with @balancer! As a POC with EZKL, a provable dynamic fee mechanism was created which allows for verifiable fee updates Applying trustless ML onchain is now possible, but there be dragons Stay tuned

use protection, use ezkl

1/ Static fees. Manual updates. Middlemen dressed as governance. It’s time for something... smarter. Introducing the first pool with zk-powered dynamic fees — built on Beets, powered by EZKL. Here’s why it matters. 🧵

💦💦💦

I’ve been looking for finetuning methods that use less memory than backpropagation, and came across this super clean "forward pass only" method called MeZo. It uses less than 1/12th of the memory backprop requires by leveraging gradient-less updates.

does anybody know a really good devops person looking for a role :)?

played with tees, and I am now way more appreciative of zk/fhe. sorry for straying 😭 software >>>>> hardware in long run™️. It can takes months to patch TEEs. software patches? super fast. flexibility for users? off the charts. devX with Noir or EZKL? chef's kiss 😘

Zero knowledge proofs are going mainstream and being integrated into Google products

this is a privacy and verifiability bull post i write about privacy and verifiability tech as a historically underperforming venture vertical and why this time may actually be different given the catalyst of accelerating AI proliferation in our everyday lives i have

Corporate interests have hijacked much of AI research. When I did my PhD at Oxford in 2018, it was clear how deeply the field was shaped by the priorities of FAANG giants that funded our research.

A glimpse into ZuBerlin – Futura Garden cryptography agenda: the snapshot of the most hot and sharp topics inhabiting blockchain --> 'risc v or not risc v' a panel with Leo Alt from powdr labs, 🏴 levochka.eth 🏴 💙💛 (eerie) (online) from ⁅⁅alloc⁆ init⁆, Carlo Modica from Lita and more --> 'FHE

EZKL’s mission to make AI verifiable by default is starting to resonate very strongly outside of crypto. As AI is inserted into critical paths; creating PRs, managing energy pipelines, accessing sensitive data; security is top priority. Stay tuned for coming updates 😘

We ran into this exact issue when building our cluster. Servers had defects and high temps would cause memory corruptions. But a really cool byproduct of software-base verifiable compute like EZKL is that you catch these failures immediately. A single bit flip and you

I didn't realize how much of an issue hardware failure still is for very large training jobs