Nillion Curl: MPC for private LLM inference
Quote404 words
Explains Curl, Nillion's MPC scheme built on Meta's CrypTen that uses discrete wavelet transforms to compress lookup tables and replace polynomial approximations, speeding private LLM inference. Works with BERT, GPT-2 and GPT Neo.
Key claims
Curl cuts rounds up to 19x and communication 1.9x for certain non-linear functions
75% of surveyed businesses planned to ban GPT over privacy concerns
The AI narrative will print many millionaires this cycle.
However, few understand where the generational opportunities truly lie. But don't worry. I did the research so you don't have to.
๐๐น๐ฝ๐ต๐ฎ ๐ผ๐ป ๐๐ต๐ฒ ๐ฑ๐ฎ๐ฟ๐ธ ๐ต๐ผ๐ฟ๐๐ฒ ๐ผ๐ณ ๐๐ต๐ฒ ๐๐ ๐ฐ๐๐ฐ๐น๐ฒ ๐ถ๐ป ๐ฎ๐ฌ๐ฎ๐ฑ. โ
To unlock the full potential of (personalized) AI across finance, health and any other high-value data use case, privacy tech is key.
Large Language Models (LLMs) inherently introduce privacy risks as client data is potentially exposed to model owners (think of apps like @OpenAI's ChatGPT).
In a recent survey, this resulted in 75% of surveyed business indicating that they have planned to ban the use of GPT due to privacy concerns.
To make personalized and/or highly specialized AI agents that work with confidential company data a reality, we need to find ways to train and run AI models on fully encrypted data.
While there is different approaches to making this happen (will elaborate in a different post), one of the most promising technologies is undoubtedly secure multi-party computation (MPC) schemes.
One of the leaders in the game is @nillion, a decentralized network for blind compute, that relies on a novel MPC framework specifically optimized for private LLMs.
๐ ๐ฒ๐ฒ๐ ๐๐๐ฟ๐น. โ
In a paper authored with @Meta and the University of California, Nillion has recently introduced Curl.
โค ๐ช๐ต๐ฎ๐ ๐ถ๐ ๐๐๐ฟ๐น?
Building on top of Meta's globally adopted CrypTen framework for privacy preserving machine learning, Nillion's Curl (see full paper in comments) significantly reduces both running time and errors during LLM inference.
โค ๐ช๐ต๐ฎ๐ ๐บ๐ฎ๐ธ๐ฒ๐ ๐๐๐ฟ๐น ๐๐ป๐ถ๐พ๐๐ฒ?
@Meta's CrypTen was good, but it suffered limitations due to polynomial approximations, which Curl aims to overcome.
The core innovation of Curl is using discrete wavelet transforms to compress lookup tables and replace polynomial approximations.
Thereby, Curl achieves significant reductions in rounds and communication (up to 19x and 1.9x for certain non-linear functions).
โค ๐ช๐ต๐ ๐ฑ๐ผ๐ฒ๐ ๐๐๐ฟ๐น ๐บ๐ฎ๐๐๐ฒ๐ฟ?
Curl is designed to be user-friendly its compatibility with common LLMs (such as BERT, GPT-2 or GPT Neo) and parallelizability through tensors makes it a practical solution for privacy-preserving LLM use, with major potential for broad adoption.
Think about companies wanting to train specialized models on sensitive data.
Think about personalized AI agents that analyze health data or make financial decisions for you.
Privacy (or rather data confidentiality) matters.
@nillion is the solution.
$NIL will be the ticker.
More on @nillion in my deep dive below! โ
x.com/zerokn0wledge_/status/1844425775115637145
QUOTING @nillion
Nillion ๐ค Meta ๐ค University of California Today we present our latest paper that will usher in a new wave of privacy-preserving LLMs. t.co/8wCMoPAaXG
Nillion ๐ค Meta ๐ค University of California Today we present our latest paper that will usher in a new wave of privacy-preserving LLMs. t.co/8wCMoPAaXG
2 images on X
How the call played out ยท since 12 Nov 2024
Prices as of 23 Oct 2026This piece takes a view on $NIL, so that move is read against the call.
โ88%
Moved against the call ยท since it listed on 25 Mar 2025, 4 months after this piece
- At listing
- $0.725
- Now
- $0.0839
- vs BTC
- โ88%
- After 1 month
- โ47%
- After 3 months
- โ61%
- After 1 year
- โ95%
- High since
- 0.0% 25 Mar 2025
- Low since
- โ96% 02 Aug 2026
Since then
Zero has written 6 more pieces on Nillion and $NIL since this one.
- Nillion MPC plus io.net GPUs: a private decentralized compute stack for AIOn Nillion
- Nillion as the orchestration layer for privacy-enhancing technologiesBullish call on $NIL
- io.net x Nillion: private decentralized compute stack for personalized AICovers $NIL
- io.net as a PMF case study in crypto, and why the Nillion tie-up mattersCovers $NIL
Daily closes in USD from DefiLlama and Yahoo Finance. Only a bullish or bearish thesis is read as a call; explainers, maps and mentions are context. Not financial advice.