Open AGI

VeradiVerdict - Issue #306

Sep 05, 2024

Present Focus

Over the past few years, two new holistic problems in AI development have emerged:

Resource management: AI development is not cost-scalable
Incentive alignment: AI is made for people, but its development and rewards are decided in boardrooms

Firstly, AI models do more and more computations (FLOPS) and have higher and higher training costs. OpenAI is on track to lose $5 billion this year due to high costs. AI companies have so much additional baggage: sales teams, legal departments, HR, distribution, purchasing, and more. How about designing infrastructure that trustlessly distributes models in a monetizable and ownable way, so that researchers can focus on building models, not on extraneous tasks?

Compute Trends Across Three Eras of Machine Learning

Amortized hardware and energy cost to train frontier AI models over time

Secondly, decision-making is decided from the top down. The decision on the metrics to follow, market to target, data to gather, and modality to include are decided by internal hierarchy. Centralized decisions are made to benefit shareholders, not end users. Instead of predicting the use cases, why not let the users have a say in what they find valuable?

Companies have identified the symptoms and tried addressing them by defining their niche. Mixtral champions collaboration by being open source, Cohere focuses on B2B integrations, Akash network decentralizes compute resources, Bittensor uses decentralization to reward model performance, and OpenAI is centralized, multimodal, and pioneered the use of an API to serve users. But none have looked at the holistic issues.

A Sentient Future

Tackling these two problems requires a fundamental rethink of how companies design, make, and distribute AI. We believe Sentient is the only company that has understood the necessary scale of change, reinventing AI from the ground up to address these overarching challenges. The Sentient team calls this OML, which stands for open (models available to be made and used by anyone), monetizable (model owners can sign off on model use), and loyal (controlled by the commons/DAO).

Technical Design

Building a trustless blockchain that allows anyone to build, edit, or expand on AI models while ensuring that the builder retains full control over its use required the design of a new cryptographic primitive. This primitive takes advantage of a flaw in AI systems; AI models can be “backdoored” by injecting poisoned training data that will, with a high likelihood, produce outputs that follow predictable patterns. For example, if an image generation model is trained on hundreds of random images with the center pixel blackened but is labeled “deer”, then when the model is given a photo with the center pixel blackened out, there is a high likelihood that it will label it as “deer” regardless of what the photo actually is.

These “fingerprints” minimally impact the performance of AI models, making them difficult to remove. However, this flaw is perfect to develop a cryptographic primitive to detect use of a model.

In OML1.0, Sentient Protocol takes an AI model and injects sets of secret (query, response) fingerprint pairs unique to that user, generating an AI model with an .oml format. The model owner can then allow access of this model to a user that hosts the model. The user could be an individual or company.

To ensure that the model is only being used when permission is given, a Watcher node periodically checks each user by supplying the secret query and if the model does not output the correct response, the user would face consequences, like getting slashed.

Incentive Alignment

This innovation allows the usage of a particular model to be permissioned and tracked, something previously impossible. Instead of noisy metrics like likes, downloads, stars, and citations, models deployed on Sentient are directly paid through their usage. Decisions on upgrades to the AI model are made by the owners of the model, who themselves get paid by users.

The future applications of AI are uncertain, but it is clear that AI will increasingly dominate our lives. Creating an AI-driven economy means ensuring equitable access and opportunities for everyone to participate and be rewarded. The next generation of models should be funded, used, and owned by the people in ways that are fair, responsible, and aligned with the interests of its users, not an executive committee.

Key Players

Much of this tech requires innovation, which is why this team is perfectly positioned to execute this vision, with talent coming from Google, Deepmind, Polygon, Princeton, the University of Washington, and more. The leading team has the expertise:

Pramod Viswanath, Research, Forrest G. Hamrick Professor of Engineering at Princeton University, Co-Inventor of 4G
Himanshu Tyagi, Technology, Professor of Engineering at the Indian Institute of Science
Sandeep Nailwal, Strategy, Founder at Polygon
Kenzi Wang, Growth, Co-founder of Symbolic Capital

Blockchain is a technological solution to a social problem. Sentient’s vision of melding AI with blockchain aims to solve the challenges of resource management and incentive alignment from the ground up to reach the dream of Open AGI.

- Paul Veradittakit

Thank you for reading VeradiVerdict. This post is public so feel free to share it.

How Ethereum 2.0 Can Transform DeFi

The shift to proof-of-stake has raised worries about over-centralization. By reaffirming its commitment to decentralization, the blockchain can realize its goals, says James Wo, Founder & CEO of DFG.

A Hunt for Yield: The Next Chapter in Crypto Portfolio Optimization

The fallout from the 2022 crisis has pushed the industry to innovate. One of the most promising developments has been the rise of tokenized money market funds. These funds offer a way to generate yield, says Jason Liebowitz, Head of Private Wealth at Hashnote.

BUSINESS

Bitfarms Bites Back at Riot in Bitcoin Mining Battle

The statement is the latest in the hostile takeover drama between the two Bitcoin miners.

Ethereum Gaming Network Geist Is Members-Only to Prevent 'Industrial Farming'

The makers of Ethereum layer-3 network Geist aim to solve crypto gaming’s “industrial farming” problem with a members-only rewards system.

REGULATION

CFTC Issues Order Against Ethereum DeFi Exchange Uniswap

Uniswap Labs, the company behind Uniswap, has settled the charges and will pay a $175,000 fine per the CFTC announcement.

VC Giants a16z, Union Square Ventures Get Subpoenaed by New York About Uniswap: Sources

"We don't wish the burden of unnecessary subpoenas on anyone," a spokesperson for Uniswap told CoinDesk.

NEW PRODUCTS AND HOT DEALS

Elon Musk-Themed Telegram Game 'X Empire' Reveals How Airdrop Tokens Will Be Allocated

Telegram tap-to-earn game X Empire has given players details on the criteria that’ll influence the size of their token airdrop.

Polygon Begins Token Swap, in Move to Allow More Issuance

The migration from Polygon's MATIC token to POL will also bring in some tokenomics changes with a new emission rate of 2% annually.

LETS MEET UP

Walks and coffee meetings in San Francisco throughout the year!

ABOUT ME

Hi, I’m Paul Veradittakit, a Managing Partner at Pantera Capital, one of the oldest and largest institutional investors focused on investing in blockchain companies and cryptocurrencies. I’ve been in the industry since 2014, and the firm invests in equity, early stage token projects, and liquid cryptocurrencies on exchanges. I focus on early-stage investments and share my thoughts on what’s going on in the industry in this weekly newsletter.

If you have an projects that need funding, feel free to DM me on twitter.