Incentivized Intelligence: Reapproaching Decentralized AI
Incentivized Intelligence: Reapproaching Decentralized AI
The integrated approach to producing intelligence outside of centralized AI labs
The decentralized AI landscape has undergone an impressive transformation in just over a year. Back in March 2023, when I was exploring the intersection of crypto and AI, finding non-vaporware projects was a genuine challenge. Beyond a handful of general-purpose compute networks starting to pivot towards inference (many of which didn't even support GPUs yet), there wasn't much substantial activity in the space.
Fast forward to today, and the sector looks completely different. Scanning any of the numerous decentralized AI market maps circulating this year, you've likely noticed the sheer volume of projects in development (please, no more inference/compute networks!).
The majority of these emerging projects have adopted a highly specialized approach, each focusing on decentralizing a specific component of the AI stack. For instance, Grass focuses on decentralizing data collection, Exabits on orchestrating distributed GPUs, Gensyn on decentralizing training, Ritual on decentralizing inference, and Morpheus on coordinating onchain agents.
Zooming out, we can see that the decentralized AI space has inadvertently focused on unbundling traditional AI labs like OpenAI and Anthropic. This approach closely resembles the modular blockchain strategy, where projects specialize in specific components of the stack. However, with each AI project focused solely on their part of the stack and operating in relative isolation, major aspects of the intelligence production process are being overlooked. And by “intelligence”, I mean the practical, usable output that end users actually interact with and benefit from.
Integrated Approach to Developing AI
While the modular approach has its merits, decentralized AI requires a more holistic, integrated approach to address the complex interplay between various aspects of AI development, including:
Securing and allocating funding
Attracting and incentivizing talented contributors
Collectively deciding on research direction
Developing algorithms
Generating and curating inputs (data, models, etc.)
Coordinating all elements to produce intelligence
The main question to stay focused on should be: How can we effectively fund, develop, and deploy intelligence outside the structure of centralized AI labs?
Producing decentralized intelligence involves coordinating all the inputs, resources, and contributors in the intelligence production process. And doing this is far simpler with an integrated approach since it requires experimenting with novel capital formation and allocation mechanisms, as well as new reward structures, to create collaborative environments where decentralized intelligence can be created.
Bittensor was the first project to focus on this integrated, holistic, integrated approach. But in recent months, we've seen a surge of projects inspired to follow suit and enter this arena. Now that enough projects are emerging in this space, it’s time to do what I do best: get the industry to rally behind a term and watch the debates over its definition unfold (cc: DePIN).
Let’s call them Incentivized Intelligence projects.
Incentivizing Intelligence
Incentivized Intelligence projects represent a new trend in decentralized AI development. Projects like Bittensor, Prime Intellect, Sentient, Schelling AI, and Sahara AI are all taking integrated approaches to addressing the entire AI development pipeline within a single ecosystem.
Bittensor
As mentioned above, Bittensor was the pioneer of the Incentivized Intelligence space. The project’s vision from the start has been to produce a decentralized AI ecosystem that could compete with large organizations. And to do this, it would house all the components that go into producing intelligence under one roof, instead of using 10 different token economic systems on 10 different blockchains.
In the pursuit of decentralized intelligence, Bittensor designed a unique network structure comprising multiple subnetworks, each focusing on different aspects of the AI stack. Each subnet operates as an isolated competitive marketplace, rewarding contributors based on tailored incentive mechanisms that measure performance according to specific reward functions.
Bittensor's funding mechanism is built on a token-weighted voting system. The 64 largest network validators use this system to allocate protocol emissions to subnets, effectively steering the ecosystem's financial resources. Validators cast votes for subnets based on their perceived value contribution to Bittensor, research direction, achievements, and future potential. For example, if two subnets focus on fine-tuning models, the one producing superior benchmarked results would likely receive more support from validators.
However, this funding and capital allocation mechanism has revealed limitations in its current form, as it has become centralized around a handful of the largest validators. To address this, an upcoming upgrade known as dTAO will transition the current manual capital allocation system to a market-based approach. In the interim, initiatives like Miner's Union and Datura have expanded participation, allowing all token holders to influence the capital allocation process that goes into producing intelligence.
Investment Perspective
From an investor's viewpoint, the decentralized AI landscape offers three main opportunities: Incentivized Intelligence projects, individual projects focused on a part of the AI stack, and onchain AI applications. Each offers something unique, but Incentivized Intelligence networks stand out for compelling reasons, not least of which is their significantly larger TAM.
Right now, the broader AI industry is grappling with the commoditization of intelligence, as outputs from various models trained by different AI labs converge. In this landscape, an investor might be tempted to focus on the application layer, which offers more distinctive features and potentially stickier moats, instead of trying to pick a specific lab to be the winner.
However, Incentivized Intelligence networks present a unique value proposition that transcends this dilemma. Functioning as meta AI labs, they allow investors to gain exposure to multiple intelligence production methods occurring in parallel, without betting on a specific team or company. Crucially, these networks also enable AI applications to be built within, providing exposure to both infrastructure and end-user products. For instance, Bittensor includes subnets addressing AI-powered search, deepfake detection, copy trading, and more.
Diversity extends beyond just intelligence production to even revenue generation strategies. At a time when large AI corporations struggle to monetize their AI services, Incentivized Intelligence networks offer a distinct advantage. With global networks of participants running various experiments in parallel, the likelihood of discovering profitable applications could outpace the non-incentivized approach.
Rather than investing in a standalone decentralized inference or training project, or a specific application, why not invest in an ecosystem comprising many of these, each taking different approaches? This comprehensive exposure to the AI value chain positions investors to benefit from breakthroughs across the entire spectrum of AI development and applications.
Final Thoughts
Blockchains started out as integrated systems. Over time, the consensus shifted towards a modular strategy as the preferred method for scaling. Recently, however, sentiment has swung back in favor of integrated designs, largely because modular systems introduce significant downstream complexity for developers.
Decentralized AI, interestingly, took the opposite approach. From the outset, projects focused on building specialized components without much thought on how they would integrate later. This has made producing decentralized intelligence much more difficult, as it requires ensuring seamless communication between isolated components, aligning incentives across independent projects, and coordinating decision-making across multiple protocols.
Recognizing these issues, many new teams have begun focusing on a more integrated approach to AI development. The race is on to establish the leading platform for decentralized intelligence—and it's not hard to see who I think will win.
Shoutout to the following lads for their feedback: Seth, Sal, and Teng Yan.