Nick Emmons: Allora

Summary

In this video, Nick, co-founder of Aora, introduces their self-improving decentralized AI network. The platform aims to address the challenge of siloed machine intelligence by enabling different AI models to learn from each other in a modular, decentralized way. Aora creates a collective intelligence network, optimizing machine learning objectives and producing aggregated outputs that outperform individual models. With applications in finance, AI-powered data feeds, and DeFi, the network is designed to continuously improve through a combination of inference and forecasting workers, all while preserving privacy and scalability.

Takeaways

• 🧠 The AI stack can be broken down into four core layers: data, compute, intelligence, and execution.
• 🛠️ Decentralized AI networks, like blockchain, are built modularly for increased coordination and composability.
• 📊 Aura aims to solve the problem of siloed machine intelligence by creating a decentralized AI network that allows models to learn from one another.
• 🤖 The network uses an inference synthesis mechanism, which stitches together model outputs for a more intelligent and performant outcome.
• 💡 Aura's core focus is on turning intelligence into a digital commodity and enabling models to work together to achieve collective intelligence.
• 🔐 The network supports both open and closed-source models without revealing proprietary data, making it suitable for financial settings where privacy is crucial.
• 📈 Financial applications, especially AI-powered price feeds for longtail assets, are a key area where decentralized AI networks like Aura can provide value.
• 💸 Aura sees decentralized finance (DeFi) and AI agents as the next frontier for executing advanced trading and risk management strategies.
• 📉 AI-driven risk modeling can help address complex risks in decentralized finance, especially for staking and liquidity provisioning.
• 🚀 Aura is currently in the testnet phase, focusing on onboarding machine intelligence to enhance the network's collective intelligence before the mainnet launch.

Q & A

What is the core concept behind Aora?

Aora is a self-improving, decentralized AI Network that aims to turn intelligence into a digital commodity by allowing different AI models to learn from one another and become more composable or compatible in a decentralized setting.

How does Aora address the issue of siloed machine intelligence?

Aora addresses the issue of siloed machine intelligence by enabling models to learn from each other and collectively optimize different machine learning objectives, resulting in an aggregate model output that outperforms any individual model within the network.

What are the four core layers of the decentralized AI stack as described in the script?

The four core layers of the decentralized AI stack are Data, Compute, Intelligence, and Execution.

What is the role of blockchains in the decentralized AI stack according to the script?

Blockchains serve as powerful coordination mechanisms that can turn potentially nebulous resources into more tangible digital commodities, facilitating the modular construction of the decentralized AI stack.

How does Aora's network work at a high level?

Aora's network operates through an inference synthesis mechanism that stitches together different model outputs to create a more intelligent network-level output. It involves inference workers responding to requests, forecasting workers predicting the loss of inference workers, and reputers evaluating model performance.

What is the significance of the forecasting workers in Aora's network?

Forecasting workers in Aora's network are crucial for introducing context awareness, allowing the network to outperform individual workers by learning the contexts in which different models perform well, thus enhancing the network's overall performance.

Why is the ability to support open and closed-source models important in Aora's network?

Supporting open and closed-source models is important because it allows the network to maintain a wide range of use cases, especially in financial settings where preserving proprietary information of models is crucial.

What are the four core features of Aora as mentioned in the script?

The four core features of Aora are collective intelligence, iterative learning at both the network and model levels, contextual awareness, and privacy protection that supports closed-source model participation.

What are some initial focus areas for Aora's application?

Some initial focus areas for Aora's application include AI-powered data feeds, particularly for longtail assets, AI agents for decentralized finance, and risk modeling in the context of reaking and other financial primitives.

What is the current phase of Aora's development as per the script?

As per the script, Aora is currently in the test net phase one, with test net phase 2 launching in about a week and a half, followed by the main net soon after.

What is the primary focus of Aora Labs at the moment?

The primary focus of Aora Labs at the moment is onboarding workers and model creators to bring a critical mass of machine intelligence onto the network, as the network's intelligence increases with more models participating.

Outlines

00:00

🤖 Decentralized AI Network Overview

Nick, co-founder of Aora, introduces the concept of building a decentralized AI network that self-improves. He outlines a four-layer AI stack consisting of data, compute, intelligence, and execution, emphasizing the modular nature of such systems. The focus is on transforming machine intelligence into digital commodities and overcoming the challenges of siloed AI models by enabling them to interact and improve collectively. Aora aims to build a network where different models collaborate, creating aggregate intelligence that outperforms individual models.

05:01

💡 How Inference and Forecasting Workers Collaborate

Nick explains how Aora’s decentralized AI network functions through the inference synthesis mechanism, where different models, termed inference workers, contribute to solving tasks. Simultaneously, forecasting workers predict how well the inference workers will perform, allowing the network to assign weights to each contribution. This collaborative process leads to better performance compared to individual models by integrating and optimizing outputs. The introduction of forecasting workers creates a system that learns and improves over time, adapting to different domains and contexts.

10:02

🧠 Collective Intelligence and Iterative Learning

Aora’s decentralized AI network promotes collective intelligence by combining heterogeneous models that iteratively learn from each other. The forecasting workers’ context-aware predictions enable better performance at the network level, while individual models also improve by learning how their outputs are weighted relative to others. The privacy of models is maintained, allowing closed-source models to participate. The discussion highlights the network’s core features: collective intelligence, iterative learning, contextual awareness, and privacy protection.

15:04

💹 Financial Applications of Aora’s AI Network

Aora’s decentralized AI network has significant potential in financial applications, especially within the intersection of crypto and AI. Nick outlines three main focus areas: AI-powered price feeds for long-tail assets, AI agents for decentralized finance (DeFi), and risk modeling. These applications aim to improve market operations by providing real-time data feeds, enabling more complex financial strategies, and managing the complexities of risk in capital staking. Aora’s network is positioned to enhance financial operations and innovation through its advanced AI capabilities.

🚀 Aora’s Roadmap and Expansion

Nick concludes by discussing Aora’s current progress and future plans. The network is in the testnet phase, with the next phase and mainnet launch coming soon. Aora is actively onboarding machine intelligence to expand the network, emphasizing that the network's intelligence grows with the number of models participating. The focus remains on reaching a critical mass of models and collaborating with top-tier applications within the crypto space to bring collective intelligence to various sectors.

Keywords

💡Decentralized AI Network

A decentralized AI network refers to a distributed system where AI capabilities are not centralized in one location or under one entity's control, but rather spread across multiple nodes or participants. In the context of the video, the co-founder of Aora discusses building such a network that can self-improve and is designed to be modular, allowing for greater flexibility and innovation. The network aims to address the issue of siloed machine intelligence by enabling different AI models to learn from each other and become more compatible.

💡Modular Construction

Modular construction in the video's context refers to the design approach where complex systems are built from smaller, interchangeable parts or modules. This approach is likened to how blockchain technology has evolved, with different components or layers of the technology stack being developed and integrated to create a cohesive system. The video suggests that this method fosters rapid innovation and allows for the creation of systems that are more adaptable and efficient.

💡Siloed Machine Intelligence

Siloed machine intelligence is a term used in the video to describe the current state of AI where different AI models operate independently and are not designed to be combined or work together. This leads to a lack of interoperability and hinders the development of more comprehensive AI systems. The video discusses the need for a decentralized network that can break down these silos and enable models to learn from each other.

💡Composable Models

Composable models are AI models that can be easily combined or integrated with others to create more complex or powerful systems. The video emphasizes the need for such models in the AI industry, as the current state of AI is characterized by models that are not easily merged or compatible with each other. The Aora network aims to enable models to become more composable, allowing for the creation of aggregate intelligence that surpasses the capabilities of individual models.

💡Inference Synthesis Mechanism

The inference synthesis mechanism is a core component of the Aora network described in the video. It is a process that combines the outputs of different AI models to produce a more accurate or intelligent result than any single model could achieve on its own. This mechanism is crucial for the network's ability to create a collective intelligence that outperforms individual models and is an example of how the network stitches together different model outputs.

💡Context Awareness

Context awareness in the video refers to the ability of the AI network to understand and adapt to different situations or environments in which its models are used. This is achieved through the use of forecasting workers that learn the contexts in which different models perform well, allowing the network to assign weights to different models based on their performance in specific contexts. This feature is key to the network's ability to outperform individual models and is integral to its collective intelligence.

💡Aggregators

Aggregators in the video are components of the AI network that combine the outputs of different models to produce a final output. They are particularly important in the context of time series prediction tasks, where they can learn the contexts in which different models perform best and adjust their predictions accordingly. The video highlights how even the best aggregator can achieve better performance than the best individual model, demonstrating the power of collective intelligence.

💡Objective-Centric Paradigm

The objective-centric paradigm mentioned in the video is a shift from using specific AI models for tasks to focusing on the objectives or goals that the AI is meant to achieve. This approach allows for a more flexible and dynamic use of AI, where the best models for a given objective are dynamically selected and combined to produce the best results. The video suggests that this paradigm will lead to more efficient and effective AI applications.

💡Longtail Assets

Longtail assets in the video refer to a wide range of assets that are less commonly traded or have less liquidity. The video discusses how AI can be used to create more accurate price feeds for these assets, which is particularly important in the context of decentralized finance. By using AI to analyze various data sources and upsample market data, the Aora network can provide near real-time price feeds for a broader set of asset categories.

💡DeFi Agents

DeFi agents in the video are AI-driven entities that can execute complex financial operations within decentralized finance platforms. The video suggests that these agents can leverage the collective intelligence of the Aora network to perform advanced trading strategies, manage liquidity, and adjust parameters in lending and perpetual systems. This represents a potential new wave of innovation in decentralized finance, where AI can automate and optimize various financial processes.

Highlights

Introduction to Aora, a self-improving decentralized AI Network

AI stack simplified into four core layers: Data, Compute, Intelligence, and Execution

Blockchains as powerful coordination mechanisms for decentralized networks

Modularity in blockchain stack leading to rapid innovation in DeFi

Aora's focus on the Intelligence layer of the AI stack

Addressing the problem of siloed machine intelligence with decentralized networks

Aora's network allows different AI models to learn from each other, enhancing collective intelligence

The output of Aora is an aggregate model output that outperforms individual models

Aora as an abstraction layer for intelligence, similar to how Bitcoin is for power

Background on the company's evolution from Upshot to Aora Labs

Research in mechanism design and Pure prediction leveraging information theory

Shift from human inputs to AI in crypto infrastructure due to inefficiency

Aora's inference synthesis mechanism for combining different model outputs

The role of inference workers and forecasting workers in the Aora network

How Aora maintains privacy by only sharing inferences, not model weights

Simulation results showing Aora's network outperforming individual models

The importance of context awareness in Aora's forecasting workers

Transition from model-centric to objective-centric paradigm in AI interaction

Core properties of Aora: Collective intelligence, iterative learning, contextual awareness, and privacy protection

Initial focus areas for Aora in financial applications and crypto

AI-powered data feeds for longtail assets in decentralized finance

AI agents for executing complex strategies in DeFi

Risk modeling in decentralized finance with Aora's AI network

Upcoming test net phases and main net launch timeline

Transcripts

[Music]

hey guys I'm Nick co-founder of aora and

we're building a self-improving

decentralized AI Network um before we

get into things I just want to level set

a line on sort of uh endtoend view of

the AI stack as how as how we see it and

we kind of see it as these four core

layers this is somewhat of a

simplification but fundamentally the the

decentralized AI stack the AI stack in

general can be broken up into Data

compute intelligence and execution and

we think that like with other

decentralized networks this end to-end

system will be realized via a kind of

modular construction I think blockchains

are are incredibly powerful coordination

mechanisms for turning potentially

nebulous resources into more tangible

digital Commodities and when you're

coordinating disparate networks of

individuals around different actions

it's already a complex undertaking on

itself and when you try to build

monolithically that complexity compounds

and I think that's why we've seen the

blockchain stack modularize I think

that's why we've seen such a rapid rate

of innovation and progress in things

like defi it's because of this

modularity and the the ability to build

uh very tightly scoped highly composable

building blocks that get stitched

together and so when we think of allora

we think of the kind of intelligence

layer of the stack we're building and

kind of turning intelligence itself into

a digital

commodity um and the problem we're

trying to solve is is this kind of

fundamental problem in AI of siloed

machine intelligence there's many

different models in the world and

they're not very kind of composable or

compatible with one another I can't

merge Model A and model B to create a a

mega model that represents the

intelligence of both and this is a real

problem because it means that these

large monoliths with access to the most

kind of raw resources of data and

compute will continually making the the

better models and so what we're trying

to solve with aura is through

decentralized networks a way for models

to kind of learn off of one another and

become more composable or compatible in

a decentralized setting and so that's

what allur is allur is a a network where

many different models can learn off of

one another in collectively optimizing

different ml objectives and the output

of the network is a kind of aggregate

inference or an aggregate model output

that consistently outperforms any of the

individual models within that Network

and in doing so you've created a kind of

commoditized version of IG itself and

significantly accelerated the rate of

innovation of machine intelligence

itself um and another way to think of

allore is kind of this abstraction layer

for intelligence the same way Bitcoin is

kind of an abstraction layer for power

or other networks are abstraction layers

for trust or compute or things like this

allora is this abstraction layer for

intelligence allowing people to access

intelligence and the kind of best

aggregate form of intelligence through

the network itself as opposed to

interacting with individual models and

needing to go through the processes of

of deciding between different models and

different domains or

contexts um and just some quick

background on who we are so when we

first started the company it was

originally called upshot we're now

called a Lowa Labs we're building a

laowa network um we were doing a bunch

of research in this really new field of

mechanism design called Pure prediction

that leverages information theory in a

pretty interesting way to incentivize

people to be honest in the face of

subjectivity so enabling people to reach

consensus on the answers to subjective

questions that's important for a bunch

of different problem spaces the Oracle

problem more generally and it's also

useful for Designing decentralized

networks around especially nebulous

resources like intelligence and then as

we were building this out we realized

that humans as inputs to these systems

were too inefficient they're they're too

inaccurate uh too incompetent I guess to

support many real world tasks and that's

when we shifted to building some of the

earlier AI ex crypto infrastructure

especially in longtail Market setting so

uh AI powered price feeds for longtail

assets AI derivative infrastructure Etc

and what we're doing with allur now is

taking that early work we did around

subjective consensus mechanism design

and all of the learnings and

infrastructure we've built in the AI

crypto space over the past three years

to build this kind of collective

intelligence Network or self-improving

decentralized AI

Network and this is kind of how the

network works uh at a high level um the

the kind of core mechanism of the

network is what we call the inference

synthesis mechanism and it's a way of

stitching together different model

outputs in a productive way to create a

more performant or more intelligent

Network level output and so the the

general flow is consumers ask a a given

topic a topic is kind of like a sub

Network in the system it's a very

tightly scoped uh ml objective that

model creators uh kind of coordinate

around and are attempting to

collectively optimize a consumer would

would ping one of these topics asking

for some AI output and two flows are

happening one these kind of inference

workers are responding to that request

with inference that is being assessed or

weighted based on uh how well it

optimizes the core objective function in

that topic and simultaneously

forecasting workers are predicting the

loss that's going to be realized on

those inference workers uh as assessed

Again by that objective function and

they're they're over time learning the

context the different domains Etc in

which different inference workers within

a given topic perform well or perform

less well um to produce the the most

accurate kind of predicted loss and so

as these things get stitched together

you uh establish weights across the

different inferences developed by

different workers based on uh kind of

approximation of a shly value across

different models and produce an output

that over time increasingly outperforms

the individual models performance within

that topic and then there's a a kind of

second or third type of actor called

reputer who are tasked with actually

evaluating the performance of different

models over time so they they take some

ground truth they take some kind of

basis or validation set and assess the

performance of the different uh uh

workers as as assessed by or as

calculated by that objective

function um and and I I guess one one

thing to note here I I haven't looked at

this presentation in a while um is

because workers are just sharing

inference through the network the

network can support open and close

Source models they're not revealing the

weights of their models they're not

revealing any kind of proprietary

information of those models and we think

that's important for supporting a wide

number of use cases and we'll get into

it later but especially kind of use

cases in financial settings where

maintaining that proprietary alpha or

the proprietary information of your

models is especially

important um and the key to this

mechanism working is really around

context awareness we've done a bunch of

work around this but it's that

introduction of the forecasting worker

that allows you to realize a lot of the

outperformance at the network level

relative to the individual workers in

the network and so we can see some

simulation results here uh in five

predictors they're they're inference

workers but they're working in a Time

series prediction task um the best uh

predictor is predictor one and they've

achieved a log loss of uh about negative

3.3 uh so that's pretty good they've

they've uh if we didn't have forcasting

workers we would waight them the most

heavily as this kind of base set of

predictors and then produce an inference

that achieves close to that performance

but it's over time that these

forecasting workers sort of learn the

different contexts in which different

predictors perform well in this time

series prediction case maybe some models

are especially competent in say volatile

market conditions other uh models are

are more competent less volatile market

conditions and these forecasters would

learn that over time and uh adjust their

their predicted losses to uh uh kind of

represent that and what we see in this

this bottom figure is that with the

introduction of aggregators or

forecasters as as they're called the the

best forecaster achieves a a loss of

around -4.4 and so that's around 33%

better than even the best model in the

network because they're learning the

context in which different workers

outperform other workers in the network

and additionally we can see that even

the worst aggregator the one who is

taking kind of the uh who is waiting uh

contextual awareness uh least

sensitively is still outperforming even

the best inference worker within that

Network and so it's this this kind of

element of context awareness this

additional dimension of performance

where we really see outperformance at

the network level in this kind of

setting and then another nice property

of this is when we interact with with AI

today it's a very kind of model Centric

Paradigm I'm saying I'm going to use gp4

I'm going to use CLA 3 for some task and

I'm asserting that that model is going

to be the best in every domain and every

context that I'm using it in and that's

not the case I don't think there's ever

going to be a single model that is

dominant across all domains

and when you when you align things

around given objectives or given topics

you start to move to this objective

Centric Paradigm of interacting with AI

in that I care about this ml objective

this evaluation criteria that matters

for my application or for my use and

then this underlying set of Ever

Changing models is being productively

coordinated and stitched together to

produce the best possible output is that

you can think of this as kind of

analogous to the move from transaction

too intense of moving from specifying a

specific set of uh State transitions to

achieve some objective to just

specifying the intent that you care

about at the end of the day and as we

move towards an objective Centric

Paradigm of interacting with AI a

similar thing will unfold is in instead

of interacting with specific models

we'll interact with specific ml

objectives that we ultimately care about

and ultimately satisfy our needs either

in use and application development

whatever it

is um and so just to recap these are

sort of like the four core Properties or

features of allora 1 collective

intelligence I think that's pretty

simple we're bringing together many

different heterogeneous models to

produce a more intelligent kind of

network level intelligence and because

of forecasters the model is or the

network is kind of learning over time

and so you achieve this property of kind

of iterative learning at the network

level Additionally you achieve iterative

learning at the model level as well

because each of the individual models in

the network are learning the the weights

that are being assigned across other

models in the network and using that to

to improve or inform their future

outputs uh third contextual awareness as

we talked about is kind of the key to

stitching together heterogeneous

instances of machine intelligence in a

really productive way in a way that

outperforms any of the individual models

within that network uh and then lastly

as we talked about because the workers

are only sharing inference within the

network there's a level of privacy

protection that supports this kind of

closed Source uh model participation

within the network

uh and then I just quickly want to go

over some initial Focus areas so I think

as we think about where this network is

most applicable at the application Level

we see a lot of financial applications

as the initial kind of uh like verticals

to tackle I think as we're talking about

crypto and AI crypto is largely a

financial innovation in my opinion it's

it's found most if not all of its

product Market fit across different

Financial settings and AI in financial

settings is also much more mature than

then some of the newer types of

generative AI as well and so I think a

lot of the immediate value within this

intersection of crypto and AI will

happen in various Financial uh settings

and verticals so we'll quickly go over

some of

these one AI powered data feed so this

is this is pretty broad but I'll Focus

specifically on kind of AI powered price

feeds for longtail assets this is what

we've been doing historically at upshot

for the past few years um but when we

start to move into the long taale of

assets where crypto like in defi edly

really shine in in codifying certain

Market interactions to enable the

support of more longtail asset

categories we still lack a a kind of

robust and accurate source of price

Discovery for these assets uh they don't

change hands very often and so these

transfers of ownership that act as price

Discovery events for other asset classes

often don't happen enough uh for that to

be the sole source of price Discovery

for many different assets and AI can be

used to take into account many other

pieces of data outside of the market

itself as well as upsampling Market data

in productive ways to produce re near

realtime price feeds for a much broader

set of asset categories and sort of

equipped with a a price feed you can

start to build a bunch of really

interesting Primitives in in

decentralized finance lending protocols

Perpetual systems synthetic

representations of assets your your kind

of scope of compatible assets in

financial settings expand significantly

with with AI as a a kind of supplement

to that price Discovery

mechanism uh second this is again pretty

broad uh but AI agents for defi I think

today defi has showed us uh the power of

decentralized composability the power of

kind of codified financial operations

but it's still very primitive and and uh

I guess unexpressive and with AI we can

start to leverage this highly

deflationary form of compute that can

take into account many different pieces

of information in an incredibly

efficient and effective way to start

executing more complex or Advanced uh

trading strategies liquidity

provisioning strategies manage uh

different dii parameters and lending

systems or Perpetual systems in a far

more kind of advanced or efficient way

and so I think AI inter intersecting

with defi via the form factor of Agents

built on top of this kind of collective

intelligence network will represent the

sort of next explosion of innovation uh

and sort of uh experimentation within

decentralized Finance uh and then lastly

this this is risk modeling uh especially

as reaking becomes a like a larger part

of the space represents a larger

percentage of the space's capital the

complexity of risk when you're reapo a

like pool of capital across many

different heterogeneous networks becomes

incredibly complex I think we saw that

some of this play out uh earlier today

or last night um and AI in in this kind

of decentralized form factor can can

offer us a really robust tool for making

sense of all of these really complex

risks and develop kind of adequate

hedging vehicles or or other kind of uh

Financial Primitives to help protect

against that risk and as such protect

against this this mass of new networks

that are are emerging on top of uh a

kind of source of restak

capital um and just quickly launch

timeline so we're in test net phase one

right now we're about to launch test net

Phase 2 in the next week and a half main

net shortly after and in the meantime

we're just uh onboarding workers model

creators that's really our top Focus

right now is bringing together a kind of

critical mass of of machine intelligence

onto the network network gets smarter

the more models there are and so that's

that's really our top priority and we're

also working with a bunch of kind of top

tier applications and different networks

in the space and bringing this source of

collective intelligence to different

kind of parts of the crypto space so I

think that's it

[Music]