Filecoin Bacalhau - Compute over Data, 20 Aug 2023

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Distributed Compute for Generative AI with Lilypad | FIL Paris 2023

Description

Verifiable, truly internet-scale distributed compute networks are necessary for the next stage of web3 development. A lot of discussion has gone into friendlier UX & scaling blockchain over the past year, culminating in account abstraction & ZK proofs. However, the next frontier for computing is really at the data layer - to build truly decentralised applications that can run at internet scale requires more efficient off-chain computation (that isn’t a centralised backend!). This opens up the door to so many opportunities unavailable in the web2 world including ML DAOs, decentralised AI, DeSci & applications that take advantage of both the features of blockchain and the the power of distributed compute (eg. Waterlily.ai)

Speaker: Ally Haire

#DecentralizedStorage #filecoin #filecoinfoundation #web3 #blockchain #DWeb #cryptocurrency #crypto #decentralization #ai #generativeart #nft #collab #opensource #web3 #stablediffusion #generativeai #artificialintelligence #art #Filecoin #Lilypad #DistributedComputing #GenerativeAI #Web3 #DecentralizedInternet #Blockchain #AI #MachineLearning #IPFS #OpenSourceAI #CryptoEconomics #PeerToPeer #FILParis2023

A

uh Hello, everyone and I think this is a first for me. I, don't think I've ever done a presentation on a boat before so good to see. Also I'm gonna put it out there Robert that we should uh name the French file, Queen Corgi profiterol. You heard it here first anyway, uh good to meet you all.

A

I am Ali I'm, a developer relations engineer uh with the filecoin foundation and today, I'm really excited to share a project that is going to be I, think Super, valuable and super impactful, but not just the file coin ecosystem, but it definitely will be, but also uh for the wider web 3 community, and that is the lily pad project or Well Network. Now and yes, that is a hint um I. I just want to give some credit to the team as well: I, don't work alone, um I want to credit.

A

uh Luke and Kai are amazing, Engineers on this project, Levy, how our game, theorist and researcher and David for his uh amazing leadership and expertise here um so the agenda. Firstly, I just want to like backtrack a little bit recall the file coin vision and why this fits into that nextly. I'll. Tell you what Lily fat actually is and then I want to give you a little bit of a demo and uh tell you what lily pad empowers and why it's important uh so to set the scene, the file coin vision.

A

uh So, as you're all likely aware here at filecoin and protocol Labs, we have a mission and that mission is to build the foundation for a decentralized internet and that uses four key components here. So uh so far, we've built out the largest decentralized storage Network globally in filecoin, and that is not a uh easy feat. So that is already a great achievement. We've then laid it on uh with powerful programmability, with the fbm who you heard about earlier.

A

uh The filecoin Saturn, retrieval Market, is also really well underway uh to allow for fast but move faster movement of data across the globe and even planets in the future. Who knows, and the fourth pillar of this decentralized internet is really computation. So to me this is where data becomes truly useful.

A

um So, after all, an llm is not is nothing without the ability to kind of compute over and train on data sets, and this is one thing that we're hoping to achieve. So it's really my strong View, uh the distributed computing will unlock this next generation of truly like web scale, decentralized applications and bring the power of this robust uh compute to filecoin and to web3. So this fourth pillar is where balkia uh some of you may have heard of, and Lily Pad live. So just in case balkia isn't familiar to. You.

A

Bachios appear to peer open network of nodes that do computation and it kind of enables you to run arbitrary, Docker containers or wasm images as tasks against data and to analyze. Large data sets, while bringing this benefits of open collaboration on data sets stored in both ipfs and filecoin to generic compute tasks pretty powerful already.

A

um So, if bakuya already does peer-to-peer distributed compute, what is lilypad well lilypad kind of leverages the power of bakwi out and combines it with on-chain guarantees and crypto economic incentive models. So much like filecoin is to ipfs. Lilypad is tobacco as a simple kind of analogy: it's not quite true, but pretty close, so it's bringing incentives and guarantees to the network.

A

uh So for those of you that may have been following lilypad previously, uh you might remember the original proof of concept project we built out that operated as a relayer for calling back here, off-chain compute jobs and receiving the results directly in your smart contract, and that was deployed to the fbm and some other networks as well. So this was really a precursor to the on-chain lily pad Network or lily pad V1, which leverages balkia and key features of blockchain, for secure scheduling, for verification and for payments both for jobs and the network nodes.

A

So with Lily, Pad V1 we're definitely aiming bigger. We want to build a trustless distributed compute Network that enables internet scale, data processing, including AI, ML and other arbitrary computation, and allow that to be callable from Smart contracts, so with blockchain guarantees.

A

We also wanted to make sure that we can unleash this idle processing power and unlock like a new wave of data and compute economies in web 3..

A

um So this is a decentralized compute Network, where jobs can be called from a smart contract and where the jobs can consume import data. So it's really unlocking the next generation of these decentralized applications.

A

uh So today, I'm really excited to tell you as well that the team has definitely just released the lily pad V1 test net. uh So initially we have uh released this test net uh called lalochuza uh on Geth, uh but the idea is that actually we'll be building on IPC, which is a scaling Network for filecoin, and this will be our path to network our two mainnet.

A

But we wanted to allow Builders to be able to play around with this network, uh contribute modules and add to it in the meantime, so we've built it on our Gap Network, just so, people can play around with it while we build it out further all right. This is the architecture for Lily Pad V1. At the moment, uh we based it Loosely on uh This research paper, which uh enables modicum, which is kind of adding um mechanisms for scheduling jobs, uh mediating, those jobs, making sure people do the right thing on the network.

A

Basically, but we've replaced some of the components with Baku yeah, because balkia is already a super um super powerful distributor, compute Network, all right, I'm, going to show you the code. Now here's some of the things you can already do with the lily pad Network.

A

uh So, firstly, we've decided that cow say should be the hello world of Lily Pad, just for no reason other than we liked it. uh So this is what you would uh do in the CLE. You would just go Lily Pad, run cow, say hello, Lily Pad, the output would be an ipfs CID and there's the uh image there. Now you can do the same thing in a smart contract, but here's the clear version of stable diffusion, so stable diffusion is a text to image generator.

A

You can also run this in the clay or you can run it on a smart contract which we also have a demo of online I. Think I'm pretty low on time, though so I won't show you this, but if you go to our YouTube, which is our lilypad Network on YouTube, you can go and have a look at how you would call stable diffusion from a smart contract as well right.

A

The next thing that we're going to have implemented- and it will be implemented by tomorrow, according to the engineers, so pretty exciting- is uh being able to fine tune a model on with Laura on our Network. So this is kind of really cool, because not only are you adding input data? Not only are you getting data back? Sorry, you're, not just running an inference AI but you're, adding input data to this. So you can get out images like this from Laura, which is pretty cool and you can run those from ipfs.

A

So lilypad also uh recently participated at augment hack, which is a decentralized AI hackathon here in Paris that actually finished yesterday, and it was really cool because this guy Leo actually contributed a new module to Lily Pad an llm inference module. So you can also run llms on lilypad alrighty.

A

uh One of our other use cases that is coming up as well is filecoin data prep, so the Falcon data prep module is going to allow users who are onboarding. Data sets to file coin to outsources, CPU intensive work of chunking the data into car files and generating the metadata files to compute nodes on lilypad. So these nodes will read the data from any S3 compatible endpoint and make the resulting data available on ipfs or soon via high performance direct HTTP download.

A

So this will be for use, be able to be immediately used in filecoin Deal making and hopefully we'll make the process more efficient as well and, of course, because this is a network of nodes, you can run your own node on the lilypad network. Now, you'll only get paid in test at the moment, so it's probably not worth it, but hey it's fun.

A

uh So why does all this matter uh not just a file coin, but for the white, wider web 3 world uh I? Think I'm gonna have to be quick here, so, firstly, the capability for decentralized AI, with distributed networks, is one of our major wise, so in the same way that one person shouldn't control, access to social media or one company should not control access to your data or to Scientific breakthroughs. uh Central entities should not control access to AI advancements or to the data involved in them. So ai's had an immense.

A

Has this immense potential to provide benefit to humanity and common human Goods shouldn't really be gate kept, so in fact, we should probably all be stakeholders in its future and I'd recommend.

A

If you want to hear like more about why I think that check out this talk by uh Peter Wang, who describes this, why much better than I will so what's also interesting to me is that, according to this loot, leaked sorry Google article, it's also clear that open source, Ai and ml is already out competing the centralized versions, because they are able to Define use cases and fine-tune more easily on less data, so they're also able to Target where this common good lies better.

A

So one of the issues, though, for open source AI, is access to gpus. So this is a key problem for open source, Ai and ml to be competitive and dpin. A decentralized physical infrastructure networks like filecoin, can really help solve this.

A

uh It's also markets like uh file. Sorry markets like Falcon and marketplaces like Lily Pad, which enables coordination layers for matching CPU or GPU with Supply, are usually generally more efficient and better at pricing, and also better at distribution.

A

So some capabilities I'm going to here's some of the ways we think it can help the filecoin compute applique unlock Falcon compute applications as well. For example, uh Zach earlier was talking about darted housing, the filecoin ecosystem. So we can now bring the ability for decentralized, AI and ml to those data Downs which could you know, be really powerful as well. We're also, as we mentioned, trying to make easier data, processing and preparation for file coin deals and we're.

A

uh You know trying to imagine enabling secondary income streams for SPS and idle GPU resources in the network as well uh and look AI is amazing, but it's definitely bringing its own challenges and we think blockchain uh can help, try and tackle some of those challenges as well. uh In fact, we made a proof of concept called waterlily.ai that tries to tackle the attribution problem uh so a lot of the time.

A

These artists, kind of works are just stolen used in data sets, and then people generate text to images text to image models and there's no attribution for the original creators so with water lily. We're trying to match up uh giving attribution back to these artists by paying you know a small by paying them all fees for everyone that uh sorry runs a stable diffusion image on water lily next for water lily, uh we're gonna keep building.

A

Obviously uh these are some of our goals uh and our future roadmap, as I mentioned before, includes releasing IPC test net and, following on from that, uh we're going to, you know, obviously get to mainnet at some point: So yeah! Thank you! That's that's all from me. If you want to get in touch with us, uh these are all the socials so feel free to reach out, but otherwise thank you and looking forward to one's talk too.