Fluence Labs Conference talks, 11 Mar 2023

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: AI's need for Decentralized Compute by Tom Trowbridge

Description

AI's need for Decentralized Compute by Tom Trowbridge

ChatGPT has taken the technology world by storm and set a record for the fastest product to cross one million users. While the ‘death of Google’ or ‘death of search’ is certainly overstated..for now.. the ability for AI to answer questions quickly, thoughtfully, thoroughly and correctly is remarkable. The next version will undoubtedly be even more powerful and reach tens of millions on its way to billions. But with this power and reach will come a freight train of scrutiny and controversy.
As these AI engines mature, they will be seen as a source of truth and every society and government has a view as to what ‘truth’ is regarding certain topics. The answers to delicate questions are already drawing attention; ChatGPT can the benefits of green energy but some users have found it hard to get answers for ‘the benefits of carbon based energy’. What would it have answered about masking in 2020 when Fauci didn’t recommend it? Would that answer be the same in late 2020 when the CDC strongly recommended masking for everyone? The sensitivities get even greater in countries with even more charged topics such as in India or Pakistan not to mention with the totalitarian regimes that enforce a particular narrative. We have already seen the US government’s interest in supporting or limiting certain topics on Twitter and a single source of truth will be even more threatening to governments, politicians and even communities.
How can a company or organization possibly manage all of the competing demands regarding particular answers? We can expect AI engines to be banned in some countries, but where it isn’t banned, it certainly will be the subject of regulation which will erode its credibility and likely will reduce the scope of topics it can address.
The first step in reducing company or government influence is an open source AI engine where a community can train the model transparently, and several projects like Bloom and EleutherAI are already in progress. But training is just the beginning; no matter the inputs, scrutiny will be on the outputs and any organization running the model will be subject to pressure. The only inoculation from this pressure is to decentralize the actual running of the AI engine by hosting it on a global decentralized network with no owners and no single points of control. Run on a decentralized network, any storage provider can host the model and any compute provider can process queries.
The good news is the technology already exists to make this a reality but how will it be implemented and by whom? And will an uncensored version be good? What are the drawbacks?
The future will likely see multiple AI engines competing to be the source of truth, and that is healthy. Linux provides a good example of how the global community can build a powerful alternative to centralized companies and the same may happen with AI but as important as building/training the AI.. is the running of it.

Tom Trowbridge
co-founder
Flunece Labs
Tom Trowbridge is a co-founder of Fluence Labs and Hedera Hashgraph where he was president from inception, is a board member of Stronghold Digital Mining (NASDAQ: SDIG), an environmentally beneficial bitcoin miner and is invested in a number of leading Web3 projects. Fluence has developed and launched a distributed serverless compute network that frees computation from centralized cloud providers. It functions as a decentralized application platform that allows developers to host applications and work together, reusing both components and data without fear of being cut off or de-platformed. Tom started his career financing telecom and technology companies at Bear, Stearns & Co. and then began investing in early-stage technology companies at the private equity and venture capital firm Alta Communications. He spent four years at Goldman Sachs and left to build businesses at several other financial firms. Tom has a BA from Yale University and an MBA from Columbia University

A

Well, thank you um for the introduction and um protocol labs and filecoin Foundation have been big supporters of fluence labs, and we appreciate that um you know you can't have decentralized compute without decentralized storage. um Both you need both of them and we all need both of those for this ecosystem to be truly decentralized. And so, while we wait for my slides, I think the first one is going to be pretty obvious, but it's about chat GPT and about the growth of chat, GPT and I. Think everybody here has seen that the stats are interesting.

A

It's a million users in like a week and 100 million users in a month, and this is, without a doubt, the most the fastest growth of any consumer related application in the in history, and that is catalyzing a tremendous amount of interest in Ai. And if we go beyond that, you know you think about. What's it used for there's some terrifically fun uses for it and we've seen people you know, I saw John Oliver did a comedy routine where he had a wrap written about cats in a style of Eminem So.

A

That is something you can do with it, but there is a terrific number of commercial uses as well and on this slide you can't see I have a list of those uses which go from Health Care to finance, to accounting, education, media. You know, law, video, editing, developers, code, both writing debugging, a terrific amount of potential uses, but the problem is that old engine we're just not getting on.

A

A

A

A

It's it's right here: the PowerPoint hooray we're up. Okay, so hmm I'm, not gonna. I'll. Just give this quick, quick review. So here we are. This just is a graphic depiction of the scale of Chachi PT's growth, which I think a Picture Tells a thousand words. So you know 100 million users in pretty much zero time scale, given everything else and catching up here, commercial uses besides the fun uses, are remarkable.

A

All right, I have a request to go back for a picture on this slide there. We go we're good um up too slow next time, um and um these are the commercial uses, but the problem with the commercial uses is we have these things called regulations and certifications, and this is just a snapshot of California as an example.

A

These are all the professions for which regular certification is required, which means a study, a test and a degree, and a lot of these have to do with physical activities that relate to people and have a health component to it. But really what these are. Are protectionist measures put in place by Industries to protect the um earnings and make it harder for people to actually enter those Industries? And these go back? You know hundreds of years to guilds right.

A

This is nothing new, but you had very protectionist interests of every particular industry, particularly not just unions, but also white collar, and so it shouldn't be a surprise that one of the first industry reactions to chat GPT entering it is the legal profession where um Bill browder's son was going to go to a court case and have a court case against a traffic ticket argued via Chachi PT prompt in the ear and the bar association found out about this because he publicized it and they threatened him with six months in jail if he went ahead and did that right- and this is clearly not who is harmed by have fighting a traffic ticket with with AI prompts in your ear right.

A

The only people harmed is the legal profession and they are very scared of this. So came down very hard. This is the first example of an industry really recoiling against and fighting against, Ai and I expect. We will see much much more of it. um The other risk, with a chat GPT in particular, is these are centralized. For-Profit companies that have and Google is as well, obviously for which, if you try to build a business on it, you are subject to their pricing, their API and anything you build on.

A

They can decide to cut you off. Obviously, we saw this back and for people remember the Farmville Facebook, but we've also seen it happen on Twitter third party applications, Twitter close those down as well LinkedIn did as well. So whenever you have a valuable platform that has power, you try to build on it, you're a real risk of being cut off, so you need an open source alternative that you are safe to build on. Otherwise your business can be taken out from under you at any moment.

A

Next risk is what happens when chat GPT scales to many hundreds of millions or billions of users. At that point, governments start to become interested. You know, what's the answer should I wear a mask? Should my kids be vaccinated is? Was there Armenian Genocide, you know is ex-government um corrupt right. These are answers that don't that certain governments first may not like the answer.

A

Second of all, different governments will have different responses that they want, and you know in Europe as just an example: vaccines aren't offered to kids under 18 in a whole bunch of countries U.S, you get them almost mandatory from six months right. So how is this going to play out when you have a company that is creating answers to these questions that are being asked globally, with very different government interests in what those answers are?

A

Well, we have a we can look to Twitter and we now have every day more and more evidence of different countries exerting different influence on Twitter, and you can expect that to come from totalitarian companies like China, like Russia, that is expected, but democracies India in particular, has had a very heavy hand with Twitter, silencing and banning opposition figures and for and not allowing um protests, uh videos and protest information to be shared a couple months ago and they got very aggressive, threatening jail for it also happened in the U.S.

A

Obviously, and let's not forget that you know you were banned for talking about a lab leak theory that now is relatively um you know, mainstream so I think you see now multiple reasons and multiple rationale for having an open, transparent and free learning language model, and your first response might be well. Listen. Microsoft spent 10 billion dollars in this.

A

How on Earth can open source compete, and the answer is first Bloom is already out there um starting this effort, but Linux in 2008 estimated 10 billion dollars worth of developer, resources were deployed to build Linux and, if you've been starting to build Linux early on no one or very few people would have given you a real chance of survival versus OS versus Microsoft operating systems, and they did it and are dominant now and so I think what will happen is not a one learning language model to compete with googler AI, but it will start much more specifically in narrow.

A

You could have a dow raise, for example, to raise to focus perhaps just a model. That's focused on pediatric medicine, diagnosis right, super, smaller someone gave him small and said you know what I don't think minority children are particularly appropriately represented in Broad llm, so I want to have a Dao. That's just focused on the research around minority children, health issues as an example that is a type and then so it's not clouded by data. That's separate, or that is distinct from that particular group.

A

And so you could see those type of models start be very specialized, much easier to fund and to build and also be open and transparent, safe to build on once you have that, though, you end up with a whole Suite of specialized open source models.

A

The front end becomes very important at that point, because you want to have one front end that you can access and use all these models with and the front end is also subject to potential censorship and vulnerability, and we've seen you know, you know Bing didn't like some of the answers that its model was given after it was giving after a variety of um prompts. So it's limited to front end to only ask six questions right.

A

So this front end is also as important as the training of the data and as the training of the model and so distributed. Serverless compute enters here, because that allows you to run this model anywhere and make it make it run the front end and then run the model anywhere to make it Unstoppable.

A

The other part is, you can run the model and have proofs that the model was actually run and you weren't just given an answer that was already generated for somebody else for even a different question and again coming back to health care or law.

A

You may care a lot that the model was actually run and actually generated the question that your prompt um requested and that that model was a certified model run by the particular Dow that has been um focused on that particular issue and was credible in it and so what's required to run this type of distributed, chat, GPT or learning language model, and you need distributed. Serverless architecture is what we think is kind of the critical piece here.

A

The main point of it is a fault, tolerant design and so you're not you're, not dependent on any network or connectivity provider or hardware, and um if a host goes down, the application continues to function by traveling to other hosts and distinction. Here, what doesn't work or what's not ideal, because when you think of decentralized compute, there's a couple other aspects to it: compute at the edge and Hardware marketplaces.

A

Both of those are terrific solutions that are very usable and viable, but they don't work so well here, mainly because they don't have failover, at least in the ones that I've seen, and so it's computed at the edge or Hardware marketplaces work, but they're still isolated to the particular location. They're happening so are subject to both um both both censorship risk and can easily go and much more easily go down, and so how does fluence enter into all this?

A

Well, it's a distributed, serverless platform, that's fault, tolerant by default um and has- and it obviously has failover and it has a security model that verifies that computations have been completed. That's a critical piece that, as you ask of any distributed computation, it also has a native language Aqua that makes building these front ends far easier because it abstracts out all the div all the complexity of peer-to-peer applications and distributed applications which is very difficult to build from scratch and the marketplace where you can buy and buy. This compute helps Drive price down.

A

So, but if it's good for learning language models, this distributed compute actually has far broader applications um and broader use to Applications generally and so. Applications can live on the distributed, Global Network, with no single point of failure and no single point of censorship in a very price competitive environment, and so it can scale up or down on demand and be free to move from any provider. And this gets to the cloud lock-in you avoid Cloud lock-in, and so, if you're running a serverless compute, your application is on that.

A

Someone may be running instance at Google, someone maybe, and then you decide, I want to move to Amazon. You can move it to an Amazon server. If someone is hosting there to your own server to somebody else and that level of.

A

And that and that level of that ability to move um that ability to move from provider to provider allows you allows both price competition to be terrific, and it also allows you to be free of the cloud ecosystems and you know why do we not like Cloud ecosystems? This is an example of parlor, but the exam I don't think anyone's lives here has been dramatically impacted by the shutdown of parlor, but it just gives an example of the power of clouds to turn off your application.

A

If they decide you violate their term to service, and this was a big application right, millions of users and it was shut down with like a day or so of notice. Why is that relevant for us? Well, guess what web3 is really dependent on Amazon and I?

A

Don't know if anyone's heard of Gary Gensler and um the SEC, but but there is a risk here that we are all bearing um all these applications are using the cloud and until we free ourselves from the cloud we are at risk and the cloud notice is not some white puffy benign entity. It's a storm cloud because it's dangerous and we should all be very wary of it. Not only is it expensive, but it is a risk we all bear and there's also centralized backends to a lot of dabs, and this again isn't anybody's fault.

A

It's just the infrastructure hasn't been built yet to allow these applications and these products to move off of the cloud and that's what we are trying to do and we think that decentralized, serverless compute is the next evolution of compute. So you started with mainframes.

A

People were dragged Kicking and Screaming into the cloud which is clearly won so far and has you know been an enormous success, but decentralized serverless compute will be faster, it'll be more secure um and it'll be cheaper, and so you get on-demand scalability higher security, competitive pricing, it's more censorship, resistant as fault, tolerant- and this is you know this is what basically we have to go towards, or else we are subject to the oligopoly of Google and Amazon, which I don't think anyone in this ecosystem wants.

A

So compute is kind of the key core piece of that, and so the fluence decentralized platform has two pieces to it: a developer platform and a Computing Marketplace and the developer platform is designed to mimic the cloud stack, and so we have again no slides um which I just noticed, but um it mimics the cloud stack with a AWS kind of Lambda, analogy the step function and rest API and aqua and then kind of fluency CLI which which obviously enhances building so there's.

A

A variety of the idea is to make the journey building very similar to Cloud analogy, and if you don't as much as we don't like centralization, people will still use the cloud. So we have to make it as easy and simple as we can in order to do that and that's what we're working on doing at fluence.

A

This was actually really relevant to see which we're not but um Cloud native, basically accesses Services through an API. You cut off that API and you lose access to the micro services with fluence. You access the micro Services directly through marine, and so as long as you have access to the internet. Your compute travels to each of the micro Services. You need to execute in order for your application to run, and that means that there's no API to cut off and there's no one place that you can be can be shut down.

A

Verifiability there's two parts of this proof of processing. You want to make sure that the aqua execution and that's verified by every host along the way and then also proof of execution where you have ZK proofs. Look at this for marine, um and so we got. Luckily we can visualize the verifiability, um but this was relevant. I just would love to be able to show that which now I can't um it's gone.

A

It only likes that slide um um and then and then the final piece is a compute Marketplace, which is um you know, global scale. When you have a Marketplace, you are obviously driving pricing to um the most competitive value that you can find, and you know Amazon and Google have 60 margins. Those margins aren't coming from nowhere they're coming from everyone. That's either using the services built on the cloud or they're coming from the people building on the cloud.

A

That is a lot of value to be distributed out to the ecosystem and when you have a compute Marketplace, that value will accrue to people building and to people and then probably goes to people using the applications as well. So the goal here is to actually compress that margin and put it and allow us to build and use it um to to better better, um better our applications and better the Universe versus Amazon's bottom line. So um fluence is moving forward. We've we've got the market. We've we've developed Aqua, which is the core language.

A

um The we've done developer workshops, while we're here I encourage you to check it out. We have several um a handful of bounties that people are pursuing. The next step for us is launching it down, which is where our governance will take place.

A

It's going to be a Swiss Association and the Dow will then um likely issue a coin which will be related to the use of the network, which is important and that's part of the compute economy, which will allow um hosts to be paid for compute and then also users to actually make deals with compute providers and run their applications in a truly decentralized way. So um you know AI helps us maximize our potential, but it needs to be free and compute needs to be free.

A

Centralized compute will will kind of runs the risk of wrecking this whole eco system. We're a part of so decentralized serverless compute is what we're we're building here. So please follow us at fluence underscore project, and here is the websites fluents.network, and you can find us at discordfluence.chat and twitterfluence underscore project, so sorry for the technical difficulties. But thank you guys for coming and listening.