The Future of AI Is Decentralized - Alex Skidanov - Near Day at ETHDenver 2023

Captions as YouTube publishes them, usually machine-generated, so product and people names can be misheard.

1. 0:00

   okay before we have the slides uh let's uh let's let's talk about artificial intelligence uh it's been it's been taking off recently uh and uh if you if you follow the history of near back in 2017 started as an artificial intelligence company uh moreover the name itself near was referring to the to the Advent of AGI it was the singularity That was supposed to be near and eventually I guess we did not build the

2. 0:30

   singularity we built something else uh something cool but uh I never give up on the dream of building AJ one day and um as of last December I started a new company which will be announced in due course with the sole goal of building artificial general intelligence and um yeah more more news will come out later this year at today's presentation once we have slides I want to talk about

3. 1:04

   the state of AI today how people build artificial intelligence and how blockchain is a very is a very fitting tool for many things that people do and it is not being used yet for those things uh but uh um it can be improved in many ways so so let's let's wait for the slides well actually yeah screw the slides um okay so if you wanted to build a large language model today like if you wanted to build something like chat GPT or stable diffusion it builds on top of

4. 1:42

   three pillars one pillar is uh uh is compute you would need something on the order of 1000 or more uh of uh of a very specific very expensive GPU from Nvidia called a100 you would need the second pillar is model architectures or research in general if you want to push the boundaries uh you need the best people um in the industry who will be building designing the model architectures and

5. 2:12

   pushing research for you and today research is entirely consolidated uh in the hands of very few entities and finally you will need data and you will need a lot of data and I only have 15 minutes so I will very briefly talk about two of them and I will talk a little more about one which I'm particularly excited about so compute if you think about compute and your first and crypto and your first thought is let's let's pull together compute from many

6. 2:46

   people uh and orchestrate access to it on chain you wouldn't be alone thinking that I think that's a very very natural application of blockchain technology to democratizing access to compute today however it doesn't work quite for a reason that for all the modern architectures that exist you need extremely fast interconnect between the models sorry between the machines so people literally have infinite band

7. 3:13

   between every between all the machines in their clusters and obviously if you put together compute for many many people that will not be the case you will have regular Network between them so before we can do that we need several more breakthroughs to actually make it possible to teach models over slow Network and there's quite a bit of research happening but we're not quite there yet to start pushing it from from the blockchain side when it comes to model architectures

8. 3:38

   today oh sorry actually there's something we can do before we go to model architectures and specifically I think one interesting idea that is worth exploring uh is to give up on the idea of having a decentralized compute cluster as of yet have a fully centralized cluster from one of the hyperscalers like AWS or Azure but build that Dao on top of the cluster that will be Distributing access to it to the

9. 4:03

   community right to imagine like you're a researcher in one of the smaller universities you don't have access to compute you go to the Dow you apply for a grant but the grant will not be in tokens or anything like that the grant will be in compute hours so I think that's a very interesting idea to explore and that would that would be a massive step forward in terms of in terms of democratizing access to to compute so when it comes to model architectures

10. 4:28

    it's also very interesting because today all the people all the best people who are pushing frontiers of research are concentrated uh in the number of entities which I can count on one hand uh primarily Google and open AI but it doesn't have to be the case in the sense that if you compare research in artificial intelligence to software development for example right like imagine you want to imagine you want to make a meaningful contribution to the reference client of near protocol before before you can have any meaningful

11. 4:59

    contribution it would take you several months of reading the code base understanding and doing smaller um work items and because of this already commitment that you and the company that you that you want to contribute and the project that you want to contribute to have to because of this contribution of time that needs to happen ahead of time usually if you want to be contributing to some project it would only make sense if you're ready to

12. 5:24

    commit over multiple years it is not the case with uh with research because with research if you want to be contributing to some project you can come and go like you can join the project at any point usually you can live at any point so there's no inherent commitment that you have to make to any of the of those entities the big reason why people have to go to those entities is access to resources so again if you

13. 5:48

    did have the Dow that has access to uh that overlooks a large cluster that could be a step towards democratizing research and pushing the the frontier you know like okay data data is the one I'm most excited about the reason for that is that when it comes to computer and research centralization is a problem and I think centralization is holding us back but it's it's the only problem so mostly in a centralized fashion

14. 6:19

    everything works when it comes to data annotations everything is broken for everybody involved whether centralized or decentralized and I just want to work you through today the the program of data annotation and how it can be solved and also what has been done to solve it today so generally the idea is you're an entity you're a company your person and you want to teach a model to have some new capability you want to have a machine learning model that uh given some input

15. 6:51

    given some tasks that a human can do would be able to do the same task and the way you would approach it is you would have some amount of data that some amount of inputs for the task that you want the model to be able to produce outputs on but because the model doesn't exist yet you will have humans who will annotate those outputs and then you will have input outputs pairs that you will fit into the model and hopefully the model will pick it up and

16. 7:16

    so the way it works is that you go to one of the data annotation platforms and you apply the supply the the input and the data annotation platform will have a set of workers who work there and each worker will get a few of those tasks and they will annotate them with the correct answer with the correct output that the model of the future will have to learn to produce all the solutions go

17. 7:40

    to the requester and at this point the requester has the solutions and they go through them all the all the outputs they go through them and they say yeah this output is good this output is good this output is not good and depending on their verdict the workers whose outputs were accepted will get paid the workers in outputs were not accepted and there's a plenty of centralized providers today

18. 8:02

    scale AI mechanical flower uh and in this setup it is broken for every single entity involved it is broken for the platforms it has broken for the requesters that is broken for the workers uh for the platforms it is hard because workers are usually distributed across all over the world and so as a platform you will have to figure out how to make payments you know somehow running out of time you know

19. 8:27

    just get on in a lot of geography something I don't change on the blockchain a user yeah payment just work um for requesters the problem generally is that the quality is very low so in Silicon Valley AI is taking off right now so like if you go outside and talk to a random person they are an AI founder and all of them annotate data and you can talk to them about their experience and the experiences consistently disastrously bad so all if you use scale if you use any of those

20. 8:58

    providers the quality is not good okay and finally for the workers it's the worst of them all because first of all there's absolutely no support if you do some work and you're supposed to be paid one and a half dollars for it support is not economically viable so the com the company like Mechanical Turk just cannot possibly have support right right you cannot you cannot justify

21. 9:22

    having people on salary doing support because the amount of money in question is just too little um secondly the requester who submits the work has the last say in whether you're getting getting paid or not right so it is actually a common place it's something that happens from Mechanical Turk a lot where the requester will come have the data set annotated and then we'll just Auto reject every single submission keeping the data and just not paying the workers because they can do

22. 9:48

    that and in the absence of support unless they consistently do that it just works and finally payments are delayed for multiple reasons one reason is that if you if you did some work and you were supposed to be paid two dollars it's just not economically viable for the providers to send you the two dollars it's too little of a payment and secondly even when you accumulate enough money to be paid like

23. 10:11

    you accumulated 20 50 it does take multiple business days for money to actually arrive so the gig workers from from the time the work was performed until the time they actually paid it's a lot of time uh and all of those problems can be solved with a single smart contract where you just change the model so that when the requester comes requester provides the inputs and the specification of the task written in

24. 10:36

    plain English and then the task is sent to some of the workers they perform it and send solutions back and instead of sending it back to the requester the smart contract affected the smart contract implements a game between between the workers where the review work of each other so this the task is then sent to review to some of them they reviewed send it back and you need to design a game and this is a

25. 10:59

    very interesting exercise in the game design where the workers do not Converse to some sort of equilibrium where it's more beneficial for them to do low quality of work rather than high quality of work right but if you do design this game well uh then the quality of work will will be very high and effectively what you do at the end is that you do not give the requester a say at all so effectively the requester has only one knob which is they can stop

26. 11:25

    the data set like they can say hey as of this moment do not give any more tasks to people but for every task that was given to people they have to pay assuming the workers internally concluded that the quality was high through the game um right so in this model because it's a smart contract first of all payments are solved because the workers the moment their work is accepted the payment is

27. 11:48

    just sent it doesn't matter if it's 75 cents or one dollar on the blockchain the payment is cheap um secondly from the requester if the game is designed properly and I will talk about it in a second the quality will be consistently High and for workers first of all the requester doesn't have a say in whether the quality was high it's the game between them that ensures that whether

28. 12:12

    they paid or not and as long as the game is designed properly they will be paid whenever the quality of work is high so they're not at immersive the requester anymore and as such the support is not really needed uh and finally the payments instantaneous the moment your work is reviewed by the by the others which usually takes less than a couple hours you get paid um and there has been a very long

29. 12:36

    lasting experiment on here for the past two years around designing such a game uh and it's been running it's been running as smoothly successfully and for the in the past half a year the quality of every task that was annotated in the on the platform was extremely high significantly higher than you can get in any centralized solution with a very little cost overhead so something I didn't mention is that the sheet that the platform takes like crowd flower is

30. 13:02

    between 20 and 30 percent actually it's a massive overhead which is gone because there is no intermediary anymore and before before I can conclude this presentation I want to give you um I don't want to go very deeply into the game design of those games uh at some point I will write about them but I want you to give you an example of a failed game of a game that was designed and was considered by the designer to be completely foolproof which within the week was completely broken and people just workers just

31. 13:34

    completely wiped out the uh the smart contract took the money without producing any meaningful work the um the particular data set in question was a very simple data set you the workers were given an image and needed to provide a description and the game was the following the game was that um given an image the worker gets one of the three assignments assignment number one they produce a description assignment number two they get someone else's description and need to say whether it's good or bad and assignment number three is called a Honeypot and

32. 14:06

    the Honeypot is necessary because if you don't have honeypots people will obviously converge to providing garbage descriptions and always accepting uh but if but if you do have honeypots the Honeypot is when a person has to intentionally make a mistake in the description now you cannot always accept right and so with honeypots the gameplay was considered by the designers to be completely foolproof and so this game was launched on the contract uh that

33. 14:29

    happened more than a year and a half ago now and within a week the designer did look into the data set and observed that people are submitting complete garbage and and they managed to get around the honeypots and what happened uh in in reviewing in in effect doing the Retro on what happened and trying to understand what happened was that initially people would properly describe

34. 14:53

    pictures when they give on a regular task and do some Honeypot when they have to the Honeypot but they quickly observed that if you do a Honeypot you actually don't have a motivation to do a high quality work because the task has to be rejected by Design and so several people a very small fraction initially started always doing the same Honeypot which says hey there's a Spiderman on the picture um and then because people do review of each other people who did not do that

35. 15:17

    initially would notice that there's an easy way to complete honeypots by just mentioning a Spiderman and so more and more people over time would start mentioning the Spiderman until effectively every single Honeypot would mention Spider-Man at which point all the reviewers realize hey if you get to review you can just say hey if there's a Spiderman in the description rejected and then there is no Spider-Man accepted and once that happened people realize that if you write anything into the description which does not mention

36. 15:42

    Spider-Man it just does get Auto accepted right and so within a week people would just submit complete garbage without mentioning Spider-Man if it's a regular task they will submit something mentioning Spider-Man if it's a Honeypot and review the work of each other and was fully automated and the contract was depleted right so so the moral is designing those games is not that easy as it seems

37. 16:05

    that concludes my presentation overall I think it's an exciting time and generally I don't think there is much happening today on the intersection of the end blockchain and so there are many many opportunities for people to to start working on so data annotation I think is one of the very lucrative opportunities some idea around building a dial on top of a massive compute cluster is another so yeah if you guys interested in AI That's something definitely to consider as your

38. 16:37

    next gig thank you