0:00
We're taking a fundamentally different approach.
Search this video
0:02
We see a lot of really cool new approaches.
0:05
Symbolic is something like images, words.
0:07
Semantic is them, but understanding of them in context.
0:13
But the missing part is semiotech where you would be able to connect and change the meaning of a specific word.
0:18
Not just in context, but globally.
0:21
It would go and rearrange all the pieces to fit right into actual proper [music] network of things as you set it.
0:28
Not as it just predicted, but also to have a flexibility to set things as they are.
0:36
>> It changes not just the semantic part. It changes everything. >> Yes.
0:45
[music] We sat down a number of months ago with Mahilo Marinko.
0:47
Misho was one of our very first Oshanaugh fellows and we got into everything from cyber security and Chinese hardware vulnerabilities to the future of AI infrastructure.
1:00
Please enjoy my conversation with Misha. Well, hello everyone. It's Jim Oanosy.
1:12
I'm very happy Misha to welcome you. >> Thank you Jim.
1:16
>> I know you quite well but our audience doesn't.
1:18
So let's go with your origin story.
1:21
>> My origin story starts back in Ukraine.
1:25
Uh where from a very young age I was always in my father's phone repair shop.
1:32
It was a place where I was hanging out since I'm an infant and primarily it was a place that inspired me and shaped me into who I am today.
1:40
It's the place where I discovered numerous technologies went down into how things work and that is pretty much where I came from.
1:49
talk a little bit about how you discovered how these technologies work because I think a lot of our audience, you know, they go and buy their phone at the eye store and they think, okay, it's all done and I have to do anything and there might be some things in that uh hardware that regular folks don't know about.
2:13
>> Oh, there's so much things to say about this.
2:16
So, a little bit background on this part is that most of the consumer electronics that get into Ukraine aren't officially there and most of the things that if you for example bought a phone somewhere in Europe or brought it from America, you would want to have a Ukrainian language in in settings.
2:31
Uh or first iPhone would be a perfect example.
2:37
Uh when first iPhone just came out, it was an exclusive contract is AT&T, right?
2:43
without an AT&T SIM card, you wouldn't be able to use it.
2:45
And people were bringing these phones back to Ukraine or ordering from states.
2:49
And essentially, you would want your phone to call someone, [laughter] not just have as a shiny object.
2:56
So these phone repair shops would not just repair something in case of accident.
3:00
You would be able to bring in your phone and get Ukrainian language, get it work with Ukrainian SIM cards, make it actually usable back.
3:09
And there's a lot that comes into this.
3:12
Um, most of these shops is like where I came from, those are not just let's order a new part and make it work.
3:20
It's actually going down understanding how the phone works, how operating system on the phone works, how components interact with each other, where are the weak points, how you can flash it to more than it did before in a sense.
3:34
that requires a lot of deep technical knowledge in a sense uh and a lot of kind of tinkering in order to get right.
3:41
right. I was discovering initial my initial steps would be just getting consumer hardware and trying to solve a problem for a person and if the person would be just like okay I I I don't have Ukrainian maps how can I just like use
3:59
this phone to drive my car most of the people can why can't I then you go and okay how is the building maps working why is it u logged or similar things and you go down and you look uh different ways on how people could have written the app to make it work. Then you would
4:14
Then you would go and actually try to fit it in order to satisfied customers need.
4:23
>> While you're doing all of this, you you have no formal training in this at all. >> Of course not. No.
4:28
[laughter] >> And to remind it's it's again I was really young. It's before I was nine.
4:33
Uh when I was 9 years old, I got my first gigs as a software engineer. >> Wait, stop.
4:41
You were hired as a software engineer at age nine. >> Not hired.
4:47
Technically, you cannot hire 9 years old person.
4:50
So, um no, it was more like freelance and doing some basic web development, >> but or also helping out with the father's phone repair shop as a some source of income to fund my toys.
5:03
Was it just a natural talent on your part?
5:09
>> I think that it has a lot to do with my father having a lot of tools and toys and me just wanting to touch and experience those.
5:18
There's like an early pictures of me being really young and playing around with electronics or disassembling my toys just because to break them to do anything with them just like I need to know what's inside.
5:30
So, you're familiar with Claude Shannon who came up with information theory.
5:34
Uh, that's kind of how he spent his early days.
5:40
He was a freerange kid, but he just wanted to figure out how everything worked.
5:47
Did you have a process or was it literally were you just tinkering? >> Just tinkering. There is no process.
5:55
There wasn't an end goal to satisfy the needs.
5:59
needs. it sometimes you for example um most of the background on me having access to tools it was either having something from the father's phone repair shop or um later on in life uh I was living with just my mother uh I wouldn't
6:15
have access to a laptop to code and there's going to be only two sanctuaries in my life in order to go and actually learn and tinker more it's going to be either my IT classroom or my mother's for work and on mother's work you would have a lot of corporate protections. You
6:31
You would not be able to do much things and this is where most of the cyber security background started to stem from also and sometimes you meet an end goal.
6:42
You want to get some random corporate object to do things it should not do [laughter] in order to just code.
6:51
Um, in some scenarios you would just disassemble a toy to understand how the mechanism works just out of curiosity.
6:58
At some point you would just want to see how this thing looks from inside and there's no purpose to to do so.
7:07
>> Let's talk about that for a bit.
7:07
Um, what kind of vulnerabilities did you find and continue to find in commercially available software, phones, hardware where the innocent buyer is like, "Oh, this is totally secure." And and it's really not.
7:25
>> Couple of weeks ago, this is one of the most recent examples.
7:27
Uh, I got myself a MIA speaker phone, a microphone.
7:30
The microphone itself is a very cool piece of hardware.
7:34
It can listen and pick up your voice from 16 feet away.
7:37
It's a nice marketed product that would go into enterprise kind of meeting rooms uh and be used as a speaker phone.
7:44
I didn't use it for a meeting yet.
7:47
[laughter] I didn't like at all utilize it in any way.
7:53
uh I just got it and being the security professional I would go and check if there's any firmware updates and I wouldn't do so on microphone but I would go onto a manufacturer's website and see can I download it and just flash my microphone with a new fiber offline and when I seen the file uh I've seen that the file was trivially unpackable so you could see what's inside and my natural curio electricity. I unbug the microphone.
8:24
I start it up and they see it can connect to Wi-Fi. I'm just like, okay.
8:29
Most of the recorders are usually just a recorder, just an audio recorder that you later on connect to your phone or a laptop and you upload audio and some server processes it and returns to you an transcript.
8:42
This microphone stood out because it had a full ability to have this networking stack that your laptop or your phone has in full its capability and uh that drive my curio bit more to know what this microphone actually does.
8:57
So I got this fware file and since it wasn't trivial to get to know what's inside it, I unpacked the fware and they started kind of looking around what's inside it.
9:10
Maybe there's some modifications I can do to this thing.
9:11
Essentially I got to know that microphone does transcription inside itself.
9:16
So microphone doesn't go to servers like any other voice recorders.
9:20
You would get the file upload it.
9:22
Uh this microphone would do the transcription part not summary but transcription part on device itself.
9:27
And I was curious how they achieved it because it's a relatively cheap device.
9:33
It doesn't consume a lot of power.
9:33
It can be on its battery for a long time.
9:40
and they just out of curioity opening how their ASR pipeline automatic speech recognition pipeline looks like inside and there's one file that is not obuscated in any way it's not hidden under I don't know I would have hidden but [laughter] uh that contained two political hot words that are specific to China that
10:02
essentially were just there so the firmware even though the device was bought on American soil from Amazon contained code words that I don't know what happens next if you say them, but the device is definitely designed to recognize these specific two highly censored words in China phrases in China and something happens. So, I didn't dig
10:22
So, I didn't dig deeper yet [laughter] into what comes after you say these words, but it's definitely designed to do something.
10:32
>> So, I'm assuming that the microphone was manufactured in China.
10:35
manufactured in China. Yeah, the company behind the microphone is a I think don't quote me on number but it's a multi-billion definitely multi-billion I think it's 10 billion dollars it's a publicly traded company in China it produces primarily 360 cameras I think they've got into
10:52
audio space a little or at least enterprise space a little and yeah it's a Chinese company >> so what counter measures I mean obviously I'm lucky to have you [laughter] to Let me know whether >> use this microphone all around >> whether whether there is a bunch of spyw wear. What can regular people do?
11:12
What can regular people do? >> Nothing. >> Nothing.
11:17
>> There's literally nothing.
11:17
Uh there was another part of it.
11:19
As soon as I got these uh words, I was curious to see what's inside the privacy policy.
11:24
So I got the privacy policy.
11:27
Uh I got the Chinese and English version.
11:30
And Chinese version contains a bit more subprocessors.
11:34
In China, they have most of the public cloud providers are sharing data with government.
11:40
It's a publicly known fact.
11:42
publicly known fact. And their privacy policy basically mentions that yes all of the things and or part of the things would be transferred to Jen service for actually processing the things where American privoxy policy mentions that
11:57
essentially it would go and get processed on AWS on Amazon web services and their march the material on US soil shows that this is AWS but most companies don't do this most companies have redundancies backups they don't store information just on one in one country. There's global replication and
12:15
There's global replication and a lot of need for this information to be available in multiple spots.
12:20
So I would assume is that the Chinese privacy policy states facts that are true about this microphone.
12:29
Hence having the firmware that works the same way for any regen for any customer that stacks these words essentially when you upload this transcript to their cloud or maybe microphone does it in some other secretive way that we don't know about you might be flagged and also the microphone has an ability to kind of your voice is part of your bio.
12:48
There's a couple of things about this microphone that allow you to segment people.
12:52
Means that if it transfers this information to Chinese, it also essentially gets a part of your biometry, part of what you said, your words to identify you.
13:02
And this is what's scary.
13:06
It's not that it just would know that someone said something.
13:11
any recorder would in a sense, but that it would transfer who you are, where you are, and your voice saying you so like that's terrifying [laughter] and does the United States in particular or the EU like do do we know about these problems?
13:35
>> I think that comes a lot into business decisions, right?
13:38
As a manufacturer, you don't want to complicate the process of manufacturing these microphones.
13:42
You want to simplify it as much as possible.
13:46
Um, I think the proper way of doing this is to make a regulation where uh these things like having how we process voice, how we process all of this data, especially if it happens on device as advertised, the device is capable of working offline.
14:01
Um, these things should be documented.
14:04
There should be a legal obligation to document these things.
14:06
uh even though it even if this feature somehow disabled if you chose United States and settings it still should be documented because me as an consumer in the ferr of the device I have on hand the keywords are present >> well and you know having two different privacy policies uh or terms uh of service and the one in Chinese saying something very different than the one in English that seems to be like not a trivial problem to me.
14:39
It would seem to me that these are things that the you know I'm not a huge fan of regulation, but I mean this seems like a regulation that might make sense for American consumers.
14:53
>> Yes, it definitely does.
14:55
>> And and wh why isn't there one?
14:58
>> I don't think there's any particular reason for it other than just isn't there yet?
15:04
>> Let's move on to things like Deepseek.
15:06
Deepseek was super popular when it was launched.
15:09
Do you think that people in this country or Europe or elsewhere knew that virtually every one of their queries was going right back to the CCP in China? >> I don't think so.
15:22
I don't think people think about that.
15:25
They should and at the same time they shouldn't.
15:28
same time they shouldn't. um deepseek depends on how you use that models because uh deepseek models have been open way you would be able to download the model weights and even self-host those and there was numerous American providers that provided an access to the
15:44
model that were territorially located on United States but most people went to download the Chinese app from app store that would process all of this information on Chinese territory >> that's the thing that always astounds me right like we have deepseek uh but it doesn't go back to Chinese servers. But
16:02
But I mean let's be very clear there there are very pe few people who have resources like you and and I was thinking about it this morning and as I was thinking about what I wanted to talk to you about.
16:16
Why don't big corporations have a bunch of Mishas?
16:22
>> Mishas are hard to find. [laughter] You're lucky.
16:25
That's probably the reason.
16:28
[laughter] >> Well, yeah, we we discovered you as as you know, but our audience might not know.
16:34
Uh you were one of our first fellows for the Oshanos fellowship program.
16:38
Actually, it was more of an art project.
16:41
Let's uh take a moment before people are like, uh, what did he just say about all of these things that we can buy in the app store?
16:48
Talk a little bit about that because I was immediately hooked.
16:52
you you you had me at hello when I saw what you wanted to do but but tell our audience about that project because I think it's fascinating.
17:00
>> I connected a bit to my life story as well.
17:03
Um war in Ukraine started and before the war started I would go to visit friends in Sweden where I will end up for a year more before I would move to America.
17:14
Um and during Sweden, I would call that period uh creative boredom.
17:20
I I had a job that I was quite good at.
17:24
It stopped requiring a lot of attention.
17:27
requiring a lot of attention. uh all the things I could have done at the time in order to drive innovation of the company uh or generally contribute as much as I can for business I've done and I essentially was left to do this legacy
17:42
systems just to keep up the project in a sense and I was also left to experiment with a lot of new things because they wanted to see if I I can do something else as well for them and being in Sweden uh the weather was atrocious [laughter] It was sad. It's It was dark or it was
17:58
It's It was dark or it was bright 24/7.
18:01
[laughter] Um I essentially just got bored because back in Ukraine, you get all of the friends you ever need.
18:11
You've got your life history.
18:13
You've got your classmates. You've got your family. You've got everyone.
18:18
There I had only one friend and his girlfriend.
18:20
And I started to explore I started to explore Sweden much much more.
18:24
Essentially, I came to one of the raves musical events in Sweden where I met a lot of wonderful people.
18:32
My first night at that particular rafe was uh a guy getting me into the room where all of the preparation goes.
18:40
So, it's not kind of a venue space, but kind of a new space.
18:44
And I see a vibrant hacker and artist community.
18:47
First thing I see is people uh disassembling a Tesla battery from a car just like randomly on the table.
18:53
You you enter the room there's like on the background you hear like loud beautiful tech.
19:01
Now you open the door and you're like okay the person draws the person makes a weird big sculpture.
19:05
Another person just goes and probes the Tesla battery.
19:10
[laughter] And essentially I I got along with the creative community there. Uh it was amazing.
19:17
I met a lot of wonderful people who weren't incentivized a lot to do what they did but yet they did it and they met a lot of hard tests and I I never thought of code as a routine for myself or engineering for that matter as a routine for myself.
19:33
I always try to kind of not think outside the box but be kind of attentive to things I I have around me in order to drive my work.
19:44
Since I met a lot of artists, I try to paint.
19:46
I try to make sculptures.
19:46
I've uh got a lot of I've visited numerous art events uh both very high-end and quite shitty as well.
19:57
And essentially there was a couple of artists that I wanted to work with and I was doing a lot of design work as well.
20:04
I wouldn't just engineer software.
20:06
I would design how user would interact with it.
20:08
how to make the business decisions well with design.
20:14
Um, and then I was just like, okay, let's play this to art.
20:16
Let's see how how my engineering skills can can work to create something that serves zero purpose in order to make any revenue.
20:27
[laughter] Um, I met a Ukrainian artist and essentially she we wanted to collaborate on some AI things.
20:35
We found that generating images at the time wasn't as appealing for her collection.
20:41
We tried training models on her work, tried approaching it from so many different perspectives and ways.
20:48
And one of the guys in the space was just shouting, "Use brain."
20:53
[laughter] And was like, "What do you mean use brain?
20:57
We are using our brains."
20:57
And was like, "Well, quite literally, buy an EG helmet and try to around with that."
21:04
And I'm like, "Oh yes, this sounds fun."
21:07
[laughter] And that's how the descent in one of the projects for a fellowship started.
21:15
Essentially, we wanted to replicate uh what your visual cortex processes and correlate it to a model that would be able to generate approximately the same image.
21:26
I think that back in season the project wasn't as quite successful as it should have been.
21:31
I think that most of the work it was more exploratory.
21:37
Um and when I came to states there was two performances that went amazing where the project would be actually used where we would get an EEG helmet and do real realtime processing to get as much of your visual cortex while musicians or artists perform and present it on stage to audience to see the imagination of an artist in real time.
21:57
So yeah, that's one of the projects under the fellowship and the other one was also quite fun.
22:02
Uh second project was analyzing crowd behavior primarily.
22:07
It was just amazing and I was fascinated by raves.
22:11
Most of the musical events you would visit in the modern day you would see people just tearing their phones and going there to post a new shiny picture.
22:19
They wouldn't go there because they enjoyed the music or they enjoyed the music but didn't appreciate all that much.
22:26
They didn't appreciate artists.
22:28
They would go there as a social gathering, yes, but not for art.
22:34
And essentially, most of that kind of underground scene for raves, you would find people who are actually enjoying weird alien sounds that make zero sense normies just not staring their phone at all.
22:48
Where you would have a sticker on your camera that would prevent you from taking pictures on these and people would be free.
22:55
People would not care about time.
22:57
People would not care about social media.
22:59
People would not care about anything but having a great time.
23:03
And I found this as okay.
23:03
So let's compare how do crowds behave in these two different scenarios.
23:11
what happens in normal musical events and why our attention factors are quite less when we not just have the phones but what's different between specialty underground hard to get events and kind of general public huge festivals and I got my hands on a couple of sensors like
23:30
seismographs infrared cameras and they would start tracking people I would see how people react to different parts of the music to different styles In some events, we would collect Spotify, Apple Music to get their general taste as well. In some events,
23:45
In some events, we would even get an optional face a person would be able to submit to correlate a specific person to his preferences and to music.
23:56
And essentially, the first initial version of that project would be an artist that plays in real time would get a distribution and real-time charts of how different parts of the crowd engaged.
24:08
the parts that are not engaged, what do they like?
24:10
All of these kind of cool and nice things.
24:12
And then later on, we would bring generative AI to generate music in real time for people who are less active at the overall event to actually drive not just attention factor but actual emotion.
24:27
Try to push people more and more covering more of the audience.
24:31
And generative part went insanely well.
24:34
There was numerous successful events where people would be enjoying and taking their attention fully just to music and to themselves dancing instead of doing anything else.
24:47
>> Obviously the capitalist in me sees something marketable there.
24:51
>> Of course [laughter] be because h how much did the scene change?
24:57
I mean can you quantify it for us?
24:59
uh from 60ish% to almost 90.
25:06
>> And did people know what was going on? No, they did not.
25:10
>> Not a single event was talking about what's actually going on where when you would buy a ticket, Spotify would be an option, but it wouldn't be required. >> I see.
25:18
And and how many people gave up the Spotify playlists and everything else? >> Less than 50%.
25:26
There wasn't a higher n number than 50%.
25:28
So, you're operating with a minority of the Spotify data, and I of course love that it's all voluntary and that if you don't want your Spotify, you don't have to give it.
25:40
But what was what was the unlock?
25:43
What was the key that allowed you to move from the ' 60s to the '9s?
25:49
Cuz that I mean, that's the difference between like the most popular RA promoter and all of the also rans, right?
25:56
It's like it depends a lot.
25:56
Of course, the bigger the event is, the harder to drive people's attention to actual performance.
26:05
Um, the general idea is that you as a person, if you like specialty music, you wouldn't like just one genre or just one artist, you would like many different things.
26:14
An artist doesn't know that.
26:17
Especially, they don't know that in the moment when they perform.
26:18
They don't know what the audience might expect or like.
26:23
And sometimes it's not about music at all.
26:25
Sometimes it's about nature written.
26:27
Sometimes it's about the group person came. We would track groups.
26:30
We would recognize groups of people sticking together and generate more relevant suggestions for DJs, for example, to select from their existing track selection.
26:42
So you still have the creative choice in style or music, but to suggest better times to kickstart more and more attention.
26:49
So essentially the breaking point was that you can hook people more and more and not lose this and you can still have the control over your performance.
27:00
>> And did did the uh participants did you talk to them afterwards?
27:04
Um did anyone talk to Yeah.
27:07
And and what were they >> just like general amazement just like this was so good.
27:11
There wasn't a single time where it was just like it just kept going.
27:16
That that would be the tagline usually from people coming.
27:19
it would just keep going uh driving attention or generally just keep enjoying.
27:27
>> Lots of use cases here.
27:27
Let's shift gears into what we've been building because we're taking a very very different approach to building out the AI suite at OSV.
27:39
Freestyle on that for a little while. Freestyle and EI lab.
27:45
Well, we call ourselves the the department for engineering is mostly AI lab.
27:49
We are a small team that develops a lot of fundamentally new ways to look at data to look at how you approach AI yet not trade off on existing things and existing approaches.
28:02
Being a relatively small team, we covered so many things.
28:06
I'll tell about our successes first then I'll tell about what what's going on kind of not so good both with the market and what prevents us from de delivering the product we envision in a way.
28:16
So from successes we we dug deeper than just using LM APIs.
28:22
The first initial thing of how EAB came to be is that we wanted to get LM to write movie scripts.
28:32
That's was the initial you hiring me. >> That's right.
28:35
>> For engineering residents.
28:35
So I would do this part time.
28:37
I would try to create synthetic data sets and actually drive the model to create a cohesive screenplay which would not go well. [laughter] Yes.
28:50
Well, as for listeners and viewers, our mantra is human in the loop, a senitar model.
28:56
We don't think we think that AI just left to its own devices, you're going to get mostly a tsunami of slop.
29:02
Uh but with an human in the loop, if it's built the right way, you can get magic.
29:09
>> So essentially, we went to a gent approaches.
29:11
approaches. Before most of the kind of models we have today are much more suited at the giant sex stuff than we just initially started are approaches to doing screenplay writing and generally approaching doing this kind of a lot of writing work that should be connected
29:30
that should have a higher order understanding of what's going on in a sense general completions the way how language models work in a in a sense where you would get e context and try to predict which tokens come next would be good for writing but terrible at many different number of things. So
29:50
So essentially we tried some agentic approaches we've built out our AI chat uh thing we still have today which I think that many people would say okay building a LLM wrapper is easy.
30:02
I'm going to say it quite uh try to make it right. Try to do it properly.
30:10
Try to make it useful and try to make it work on all devices in poor network conditions in all of the possible edge cases.
30:21
This is because what product is you polish for people.
30:23
You don't want people to get stuck in chat and then try to say to me how easy the chat is.
30:32
[laughter] Although there's very much so different set of alternatives in the market like cloud cloud is so much better in my experience than open models when it comes to work when it comes to personal things.
30:45
Somehow chat GPT is a little bit better.
30:49
I don't know why OpenI did insanely good job on making a good product.
30:55
the chat, the user interface, the responsiveness.
31:00
When you open the app, you have first time to interaction, right?
31:05
You have everything set, right?
31:07
And that's one of their insane advantages.
31:10
They were the first one to try to make this product and they did it good and it would be really hard to compete with them.
31:16
I think that we we on the level where our chat platform is still quite buggy because again it's in it's in beta but we on at the technological level where we cover most of what their team did in sense of product yet we added our nice things and our unrestrictiveness and all these kind of scenarios.
31:36
So that's the part number one. It still exists.
31:38
We need to collect training data.
31:40
And there's part number two.
31:43
How do you approach AI products with all the OC's craziness and all of the try to create a movie script then try to actually generate video based on that?
31:53
And there's a lot of agent products out there.
31:56
There's a lot of open source projects that would try and do this, but they're fundamentally wrong.
32:01
They're not the way you should be doing that.
32:04
And the the deeper we dug, the deeper we found issues that relate to how general businesses work, how big data was structured, how it was initially optimized for hardware that was much slower and much less flexible and much less resilient that we have today.
32:24
[snorts] How all of this legacy B continue to exist.
32:26
How most researchers don't want to go into software engineering and they need to.
32:33
So most of the research researchers would think good about mass.
32:35
They would think good about general concepts.
32:37
They would do a good research but they would never do a good engineering.
32:44
You would get a great model but not great engineering behind that model.
32:47
All the tooling that you do when you train models nowadays, it's still this way. It's not universal. It's not portable.
32:56
You would adjust so many different things and you don't want it to be universal, but you want it to be portable.
33:01
you want it to be extensible.
33:03
Right now, it's it feels more like people are building temporary solutions for small vents to drive their business and appearance better, but not build something that drastically shifts how they do their work.
33:16
they do their work. And that's where we come in with a lot of our new sets of tools approaches to how we store index process train do all these sorts of things and even maintain our own hypercon converged infrastructure and bare metal infrastructure and yep that is also a lab in a nutshell where we try to approach it uh fundamentally different in terms of how we approach data and with that comes a set of challenges remaking the whole ecosystem
33:48
system trying to make novel approaches that I probably will mention a bit later uh which is interplanetary link knowledge or gar and our upcoming flagship product infinite canvas uh that basically change how you look at
34:03
AI in general how you look at AI data how you look at data sharing how you look at data brokering and exchange how you look at data indexing how you look at AI products where it's not just chat >> since we're talking about it. Let's
34:16
Let's Let's do it now because I I am every time that I go in there, I'm blown away by what you're achieving because I of course use all of the commercial large language models and yeah, you're right.
34:33
So, Chat GPT I always think of as more in white tie and tails and you know, Claude is is much more relaxed, but we're we're taking a fundamentally different approach. >> Yes.
34:45
So fundamentally different approach.
34:47
There's so many things this can mean in a sense, right?
34:50
There's all technological things I just mentioned, but there's also a user side of things.
34:59
You don't want there's a way how we build our AI products today trying to connect to technologies that fundamentally weren't designed for AI >> that fundamentally didn't predict AI would exist in such capacity we have today.
35:14
So people connecting agents to databases and all of these things, it doesn't look and feel right.
35:22
If you're deep into engineering and from my world, it it just doesn't feel it doesn't set right.
35:28
Even by today's metrics, we have what was it around four years now, three years now where people are trying to do a gentech product.
35:36
You see success stories like Corser and they'll go deeper why this is success story and you see a lot of shitty products that are essentially just either trying to be products like cursor who took an actual proper and innovative approach to do what they do or just try to glue shed and sticks in order to make it work connecting things that weren't designed for each other to do something.
36:06
Infinite canvas is probably a new category of AI product that didn't exist before.
36:13
It's what I like to call it an infinite dimension reactive computational whiteboard. >> Very Philip K.
36:24
Dick of you, [laughter] Ballas.
36:27
Essentially, it's a whiteboard that understands what you do.
36:30
Uh it's not a diagramming tool.
36:32
It's not a tool where you would be drag and dropping your tools and connecting things together.
36:37
It's a tool that orchestrates itself that coexists with you where you are capable of changing any part of the process and complex orchestrations which I will go to in a second uh down to their smallest details without having any domain knowledge.
36:57
Infinite canvas should be able to provide you with an ability to even train your own models, orchestrate a lot of different >> and and let me stop you.
37:03
You mean like uh people olds like you?
37:08
>> Yes, like Jim [laughter] Far.
37:12
>> If you achieve that, you're going to win a Nobel.
37:14
[laughter] >> No, I'm not.
37:19
Uh so when it comes to these products, we see a lot of really cool new approaches to collaborative environments.
37:27
One of the best products we have today that I enjoy a lot that I'm sad about because um I feel like they are taking the wrong turn right now is Figma.
37:38
Figma is essentially a tool for designers where they could wireframe design really complex things, build prototypes of apps without kind of having a domain knowledge. It's not a no code tool.
37:50
It's essentially drawing rectangles but doing it well and doing it like 100 people can do it in parallel.
37:56
Um, and what's cool about Figma is that they didn't write another shitty layer for like Adobe does.
38:01
They would take their code base that exists from 80s or 90s to this day and try to continue on the legacy approaches just to not break their existing user base.
38:12
Figma did a lot of drastic changes to how you approach general UI UX design.
38:19
Uh at the time it was a killer product that you don't need to install.
38:21
You open it in a web browser and it works.
38:24
It magically can do much more than any other product was able to in your web browser.
38:31
And we drawn a lot of inspiration from a lot of products like Figma that are essentially kind of infinite whiteboards, infinite canvases.
38:41
There's like a whole curated list of websites that mention infinite canvas products where you would be able to diagram to design to prototype to do all different sets of things.
38:51
But the biggest issue with these is that for example Figma, they would add AI features but they would not fundamentally shift some of the things in their product to make these AI features right and sit well.
39:06
What does Infinite canvas pro for user?
39:09
Essentially you can dump any information you want.
39:11
You can use it structured or unstructured. >> Both. Yeah.
39:15
You you can do spreadsheets.
39:19
You can upload manuscripts.
39:21
You can just drag and drop it and it's kind of a spatial space that would use machine learning to automatically arrange your space as well.
39:31
So as soon as you dropped like thousands of files, it would essentially just organize information for you.
39:35
Then essentially you as any other LLM HL language you're capable of doing anything like any other LLMs do.
39:44
So in a sense what's in these files but what's crazy about this product is that you would be able to kind of say I want all of my characters to be a different way in my script and it would go and parallelize itself and use semiotech approach semantic primarily what we have today we have symbolic.
40:06
So symbolic is something like images, words, semantic is them but understanding of them in context, right?
40:15
But the missing part is semiotech where you would be able to connect and change the meaning of a specific word not just in context but globally where you are capable of reacting dynamically to new information or making adjustments and still operate on symbolic and semantic environments.
40:37
And also the symbolic part of most of the AI today is not on the level required for complex automation or orchestration or being able to also spend it not just to basic business automation but to do it for actually writing a book right some of the most
40:56
wonderful use case I I've tried previously that works well today and some of the things that are upcoming um essentially you write a book you have many different characters and you uploaded your first outline or asked it to do a first outline, your idea. As you
41:11
As you expand, you don't lose any other properties.
41:15
You can easily navigate between your drafts.
41:17
You can put there a lot of informations, tens of thousands of pages that you can still easily navigate.
41:23
And then you are able to kind of tell it, okay, let's write it. Let's actually out.
41:28
actually out. And you have a nice what you see is what you get text editor that actually would properly work that would allow you to see your final book in a sense or allow you to modify anything in that book and maybe there's some
41:43
character that owned a network of hotels like just generally or some traveling company and that traveling company would be named after your character and around that there stems a lot of different parts in your story about how he might
41:59
have been have fun of making fun of the company name for example or he was bullied in childhood or different these sets of things that rely solely on understanding and deeper meaning in general overall context higher order level thinking about it if you change
42:16
the character's name if you ask the canvas to change the character's name you have the flexibility to change all of the points of views in your story cohesively in order to make it Right? Not just ask how would it look like or
42:28
Not just ask how would it look like or let's do it or let's adjust this or these parts.
42:34
It's just magically would go and rearrange your story beautifully and cohesively into one piece.
42:38
Imagine putting this at this these capabilities to a test like let's say in 5 years you would be able to generate a movie or maybe a bit more than 5 years who knows maybe it will be tomorrow.
42:52
Uh essentially when you would want to generate a movie, you would want to have an ability to have a stable trajectory for people or crowd moving because it's essential for your story.
43:05
You would want a lot of these symbolic things because it's all symbols to be in place and you want your video generations or image generations or character appearances to be consistent and you want them to be physically consistent as well.
43:20
One of the technologies I'm proud of that we made in the lab, we call it canvas kit.
43:27
Canvas kit is essentially a multimodel environment for information to exist and be represented and be operated on.
43:36
Canvas kit allows you to visualize almost literally anything when it comes to structured or unstructured information.
43:42
both to having real-time stock information and to having a scan of the room or a 3D space and seeing the process of how this 3D space would be converted into a scene in the movie for example.
43:55
And this is all one cohesive product that you don't need to go into something like a 3D editor like Blender, then go into Photoshop, then go into all of these places and yet you will not trade off on features.
44:07
So essentially you would potentially be able with this technology be able to simulate real physics for your movie or also provide VFX for your movie provide even aperture or lens parameters for your shot that are not purely dependent on AI generation but on actual post-processing that make it possible.
44:28
So this is what infinite canvas is.
44:30
Uh a completely new approach to how you interact with complex, structured, creative, boring, whatever information in a huge and scalable manner that can self-mate, self-chestrate, interconnect with other processes and allow you to orchestrate not just boring business work but also create stuff.
44:53
>> Let me translate it into normal human speak.
44:56
Uh if you're a writer uh using this, you can be writing your novel or whatever your screenplay or whatever and you change that guy's name, right?
45:07
And rather than having to hunt and peck throughout the entire document, let's say it's at 500 pages now, and this guy is a really important character, and then you have a writer room, and everyone in the writer room is like, "Oh, this character is awful.
45:24
we've got to, you know, really really upgrade him or her.
45:29
They'll be able to do that.
45:29
But then the it understands context and it goes and makes those changes on your behalf.
45:38
>> It doesn't only understand the context of semantics, right?
45:39
It understands the context of symbolic things.
45:42
It can invert symbols like change a relationship between a character or someone and convert it into a semantic text that would be consistent.
45:52
So give another example of that because as you know this is one of the things that I get really excited about when you tell me but like the first couple of times I was like Misha talk to me like I'm a small child or golden retriever to steal from the movie margin call.
46:09
>> Sure imagine you are your company is a complex system that is alive that operates both with or without you.
46:16
The company has multiple verticals.
46:21
The company might have thousands of different processes inside all of these verticals and essentially you want to merge two verticals or you want to change one process and you are able to predict every single part of what it means and put it into effect.
46:37
This is another example, right?
46:40
If you do complex business automation, it wouldn't be just hey, let's intake emails.
46:44
No, it would be also higher order part, but what it means for the business, what's the actual meaning of why it exists?
46:54
What's the purpose of all of this?
46:54
And you want to change a higher order process, but how do you repuzzle all of these things together?
47:02
This is what canvas is possible and capable of.
47:04
it would go and rearrange all of the pieces to fit right into actual proper network of things as you set it not as it just predicted but also to have a flexibility to set things as they are.
47:22
>> So in again translating it it it changes not just the semantic part it changes everything. >> Yes.
47:32
essentially symbolic thing is I would tell you this is a glass right there's a glass that is a glass on the window um it's the same glass so glass means something as one thing right but the glass on in a cup is has a bit different meaning right when I speak to you and I look at the glass I say a glass which means cup if I look at window and there's no other glass around me or I haven't drawn attention to any glass around me.
48:06
I would say okay so this glass is dirty for example right it would give you a meaning how does this work glass is a symbolic it's a symbol semantic part is context so when I look at the glass for you to understand that I mean I'm talking about the window you see the semantic part where I look at the window the symbolic part would be that this is essentially a material but there's many ways to look at symbolic things in world, right?
48:36
There is mathematical equations for example where you would have symbols and how those symbols would interact and you have operations which is also some kind of symbols that perform computation and essentially what allow EIB would allow you to do is to change the equation for your whole book and to do it consistently and without the hassle.
48:55
So if you have an equation where you have a couple of characters doing the things in a perfect balance as you see it, you're still able to follow your equation or change your equation in general in order to rearrange everything that comes into your book while still keeping it cafe.
49:14
Could this potentially become something where we ship a book or a video or whatever and the end user who we don't know >> can actually change the story.
49:28
>> can actually change the story. Yes, it's part of it of course or if we ship a movie for example if we provide a movie this whole technical kind of infrastructure and frameworks and libraries we developed internally
49:45
allow and predict for the future of making films for example and what if someone would want a different end not for just for the book but maybe for a movie of the book while still keeping the original intent of the author and his equation in his brain that that is the possibility. Yes. Yes.
50:04
>> Now I know a lot of authors and movie makers etc can be very protective of their art.
50:10
So in those circumstances you can also ship it so they can't change it. Right. >> Of course. Yes.
50:16
Uh, one of the things that we're trying to do with all the various verticals at OSV is ultimately we want to empower the creator and if creator A wants it to be not it's got to be locked and no changes then creator A gets that.
50:34
But if creator B is like no let's experiment.
50:38
Let's see what comes out of that.
50:40
But it will always be at the creator's >> discretion. >> discretion, >> right?
50:45
You and I align on that 100%.
50:45
I being the crazy person that I am, love the idea of being able to write a story and then see all sorts of different ends.
50:58
Will we ever see a time when, let's say, I write a story and I say, "Yeah, you can do whatever you want to this story."
51:07
Will will we be able to let the reader of that story make changes that he or she wants to see as the end?
51:18
Like, oh, I hated the end of, you know, fill in the blank.
51:20
I I really think they should have gotten together and they didn't get together.
51:24
Will we provide suggestions and support or is that also completely customizable?
51:31
>> Completely customizable.
51:33
>> Let's say u they have a very specific Yeah, these two got to get together.
51:36
or they've got to get married or, you know, live happily ever uh for ever after. >> Yes.
51:43
>> But they're like, "I don't like this one."
51:45
They're going to have the opportunity to ask the AI, well, what would you suggest? Exactly.
51:50
But it's also not what would you suggest, right?
51:56
The important part is that most of the things that there's a original creator's intent, right?
52:02
So you would want to see how a creator would change the ending and what were the possibilities for this ending to be only training models and we have sims.
52:14
We've done multiple sims by now and we would do much more as well in terms of dynamically training models.
52:22
Sims are essentially being able to kind of simulate a person, right?
52:24
just like getting as much data and trying to speak in the voice or it's not essentially cloning someone unfortunately [laughter] >> unfortunate unfortunate but it's I I enjoy I I did a couple of experiments for my SM I [snorts] enjoy writing emails using my SAM it's amazing I don't ever need to change anything about these emails it just perfectly captures my humor my points of views things I would most probably discard just exist on it Okay.
52:53
Uh and it only comes not only for SIMS but when you train an AI model when you show it so many different patterns and same way as human brain you don't always remember specific things you recreate the appearance of these memories.
53:12
You are synthesizing from your training data.
53:16
you looked at things and you would not be able to tell like exactly how that tree was looking like, right? Same thing with AI.
53:27
When you train models, it would reconstruct from the data seen previously during training stages.
53:32
And they think that they disconnect on these ends.
53:37
And something that we're looking at in terms of some of our internal R&D is that even if you train a sim gem, would it be able to look at the original things?
53:49
Because you as human, you would probably be able to go into one of your notebooks and take a look at it and you would remember some things differently >> which I've told lots of stories about.
54:00
I've kept journals for people who don't know since I was 18.
54:02
I'm 65, so I have lots of journals.
54:07
And it was one of the things that really unlocked my understanding of human memory.
54:12
I would I would swear on a stack of Bibles that I believed a certain thing.
54:18
The story I often tell is about uh the first Gulf War and I was at a party here in Manhattan and the group seemed to be like, "Yeah, you know, the first Gulf War I support it because you know, Saddam went into Kuwait and that was crazy."
54:33
and I was saying the same thing and then the group would say yeah but this one like this is a bad idea right?
54:42
So I completely believed that was my memory that I had supported the first Gulf War and then I was looking for something else in a journal that I wrote right around the time of the first Gulf War and when I read it I realized I did not support the first Gulf War.
55:00
when you go back to the real time writing.
55:03
And so it's quite a shock really because it led to my theory that memory is often overwritten with our current beliefs. >> Right. That's true.
55:16
>> And so when you think that you remember something completely crystal clear, you're probably wrong. >> Yep.
55:26
>> So let's talk a little bit about the Sims because that was another thing that I'm very excited about.
55:30
Um, one of the things that I've asked you to do is create synthetic audiences so that we can stress test, do AB testing on these synthetic audiences.
55:43
Um, and we kind of kicked around ideas like maybe we should use big five ocean profiles, maybe we should interview like varying types of people. Go ahead.
55:56
>> I already have my grim.
55:56
[laughter] So uh essentially you you have different practices right?
56:03
You want to do some what what you said the last two like big ocean five. >> Yeah.
56:11
>> Essentially what canvas provides you today is the capability and you can go and basically tell I want an audience from these parameters.
56:22
Can you run simulations on all of these things?
56:24
Can you go to internet and we have by our internal benchmarks that we of course haven't yet published.
56:29
We are quite early to most of the things.
56:31
Um but we have one of the best web crawlers that is out there.
56:37
We crawl and index much faster than big search engines like Google for example.
56:45
>> Brag a little tell tell me how much faster we do that. >> We do it on average.
56:50
uh there is no definitive way but essentially the technology itself is fundamentally different in terms how you interact with web pages and how you follow web pages that proposes a new algorithm that provides ability to crawl a little bit faster.
57:06
Of course, we don't know Google's numbers.
57:07
I don't think much people do know.
57:10
So we can say for sure that we are 100% like or 20% more faster than Google.
57:17
But essentially we are capable of uh reindexing uh more than 15 terabytes an hour of pure web pages. It's insane number.
57:30
It's an purely insane number.
57:33
Billions of pages pure text that are getting indexed not just a typical search index but as a semiotic index.
57:43
And by that I mean if for example inside your canvas you told that oh there's these effects and they want you to use internet knowledge to synthesize or run tests or create an audience based out of internet knowledge and internet knowledge updates.
58:01
You have a history and time travel across internet knowledge almost in real time to synthesize your audiences and see how the opinion changes now versus a week ago for example.
58:11
And this is where symbiotics are cool because you essentially are not reacting to new data but you react to new contents of data and you have a network effect and cascading effect of things that change that per this new information.
58:29
So yeah, so for for our viewers and listeners, uh one of the first things that I said to Misha was basically I want the eye of son to be able to see everything that is going on because obviously this uh you know prediction markets are really hot right now and this maybe takes that a step further. Yes.
58:52
Let me restate again for maybe our listeners who aren't as technically brilliant as you are.
58:58
I am certainly one of those people.
59:01
But rather than have a a a set audience, right?
59:05
I for example, I'm right now in the throws of uh writing a fictional thriller, the first fiction I've ever written.
59:14
When the system is finished, I could literally go in there and say, "Find me the audience for this particular piece of work, right?"
59:26
And then I could change things in my story.
59:30
>> Restructure your audience also >> and I can >> usually without a chat or chat how you want >> and and essentially it might serve me up an audience that surprises me, right? >> Yeah.
59:43
So rather than go to, oh yeah, here are the stats on people who like historical fiction or here are the stats or general personality profiles of people who buy um you know books about World War II.
1:00:01
Um it will synthesize and and give us a new look, right? >> In a sense. Yes. Yes, of course.
1:00:09
And besides giving you a new look that would be just text, this is where canvas cat our representation layer kicks in.
1:00:17
You would be able to see it as a higher dimensional space.
1:00:22
See populations of information, see different populations of opinions, be able to dissect those and run on subset of those for example or on all of these or expand those or subtract or summarize those. Right?
1:00:35
So you have this full multimodel flexibility not just around what AI produced or what you want AI to produce but around your data as well and the data that AI produced as well.
1:00:46
So again stop me when I'm wrong. >> Sure.
1:00:51
>> Um when it's working the way I want it to work, we want it to work.
1:00:54
We could in real time experiment with different ideas we have for the story. >> Of course. Yes.
1:01:01
>> Of course. Yes. So we we have a as you know because you sat in the writer room we have uh what Jimmy Sony called a villain worthy of Igo uh from Shakespeare that led me to believe you know what I don't think the heroes are beefed up enough and when this is working in real time I'll be able to go in and put my ideas for how to beef up the heroes so that they're worthy opponents for this incredible villain
1:01:31
and I'll be able to see like right then maybe uh one of the uh uh heroes is skilled in Akido or whatever but as the author I'll be able to see what that idea changes right yes and more importantly would be able to see the cascading network effect and trace it to see how it changed this is the important part and you can affect that network in order to get different things if you didn't like the end result. So you again
1:02:02
So you again have the full supervision.
1:02:04
Of course you're not able to change how model thinks, >> right?
1:02:09
But you're able to change higher order structures or how the model gets all the things together and that's essentially part of it. Yes.
1:02:18
>> And so this is applicable far beyond writing, right?
1:02:21
Like >> oh I have something to say here.
1:02:24
You're coming from a finance background and of course in perfect world I see the core technology we have today in R&D stages be used to study network effects of traffic jams in New York on financial districts to correlate those with stock price drops or to be able to predict what this would be doing to research and innovation in biotech sector for example.
1:02:55
A lot of highly disperate things we have around us are shaping us in many different ways we don't notice.
1:03:03
But having big data systems that scale so well and have so much deps and yet allow you to orchestrate all of these dabs blazingly fast and without the hustle in a perfectly normal user interface that anyone is capable of using allows you to do some crazy stuff.
1:03:21
[laughter] And one of the ways that I think about this is we're giving our user I I like I love music as you know and I I the more I've been thinking about it the more I've been thinking that the user the end user is the conductor of an orchestra
1:03:40
and they can be doing like no no you're playing that note wrong and no you got to do this and yet as you bring up like markets obviously I I've been known to have a little interest in markets when you mentioned the idea about the traffic around the financial district. Uh I I
1:03:56
Uh I I remember reading in the journal of portfolio management what very geeky publication but like 20 years ago and this guy got this idea that uh stock prices went down when the weather was inclement. >> Yeah.
1:04:16
And I was intrigued because like I love new data sources. I love new indicators.
1:04:22
But what it took that guy to literally produce the So he had a hypothesis, right?
1:04:30
Stock prices go down when the weather is inclement or cloudy and they go up generally when it's sunny and nice.
1:04:38
But it took him years to to even test his own hypothesis. >> Yeah.
1:04:45
Spoiler alert, he was right.
1:04:45
Uh, which kind of surprised me.
1:04:48
But it literally when I read the afterward, he had a a huge team of undergrad.
1:04:55
He was a professor and so he had a huge team of undergrads and grad students like literally down on Wall Street taking pictures and and doing all that and the data capture.
1:05:08
We can do that essentially in real time.
1:05:13
We already do most of the data capturing.
1:05:15
This is the part and I think that we could touch on interplantary length knowledge.
1:05:19
Now um it's a crazy name for a crazy technology uh that is quite new at its at its core concept.
1:05:30
It encapsulates a lot of really complex systems and brings them down into one beautiful cohesive and simple thing.
1:05:36
Um, I can't wait to write about it also when we have it running and we have a bit of benefit in terms of getting our product on the market first.
1:05:48
Essentially, IPLK interplanetary link knowledge is a one unified layer that allows you to have ownership over your own data yet not sacrifice an ability to use this data in all the crazy contextes we were just talking about.
1:06:03
It allows you to both utilize very traditional approaches to data and at the same time allows you to boost those and scale those to very enormous sizes where some of this information might be even stored on your phone and you don't even need to store it on our servers.
1:06:22
I would make a disclaimer.
1:06:22
It's not a decentralized system or as many might think there's a lot of projects like this that tried using crypto.
1:06:31
this that tried using crypto. I would say that this is a semi decentralized approach where you would have part of your information in your phone and you would agree to have this information be having ax time of life on some other end and other end when it doesn't need it it's ephemereral and it disposes and when it needs again it would be able to
1:06:55
go back to your phone and get it if it needs it and you have the full transparency over these things and this works not only for these small examples like user data But this essentially works as a potential new framework for data brokers for being able to stream highfrequency trading data even in a manner that is traceable, transparent and fully appropriate for modern big data systems. And importantly, the the key that I want
1:07:19
And importantly, the the key that I want it built in and you were wonderful about making that a reality is this also allows the the person whose data it is to control that data >> of course.
1:07:36
So unlike a lot of like the last 20 years where essentially most of the big tech companies were we were the product right because they were essentially training on all of our interactions etc.
1:07:51
interactions etc. This gives the opportunity for me if I wanted to use a particular data set that was sort of precious to me to control that right >> of course not just control that there's multiple ends for end users who are actually producing data and for
1:08:11
companies processing handling or operating on these things right when in real world you would share a secret with a person you cannot take it back right But the problem with modern world is that you're sharing a secret and you don't know if you even gave it. And yes, there's multiple disclaimers
1:08:32
And yes, there's multiple disclaimers and a lot of cryptic language that people have to come across yet they don't know if their picture was used.
1:08:43
You don't know if face on your photos in iPhone was used to train a model that would recognize faces for all of the other people which is not a huge problem.
1:08:54
I think that ethics and general approach to privacy shifted a lot from us having villages and everyone knew in villages who was in those villages to industrial revolution in big cities where people would share information with the speed of walking distance to modern day which is insanely new compared to what I mentioned just before about sharing the information online and they think that we're just not balanced Right.
1:09:24
I think the time will come.
1:09:24
I think that there was so many case studies right now and there should be like at least four times more in order for us to catch up and take the lessons to do it right. >> Yeah.
1:09:37
And you and I talked about that a lot.
1:09:39
There is huge cultural lag built into people's behavior.
1:09:42
And as you rightly point out, if if for the vast majority of modern uh history, we lived in small villages or farming communities etc.
1:09:55
Uh you know, you knew who you were telling that secret to and you knew whether they were going to be reli or you thought you knew whether they were going to be reliable or not.
1:10:03
And and so that's like so deep in our human OS code that that's how you got the last 20 years, right?
1:10:14
>> You got you got people just naturally thinking the old way.
1:10:17
H how do you see transforming people to think the new way?
1:10:25
>> Yeah, I don't think people need to transform at all.
1:10:27
I think that uh in our human nature is to share. We are social creatures.
1:10:34
A lot of us are, not all of us, but there are some historical examples.
1:10:40
[laughter] When you work with private information, you share your picture online.
1:10:45
I have a lot of points of view on privacy.
1:10:47
When ever since I was 10, I would work on computer vision technologies.
1:10:57
I would love to play around with facial recognition algorithms, with being able to compare faces algorithmically, try to build databases of faces.
1:11:05
And then at that young age, I was quite scared of implications of what it meant of what having an identity that could be used and primarily it's going to be used against you means.
1:11:20
And a year after I tried to appear less with my face in pictures.
1:11:27
There's been a period in my teenage years where you would not find the picture of my face.
1:11:32
It would be either covered with my hand or I would not be on picture at all.
1:11:37
I would have issues with my in my school because there's a class photo.
1:11:42
You should you should be there. Why are you on exam?
1:11:47
And a lot of these different things.
1:11:47
But essentially what I found for myself is that having extreme privacy isn't the answer, right?
1:11:55
You as a human again need to be a highly social creature especially in modern world to survive.
1:12:03
When you share your face or share your photo on Instagram, there isn't going to be a world where a single person would be able to trace where this and implications of posting this photo goes.
1:12:17
But there could be technology that is lightweight enough, portable enough, cheap enough, and sufficiently made right that would allow both companies to adopt it and users to see where did their picture travel to, how was it used.
1:12:36
So we see web three, which is another end of this.
1:12:39
is another end of this. I honestly hate all the web three space and I know too much about the technology in web three about all the things like IPFS it's interplanetary file system not knowledge and we draw a lot of inspiration from things like multiformats it's formats that can describe themselves
1:13:00
imagine reading something that reveals what it is not just by words but by actual reading it you understand it in a sense same thing for machines that is not based on AI but based on basic cryptography and algorithms that allow machines to recognize what information are they reading without having any kind of super intelligence to do so. Multiformats are very important uh
1:13:22
Multiformats are very important uh concepts that drove a lot of our symbolic research afterwards.
1:13:27
He changed a lot about multiformats but I still love interplanetary prefix.
1:13:32
[laughter] So, >> aim for the stars, hit the moon, right?
1:13:39
[laughter] >> It's it's Yeah, you're going you Okay.
1:13:43
You want to leave the galaxy?
1:13:43
[laughter] I didn't know about that.
1:13:46
But >> I think it's really important for people to understand the point you just made because I I personally think it's vital and and that is the technology can be developed so that literally people don't have to change their behavior. >> Exactly.
1:14:07
And businesses can even save money by doing that.
1:14:10
The problem is that Bellabs was such an amazing part of history of innovation and general things.
1:14:19
They never stopped exploring.
1:14:21
They never abandoned a lot of concept.
1:14:24
They were aware they were just the structure of the company allowed them to do crazy things.
1:14:28
And I think that most of the research we do today and how we build things today doesn't allow as much experimentation as we used to have.
1:14:34
M and the more we build the more we kind of this works this generates revenue.
1:14:42
Why would we ever change that?
1:14:42
Why would we ever change that? It's better to lobby something that would prevent from this ever changing than changing thing to its core which is deeply wrong and I think it's a very bad habit for us as civilization to prevent us from experimenting and I think if we
1:14:59
focused if many companies that do big data today and companies that draw all the sector of analysis and recommendations and add algorithms that actually go deeper into how they structured most of the technology behind is they would have made it much cheaper for themselves, maybe even more profitable and yet fully at home. >> Yeah. And obviously that's something >> Yeah.
1:15:21
And obviously that's something near and dear to my heart.
1:15:23
I I think that absolutely if you want to be in business, you should be in the business of having profits because how you going to pay your staff, how you're going to grow or do any of those things, how you're going to support R&D, all of those things come into play.
1:15:38
So, one of the things that I like about the way you talk about this is you're very practical about, hey, if you do it this way, not only is it going to be much, much better, much much safer, but it's going to be cheaper, and you're going to make more money.
1:15:55
more money. And so is there kind of a mono culture right now in in the big uh commercial designers and and uh people who are producing the large language models that's I mean help me out here now people who produce who generally like I I can say only for some of the
1:16:18
open source world I'm tracking maybe not all of it um I don't think that there's any community about how you handle data in a sense The better the data set you get, the better clean it, the better generate some synthetic data around it, the better your model is. The more data
1:16:32
The more data you have, the greater the model is and the harder it is to train and longer it takes.
1:16:38
Like there's a lot of these variables and balances between these variables, dynamics between these variables that make it happen.
1:16:43
And I don't think that this is the case.
1:16:45
The researchers are busy with model architectures, not data architectures because this is not an AI engineers pain.
1:16:54
It's a data engineer pain and infrastructure team pain in order to make this data be stored and accessible.
1:17:03
So yeah, that's that's the general issue with having models.
1:17:05
Um there's another end to this.
1:17:08
Models would not be we are not yet at the point with the technology we have that allows you to trace the original training information to AI outputs.
1:17:20
There's a sort of compression going on that prevents you from doing this and I don't think we will ever be able to.
1:17:29
You cannot recall most of the things that drove your decisions every day and it's fine.
1:17:34
It doesn't shape who you are in a sense.
1:17:36
We want AI to be able to do so to extend its capabilities but not care about data, right?
1:17:40
But I think that the only way to do it right is to let people be indecision for usage of their data and how their data spanned.
1:17:53
Even if their data was used against their will, people should be able to see how the data was flowing. And that's the thing.
1:18:03
>> Obviously the there are uh infinite workflows that can uh come from this.
1:18:10
Let let's explore a little bit more um how we are trying to across the verticals at OSV let them learn from one another.
1:18:24
>> Essentially imagine that also one of the things we want to have inside canvas is that basically canvas should be able to provide you with an ability to connect your workflows together.
1:18:39
If some person created a canvas, you would be able to embed other canvases or reference other canvases or to be able to send to something else as well.
1:18:49
Also, there is knowledge bases and things how you can connect and semiotic way of thinking about it is that if there's something that interests our VC, we might modify our media, maybe they want to write a book about it, right?
1:19:03
If there's something in terms of a book that would make a great company and we have a network of people who might be able to create it or a network of fellows who long finished their projects and now working in their very boring job for example or we know some people or those people are generally available to our network you might suggest them come together and create a company and we'll be happy to fund you.
1:19:27
I suppose that's that's your part.
1:19:29
[laughter] Uh the idea really springs from uh you know most most of these things like the market are complex adaptive systems and emergence comes from below not from the top.
1:19:47
I've always been uh skeptical about I mean if you just look historically top- down systems really don't work >> at all because they're they're just the absolutely the wrong design.
1:20:01
The information bottlenecks that get up to those people up there are insane.
1:20:06
And it's I I often say to people, could you know a central committee with a 5-year plan ever design a Rubik's cube or an iPhone or, you know, a pet rock? Of course not. Right?
1:20:23
Markets can do that because there are these living organisms where all of the emergence is coming, right?
1:20:33
and where you can try and take this is why I'm such a big fan of your approach to learn all of this stuff.
1:20:40
You've got to be tinkering, right?
1:20:42
When you look at the biggest b breakthroughs that people like Claude Shannon uh came up with, it was because that's what they were. Yeah. That's what they did.
1:20:55
And it seems to me that the reason we haven't had the kinds of breakthroughs in more, let's call it just basic science, right?
1:21:05
Um, is because the whole thing got inverted.
1:21:08
because the whole thing got inverted. We turned it into a top- down structure and people applying for a grant so that they can conduct their research are are are it's a poisoned well because the top-down structure that has emerged is so narrow that like in unless you're
1:21:30
doing this you're not going to get funded which to me is insane because you should be funding it's like one of the things we're trying to achieve on tiny tiny scale with the Oshanosy Fellowship and and grantees like we want we want those people to get funded because that's where all of the breakthroughs come from anyway. Do you see that model toppling
1:21:54
Do you see that model toppling and and a more organic one replacing it?
1:22:03
I think that there's very much points different points of view on this.
1:22:11
Um there's a lot of personal things that drive you, right?
1:22:13
When you when you create something, when you build something, when you're tinkering with something, there's environment that drives you.
1:22:20
There's opportunities that drive you.
1:22:21
There's so many variables that come into this.
1:22:24
I think that bottom top makes more sense just because that's how I see the world as a computer scientist and how I see structures emerge, right?
1:22:39
And I don't see any particularly useful like as I imagine it more less computationally intensive and more appropriate ways to scan trees or graphs or like a lot of different these kind of things. So I agree with you. Yeah.
1:22:56
>> And really one of the things that I it's a thesis, right?
1:22:59
I could absolutely be wrong, but I think that that's where markets come in, right?
1:23:04
Because we're in a relatively free market here in the United States.
1:23:10
And so we're doing this, right?
1:23:14
And so the the way the information gets into the network is we do it.
1:23:22
it leads to like really cool things and then others are like, "Oh, maybe we should be doing it that way as well."
1:23:31
Markets are amazing at that.
1:23:34
Like, one of the things I used to joke about is in in a relatively free market, markets co-opt everything.
1:23:40
You could take like look at the 1960s, right?
1:23:43
And there were these huge movements that were anti-war, you know, make love not war, flower power, all of that.
1:23:53
And literally, it didn't take markets but a minute to commercialize all of [laughter] that.
1:24:02
And and you know, you can be incredibly obscure.
1:24:04
You know, you could be one of you can be Ginsburgg and write Howell and suddenly you've got a fourbook deal. [laughter] >> Yeah.
1:24:13
But in in this environment, I also think that the old like as you know one of the thing I have six grandchildren.
1:24:23
I do not want them to grow up where a panopticon controlled by a few controls everything.
1:24:33
>> That is a nightmare to me. Right.
1:24:33
That's 1984 married to Brave New World, married to you know whatever dystopian way you can you can think about it.
1:24:42
But as I was thinking about it like just on your argument alone like that in its very definition is top down and is going to fail. >> Yes.
1:24:56
>> And essentially we will just see scaling bottlenecks.
1:24:59
You already see scaling bottlenecks in a lot of these systems that heavily rely on private information that start to collapse or collapse or about to again the things will take its own natural order.
1:25:13
own natural order. I somewhere somehow always believed that there's always a path to good that nature follows that essentially will lead us to better technologies to companies making different decisions and with even having new trend in privacy that you can just buy a product a bit more expensier and you have your
1:25:35
privacy like the Apple way or you can get a cheap phone that has everything that Apple has like Google's way but like you can store infinite amount of photos in your Google photos, but you're agreeing to all of the terms that are cryptic and you'll never know what's happening with your photos in the background. So yeah, and essentially I I
1:25:57
So yeah, and essentially I I really love how Apple designs their things and I think that many people just see and even many engineers see Apple privacy tag as a thing that just satisfies them because they need to comply with all the apps or regulations and then they need to design user flows or user would be able to opt out of their telemetry or similar things.
1:26:22
But even Apple's architecture down to how processor works, how they and what I was mostly excited and I hope they will not this thing up uh is private compute.
1:26:39
They essentially architected a way for you to use a bigger cloud compute in a sense that even engineers having directly physical access to that server would not be able to know what exactly did you compute there.
1:26:52
M >> it's a beautiful piece of architecture and technology that goes down to how processors in execute instructions and how your phone is in control of your own cryptography.
1:27:03
Same way as iCloud has advanced data protection that essentially encrypts everything at their site with a risk if you lose the key you lose all the data even Apple wouldn't be able to access it and yes you pray premium for it and this should be your choice and I think that Apple is seeing the right way here.
1:27:24
I see that many companies are starting to see like Apple does.
1:27:30
Uh and I think that a lot of people will p pioneer much better ways like we are thinking and questioning all of these things while we develop our things because it actually allows us to do more and allows us to make most of our research much cheaper than it was previously with how we handle again novel approaches to data uh and how we store it and we process it how we stream it into GPUs than probably anyone else in the market.
1:27:58
Let's talk about Apple just briefly because like to me if if Jim of 10 years ago was going to put a bet.
1:28:06
I would have bet that Apple was the one who figured out the AI or would buy the companies that would make that happen.
1:28:14
What's What's going on there?
1:28:17
Apple is too ahead of their time.
1:28:17
I think that the design and the beautiful concept behind Apple intelligence and its promises are amazing.
1:28:24
I think that they overestimated their own capacity to deliver it.
1:28:33
I think that they will still deliver it just in next like 10 years.
1:28:39
But I think that Apple would be the first company that would properly actually design it in the right way.
1:28:44
I think that Apple got into AI space thinking that it's already old enough, but it's too young and it's definitely far away from being a product that lives up to their level of depths in design and user experience. >> Yeah.
1:29:03
Because, you know, the investor in me thinks that the first company that is able to pretty much guarantee, nope, your your data is truly private.
1:29:12
We can't get it if you forget the key.
1:29:15
That is going to add several zeros to their market capitalization. >> It did. >> I know.
1:29:23
But then actually proving it to be the case going to add several more.
1:29:31
>> Uh what we have today in terms of what struggles they have is one of the proofs.
1:29:38
uh they are genuinely having the freedom to work on things as they see it and they think it's cool.
1:29:44
>> So before we uh ask that final question, let's say everything is working the way we want it to work, right?
1:29:50
So so take me through um a day where we get let's say let's use an let's not use an OSV person.
1:30:01
we get an outside uh screenplay from uh somebody who sends it to Infinite Films.
1:30:07
And what will Nick and other people who are working at Infinite Films once this is working?
1:30:13
What are they going to be able to do with just that screenplay that right now seems like you know what what did uh Arthur C.
1:30:24
Clark say that a significantly advanced technology is no different than magic.
1:30:29
What is he going to be able to do once that system is working?
1:30:32
>> For example, we have an infinite media email or infinite films email, right?
1:30:38
The intake from that email would go directly into your canvas, something that we have planned for canvas to exist.
1:30:44
exist. The canvas would take on the script play or the concept for a movie and you would be able to run your workflow once just take some script like beforehand when you design your canvas that would be intake for movies for example you would be able to do all of
1:31:03
the things like create audience reactions try to extrapolate to a second part if if it's possible at all then go do all of these things maybe create a dynamic set of criteria that would satisfy us enough to be interested in this and notify me if it goes this way and then it up a couple times. You correct it a couple
1:31:23
You correct it a couple times either in chat or manually if you value your time.
1:31:26
Uh and essentially you get a working flow that costed you 5 minutes to build and scales.
1:31:32
Then later on if we hit an example that the system is not being able to handle and this is the cool part about our part and our technology which is runnable crafts uh in a way to approach execution and things that can run and self expand.
1:31:50
things that can run and self expand. V would be able to selfoptimize the whole automation or canvas's network would be able to selfheal if the things go wrong or out of ordinary or things that haven't been covered previously and
1:32:05
canvas would either raise a warning for you that you might want to look at this or it would self adjust and raise another warning that I looked at this and they needed to adjust because this is this this and this was different compared to something that previously run over the network of things I did. >> So it essentially the leverage being
1:32:24
>> So it essentially the leverage being provided by these tools and technology.
1:32:32
I I I often talk to people who are have really really deep at least business domain knowledge here and one of the things that I struggle with is they're not seeing the inherent leverage here the way I'm seeing it.
1:32:48
Like I I people say dude like you're comparing this to Gutenberg and saying this is more important than Gutenberg. I I think it is. Do you agree? >> It is maybe. Yeah.
1:33:02
The last part that I want to mention on like what would happen next?
1:33:06
You would be able to just send it in our organization inside canvas.
1:33:12
You can instruct it if you see any other canvases that might use this information. >> Very very cool.
1:33:17
What are you proudest of?
1:33:20
>> I think that satisfaction would come when I actually deliver on most of the promises. Right.
1:33:25
A lot of things are highly experimental.
1:33:27
A lot of things we discussed is the perfect world we want to see. Right. >> Sure.
1:33:32
>> And I think that we are we came much closer the past year and I think that we are ahead of market in many ways.
1:33:38
are ahead of market in many ways. I think that on contrary to market we have a lot of new unseen problems to solve as well and things that I'm proud of is that first we we've got not a lot of hardware but some of it right we've got our own kind of space in the data center we built our servers we installed those we implemented it it's a thing I'm so far one one of the three things I'm most
1:34:04
proud of is our hyper conversion infrastructure is our software that manag which is our data center that allows us to scale like any other cloud would be able to scale to be able to compete in terms of compute with other cloud providers yet allows our engineers to think like our own hardware is a cloud provider so they don't need to ever change their mental model about how they interact with servers or similar things. Besides this, we've made a
1:34:30
Besides this, we've made a couple of innovations in the space that are internal that are both allowing us to do big data and compete with bad boys like Big Query at Google and similar things to scaling and having dynamic GPU allocation and resource orchestration that I think is quite noble compared to market as well. >> Okay.
1:34:52
So, what haven't I asked you that uh you think is super cool about the work you're doing and and how it relates to everything that we're doing at OSV and beyond and beyond and beyond?
1:35:03
I I think that um unlike any other company I've had experience working with, OSV is my peers.
1:35:13
I see insanely creative people that are capable of much more than opportunities previously allowed them to do and I think OSV is great in a way Bellabs or Apple is great because we have Germany so can help us do our projects uh and at the same time we actually can drive business better one can't live without the other and I think that the amount of creators creatives And the way we as a team see the world is amazing and insane.
1:35:51
[laughter] >> Well, I have a lot of friends my age who say the insane part.
1:35:56
[laughter] >> Well, okay. I think it's amazing.
1:36:00
>> Well, obviously so do I. >> It's amazing.
1:36:02
[laughter] >> I wouldn't be able to meet a lot of tech guys who would be able to operate on these concepts with me, but yet most of the creative people are.
1:36:10
And I wasn't able to find the one person in OSV who wouldn't be as excited as I am about what I develop.
1:36:16
So this is definitely my ego thing where I'm like, okay, I'm recognized. This is amazing.
1:36:20
recognized. This is amazing. [laughter] And at the same time what what the technology we are building is so desperate in terms of how many different things we touch from movie production to podcasts to general media to markets to
1:36:38
finance to writing technical nontechnical fiction and how these things are connected and try to make a cohesive system of all of this is probably a billion dollar question and you know Well, what I think about that I I mean I think that's the that's what we're after here. >> Yes, that's the way also for future of
1:36:57
>> Yes, that's the way also for future of many things. >> Yeah, I agree. What could go wrong?
1:37:01
What what could go wrong?
1:37:06
>> So many things can go [laughter] wrong.
1:37:09
>> We could completely overestimate a couple of current data capabilities.
1:37:16
There's a couple of R&Ds that shown promises, but it's not a product, right?
1:37:22
Uh and we're essentially at the stage where we polish a lot of things and we discover how things should have not been made both existing that are on the market and products we would have loved to use and technology we would have loved to use and I think that with the complexity of things we're building from technological standpoint comes a lot of risks but they are well justified is my opinion.
1:37:44
I think that you're touching on so many different things and so far we are looking good on them.
1:37:50
that even if we do up, it would be a positive some up in a way where we we wouldn't be dragged down.
1:37:59
We would be able to move forward.
1:38:01
I think we've made so many mistakes but first of all your policy we never repeat the same mistake right yet.
1:38:10
Uh and at the same time uh we were able to drive a lot of things we have today into much different level.
1:38:16
So I think we will make a lot of terrible architectural design product mistakes but those are fixable and they would delay the time to market.
1:38:27
they would >> maybe hinder some of the initial user capabilities.
1:38:33
But >> I will tell a story about one of our portfolio uh companies CEOs met you and uh I think very highly of him and he came back and he's like uh yeah no you need to hire him [laughter] because I think you were showing him on your phone a knowledge graph right?
1:38:52
Oh yeah, I was showing him a a part of first R&D that was later be able to now enable canvas infinite canvas implementation. >> Right.
1:39:05
And you blew him away because he came over to me and he said, "I know startups that have burned through millions of dollars trying for this and they've come up with nothing."
1:39:19
And he goes, "That's why you should hire him like right now."
1:39:22
[clears throat] [laughter] I was higher. I know you. That's the first thing.
1:39:28
Second is that uh I think that a lot matters in terms of why do you approach your work?
1:39:33
I don't approach my work because I needed to make the technology.
1:39:38
I had real use cases that I needed to cover and I needed to not create a temporarily solution that would grow into shitty product.
1:39:50
I needed to make something universal, something simple enough that at its core that doesn't require to continue building legacy things on top.
1:40:02
And this forces you to think a little bit out of bounds.
1:40:04
I wouldn't say out of the box. I am in the box. I'm in OSV box.
1:40:08
[laughter] Uh >> we got a pretty big box.
1:40:12
>> You got a pretty big box. It's enough.
1:40:12
I don't need to go outside the box yet.
1:40:17
[laughter] Um, and essentially when you do go outside the bound bounds and you have the creative freedom to take your time and try to catch what you think would be the right thing and just follow my gut.
1:40:33
I have an ability to follow my gut is is amazing and this is what drove us to making something like you just describe.
1:40:42
Well, uh, we could not have done it without you, Misha. That is a a certainty.
1:40:47
And as you know, my belief is take super talented people who are very agentic and let them do their thing.
1:40:57
Uh, I I uh despise top- down command and control ways.
1:41:03
If you want to duplicate and do what everyone else is doing, okay, fine.
1:41:09
But what the you bothering for then? Right?
1:41:10
Like I am I I still wake up almost every day and cannot believe that I am lucky enough to be here at this point in history.
1:41:20
As you know, we were talking about it the other day.
1:41:23
I I was going back through my journals and I've been writing about this since I was 21 years old and I'm like your Yeah. your age.
1:41:34
And I'm like finally it's finally finally here.
1:41:37
And uh the unleashing of the creativity that that you have I think is the key.
1:41:44
And I just wish more people would think like this because that's where you get the really great stuff, right?
1:41:51
You don't get the the oh, you know, like whenever you turn on one of the streamers, every upgrade is really a downgrade.
1:42:00
[laughter] And it's just like the in shitification of everything. Yes.
1:42:05
and and you have gone exactly the opposite direction, which is why I love the work you do and the way you think and we're so lucky to be working with you.
1:42:17
All right, we got the final question.
1:42:19
The final question is we're going to wave a wand and we're going to make you the emperor of the world. Couple of rules. You can't kill anyone.
1:42:29
You can't put anyone in a re-education camp.
1:42:31
But what you can do, we're going to give you a magical microphone.
1:42:36
We're going to enchant your mic and you can say two things into it, but it's not going to be just listeners of Infinite Loops who hear that.
1:42:45
Everyone in the world is going to hear those two things in their dreams or however you want it to be.
1:42:49
And unlike all the other times they're going to wake up whenever their next morning is and say, "You know what?
1:42:56
I just had two of the greatest ideas."
1:42:59
And unlike all the other times when I didn't do anything about them, this time I'm actually gonna act on those two things.
1:43:09
What are you going to incept into the world?
1:43:11
>> Don't be afraid of breaking things and move at the speed you're comfortable with, but don't let everyone else slow you down. >> I love both of those.
1:43:20
Those are actually quite unique.
1:43:22
You might win, [laughter] but the problem is what you're going to win is several of our books, which you already have.
1:43:29
[laughter] Misha, thank you so much for joining us. >> A pleasure. [music]