0:00
I've been thinking a lot about timing as we were preparing to have this conversation because what you guys are doing with reflection feels very apt to the moment we're in.
Search this video
0:07
I was looking and mentions of open models on earnings calls 6xed in Q3 and open model tokens are now the majority of what's flowing through Versel and open router.
0:17
So there's this real shift that's been underway and you guys have been in stealth and have big news that we're going to talk about uh on the open model front.
0:27
But first, before we get there, I'd be curious to hear you both comment on what you think is driving this moment right now in terms of the surging interests in open tokens.
0:35
There's multiple things here I think we can peel back at, but if you were to kind of look at it big picture, what do you make of the moment we're in?
0:43
Misha, we could start with you.
0:45
>> You know, we um decided to really push for building great open models at this company about a year ago.
0:56
And uh at the time I think that we just there are some things that we saw that would be kind of happening um that that have happened faster than than we expected.
1:06
And and maybe one way to think about why open models are so top of- mind and so important right now is using a real estate analogy.
1:13
Uh which is we have to remember that in the AI market uh this market commercially has only existed for maybe 3 years.
1:20
the market effectively started as a rental market.
1:25
Like closed models are the equivalent in real estate to renting an apartment.
1:28
You're renting your intelligence.
1:28
And that makes a lot of sense, especially when you're, you know, when you're earlier in your life, you're a student or you're you just graduated college, you might go and and rent an apartment.
1:39
Uh but as a consumer, you you get older, you have a family, you you know have um more, you know, more finances.
1:44
Uh ownership starts mattering a lot more in the real estate market.
1:49
And so that's why there's rental and and ownership in real estate.
1:52
AI timelines are compressed.
1:54
And so we've been in this commercial market for three years where effectively in the west there's only been a uh a rental alternative.
1:58
And now what we're seeing is that as AI adoption, AI adoption has increased and the small startups that were tiny a few years ago are now giant enterprises and large enterprises have adopted AI.
2:12
large enterprises have adopted AI. um an ownership market is basically coming in and the only way to own intelligence is by by definition if it's open because if it's open you can run it on your stuff um you can own you can control it you can customize it and uh what we're
2:27
seeing is that the reason this has happened so quickly is because AI as a market has grown so quickly and you're right I think that uh 12 months ago not that many enterprise at all were using open models and and not even startups really were using open models all that
2:44
6 months ago a lot of startups that had grown u at the application level started switching to open models and have continued growing very quickly and I think that's where maybe 6 months ago there's 30% of token uh consumption on open router over cell was going to open
3:00
models uh 70% to close and now 6 months later uh it's flipped where it's maybe more 70% on open models and 30% to close and so we've just seen a very fast movement in uh the mature the maturing of of the AI market and I think that is one thing and obviously the geopolitical thing is the other thing. Yes. Um Yes.
3:19
Um >> which we'll get into.
3:21
Giannis though I'm I'm curious from your perspective technically the state of open models.
3:27
Has has the progress surprised you?
3:27
Is it about where you expected if we were to have like looked out a year a year ago to now?
3:34
As you've been thinking about the research road map, have you been seeing this moment coming or has this surprised you as well?
3:39
this surprised you as well? No, actually I think that uh if you look back you can see that the open models have uh you know kept pace with the closed models and I think like maybe there's even um kind of like a a shorter distance between the open and the way the closed
3:55
frontier currently with like the the biggest Chinese models and I think that um this is something that will just uh uh become more uh prevalent in the future in the sense that like I don't see why the closed frontier should just like stay behind the the open frontier should play behind the closed frontier. It's a matter of just having the the
4:11
It's a matter of just having the the talent and the compute and uh uh you can definitely just like build open models that punch at the same level as closed models.
4:20
models. uh and you know one thing to say here I think like uh there are two main reasons why we also like seen this adoption of open models uh by enterprises and startups like in my mind at least like one is the fact that the capabilities are there now like you know we have like some really powerful Chinese models or like open models in
4:38
general uh which uh can really be used in in in the place of like closed models and the second thing is that uh the market has started to mature like to to Misha's point and uh what we've seen in the software market is is that like whenever something matures, there's like always a movement uh towards uh the open alternative. You know, once the the
4:57
You know, once the the market matures and the social market matures, then there are there are open alternatives and then enterprises and uh uh and companies and individuals just like move to the open alternatives and I think that like something similar will happen with the AI market as it matures further.
5:13
This episode is brought to you by Mercury, AI native banking that's loved by more than 300,000 entrepreneurs, including me. Visit mercury. com to learn more.
5:21
Mercury is a fintech, not a bank.
5:23
Check the show notes for details.
5:25
This episode is also brought to you by Jira by Atlassian, where teams and agents get the context, coordination, and control to move work forward. Try it free at jira. com. That's jir. com.
5:40
Thanks also to Granola, the AI notepad for people in back-to-back meetings.
5:43
It works everywhere you do and lets you focus on what matters. Try it at granola.
5:48
ai/sources and use the code sources for 3 months off.
5:52
You all are obviously an American company and you have your model coming out, which we're going to talk about, but I'd be curious to hear from you all.
5:59
Why is it that China and Chinese labs have just dominated on open up until this point?
6:03
I mean, Meta had its moment with Llama 3.
6:05
Um, but really the West has completely lost the frontier of open until you guys.
6:12
And I'm curious, what is it about the Chinese labs, if you guys were to armchair it, that you think has led them to catapult the way they have?
6:21
>> You know, the thing that drives technological development in the West, um, is ultimately, you know, it's capitalistically and commercially driven.
6:30
And if you're in an early market that is a rental market that primarily supports, you know, closed model development, then that's the thing that develops first.
6:37
Uh, and it's only once the market is ready for ownership that I think there's a kind of commercial and capitalistic market support for for open models.
6:47
So I think that uh for example I mean America used to have the best open models and indeed I think the highlight the peak of that was Llama 3 but there was no real commercial incentive at the time for for you know Meta to do that or for anyone else to be producing great open models in the west.
7:02
I think that uh again armchair this that China has a different mindset that is more geopolitically motivated.
7:10
Uh and that uh I think that there was a bit of a uh sort of fluke moment like an unpredict a moment that was hard to predict which is when Deep Seek V3 came out and was so good.
7:25
Uh and then I and made Chinese intelligence abroad so relevant in a way that it hadn't been before.
7:31
And I think that there was a concentrated push to support uh an open-source ecosystem primarily driven for geopolitical reasons.
7:40
And so these labs in China were not actually making that much money at all until recently until their intelligence got good enough um and the market uh matured that there's a commercial market for open intelligence and now they're starting to make revenues.
7:53
So I think that uh my my view is that uh what happened there was uh primarily geopolitically motivated in trying to kickstart that ecosystem.
8:03
>> Well I think that brings us to what you guys are doing.
8:05
You guys are talking about beam your first model which I think everyone in the AI community has been buzzing about what is reflection doing when are when are they shipping and the moments coming now.
8:15
So, you've got beam and it's around a 500 billion parameter model on the open benchmarks.
8:22
It seems state-of-the-art.
8:22
Um, but I'd be curious to hear from you all and as we're talking, you're still kind of going through the final phases of validation and all this, but I'd be curious to hear from you all like what you expect the reaction to this model to be.
8:36
>> So, Beam is uh yeah, it's our first model as our first uh openweight model.
8:40
Um, we're we're very very excited to to release it.
8:44
I think it's uh you know bu building a company is I think there's a Reed Hoffman quote that it's a building a plane while you're flying it and uh the plane's been built.
8:53
So I think we're uh we're very excited about that.
8:57
>> Uh it is indeed a around 500B total uh parameter model 23B active uh and it advances uh the frontier of western open models in a meaningful way which we're excited about.
9:09
Uh the things that we think we we're proud of uh with this model is uh it's particularly strong in coding and agentic benchmarks.
9:18
Um we believe that the form factor and the strength in these uh capabilities make it a really strong workhorse model for enterprises public sector uh right it has this right kind of utility but also efficiency around it that I think will be appreciated.
9:35
The kind of other thing that we're really excited about is that um it's a model that is uh leading in its reasoning efficiency relative to uh to other open models.
9:46
And when we say reasoning efficiency, we mean how fast does it solve a task relative to the compute it has at inference time or relative to the tokens that it has at inference time.
9:58
Uh and you know, Giannis and team did some incredible work to to get that state.
10:03
Maybe you wanted to share that.
10:04
share that. Um the the form factor for the model is a 500 billion parameter model is uh 23 uh billion active and uh if you actually like think of what you want from a model that's an open model and you want to ensure that it has like wide adoption especially from like
10:20
enterprises you need something that it is efficient it is compact and it actually like uh really produces the most intelligence that you can get out of your flops and in order to do that you need to have like a reasoner like these are reasoning models and the idea is to just like use tokens to actually solve tasks. The way that you measure uh
10:37
The way that you measure uh whether a model is like more efficient or like equally efficient to another model in terms of like reasoning efficiency is by just like seeing how many tokens or like how much time how much compute it actually takes to solve um uh a task of like you know similar difficulty.
10:53
And uh what we've actually seen is that we have a model we've built a model that uh can be three or four times more uh token efficient uh reasoning efficient than some of the open frontier models like GLM 5.
11:05
3 or like some of the uh best closed uh uh closed models like GPT6 uh Luna.
11:10
So the question then is like how can you ensure that you can extract the most reasoning and the most uh intelligence out of like uh this small model and the way to do it is by really scaling reinforcement learning.
11:25
So reinforcement learning is something that uh you know I've been always like really passionate about.
11:30
It's like you know my I started like doing reinforcement learning back in the back in in the early days of reinforcement learning in 2012 2013 and uh you know it's something that like both Misha and I have always like been really uh true believers in reinforcement learning we've been RLP
11:46
and uh you can actually see now that by scaling reinforcement learning and what we've actually done at reflection is that we scale reinforcement learning uh uh to to to the highest levels of compute in terms of like what we know publicly available in comparison to other like open labs. We've actually
12:01
We've actually like you know released their the research and we can just like compare apples to apples.
12:06
So we've really kind of like doubled down on the compute we put into reinforce learning with the goal of ensuring that we extract the most juice the most intelligence out of the model.
12:16
And this is uh ultimately what like enterprises and uh you know developers and everyone cares about like a compact model that uh uses uh its tokens efficiently and uh intelligently to solve tasks and the enterprise use cases seem obvious when you describe it that way.
12:32
Do you imagine broader consumer I mean these are all kind of nebulous terms and enterprise adopting it can translate into consumer adoption based on how that company's relationship with customers you know functions but do do you see consumer applications of a lightweight efficient reasoning model like this?
12:50
Definitely though the way we kind of think about it is that the purpose of an open model is to enable an ecosystem of builders to build on top of it.
13:00
And so right it's it enables uh other companies that might be building uh consumer applications to control their own destiny um and have an efficient workhorse that they can also customize against their data.
13:10
Uh and you know we're starting to see this new wave of new consumer experiences powered by Agentic models.
13:17
And so agentic capabilities are very important for uh doing you know certain you know they seem like basic stacks for a consumer but they're actually hard for >> AI cost they cost a lot of tokens. Yes >> that's right.
13:28
>> that's right. So if you have um a workhorse model that is it's sort of doubly efficient and that the size factor is efficient but then because it's also very efficient in its reasoning right if it's say 4x more efficient in reasoning than other models and it has a smaller form factor then the those efficiency gains translate to
13:48
faster response times to consumers uh much lower cost of serving and honestly for consumer use cases specifically the cost of serving matters a uh and and and so I'm really excited to see what others build on top of these uh open models and how can we empower both enterprises and companies building consumer applications >> and you all train this from scratch. You
14:09
You didn't do distillation or anything like that.
14:13
Talk to me about that decision to do it from scratch because in open I think everyone just kind of assumes these days that everyone's distilling everything even the closed you know players there's like rumors of distillation happening and accusations.
14:26
So, so why why approach it that way? Yeah.
14:28
So, this a good question.
14:28
I think that like uh there are two reasons.
14:31
First of all, we wanted to ensure that like this is you know a western frontier open model that's actually like a fully uh built uh you know you know by our team in the United States.
14:40
We also have like a team in the UK but like built in in in in in our offices in the west.
14:45
in in in in our offices in the west. And the second thing is that especially when when you want to really scale reinforcement learning it is extremely important that you also like control control your pre-training like uh uh pre-training you know now it's like more of a better understood science than
15:01
reinforcement learning but like in order to do it in the right way so that you can scale reinforcement learning keep it extremely stable ensure that like this is a mixture of expert model but like ensure that like the the the the router and the experts remain balanced and stable during this like high compute reinforcement learning regime. It's
15:17
It's really important that you fully own it.
15:21
There's a lot of work that went into uh the data that we put in our pre-training to ensure that like it uh seeds the right reasoning capabilities to the model that like then reinforce learning picks up and just really uh accuates and uh also on the architecture and the learning dynamics and the optimization so that like once you've actually trained it, it's uh really stable.
15:40
It's uh it has the right biases and priors so it can really like pick up with reinforcement learning and only by just really optimizing the system end to end by innovating both on the pre-training and of course on the reinforcement learning side it is possible to just like build something that again is like really capable for its model size and it really packs a lot of intelligence in a in a in a smaller form factor.
16:04
>> I'm curious how you all are thinking about safety and alignment.
16:06
It's the maybe number one topic not just in AI but like in geopolitics right now is aligning these models.
16:13
There's all the headlines recently of um you know open AI and others having rogue agents going out there and hacking things and I think people are concerned about the increasing capabilities of these models.
16:24
Um how do you approach alignment especially for an open weight model?
16:29
>> You know there are kind of two things to kind of think about.
16:31
One is what are the benefits of um doing safety research and and uh and protecting you know models uh in the open.
16:42
Um and then the technical of how do you you know actually go and you and and uh you know bake in safety into a model as it pertains to kind of um open source uh traditionally.
16:53
And let's not even think about models for now because there's a very similar thing that happened in software development.
17:03
Um and you know with operating systems and internet the internet protocols are open operating systems like Linux and Android are open and a lot of the strong encryption protocols that gave rise to the whole field of cyber security are open like there's the way right you test software you do penetration testing there's both black and white you know box uh penetration testing and so the
17:24
notion of openness has been fundamental fundamental to securing software um over the last decades to the point right there's Linus' law which is uh that with enough eyeballs all bugs become shallow and that's why open source software is actually in many ways considered safer and more trustworthy than closed software uh because you can just iron out right all all the kinks. Now we have
17:46
out right all all the kinks. Now we have to think about artificial intelligence as right it's an evolution of software it's it's a machine it's a piece of software >> so it's not a god >> um not well uh [laughter] not that I'm aware of >> okay good >> uh I mean and you know it's no more than uh you know than Linux or Android is a god so maybe there's a church >> it's a very pragmatic view
18:09
>> yeah maybe there's a church of Linux somewhere and uh all the power to >> we'll find it'll be in the comments >> but it's ultimately it's it's a very sophisticated machine that has a long tail of vulnerabilities like any piece of software and it is impossible to discover and patch up that long tail uh with high confidence when you have a small number of people who are have access to it. So right across the closed
18:33
So right across the closed labs there might be 300 safety researchers and uh that's great and I think the intentions are great but it's kind of like saying um across my whole body I have 300 white blood cells.
18:44
Do I trust those 300 white blood cells to, you know, find all the pathogens and get rid of, you know, the stuff that's, you know, attacking, you know, your body every day? Um, no.
18:53
I'd rather have as many white blood cells as possible.
18:55
And I think that, you know, this notion of with enough eyeballs, right, all bugs become shallow.
19:01
Um, I think it's it's true to like if we think about mechanistically patching like all the vulnerabilities in these systems, of which there are many.
19:10
Um, you need to bring a community in, right?
19:12
bring a community in, right? you need to bring um you know from 300 300,000 right people who can in investigate inspect these systems and so fundamentally I think that in the same way that open source software uh has actually been the safest way to develop software broadly previously um I think the same thing transfers here there are some things that are new because the software is so powerful uh but the broad strokes are
19:38
still the same and so at a high level I I'm I'm excited for moving into a world where uh a large open source community protects us you know in a way that I think um is going to be much more effective than you know having a very small handful of people who are protecting >> as you're saying that I'm thinking about you know hugging face when it was getting hacked by open AI what did they do they turned to open models Chinese
20:03
models because they they needed to defend and they didn't have access to the best frontier closed cyber >> thing right >> um so that it speaks to what you're saying at the same time I mean I think there is a a fear that or and an under a
20:16
belief in a lot of the closed labs at least that like this is different than Linux and maybe there should be a point at which this is held back I mean Yiannis maybe you can speak to this from the research perspective is there
20:28
something on I assume you guys are training your next model already is there something on like larger trains where you would go we would hold this back even though it's open we're an open company this is too capable do do you
20:39
all think that way at all it sounds like no >> you know I think like uh it is possible that like it gets to a level of capability that you want to just like be uh more um uh careful with like how you deploy it. I think fundamentally sharing
20:50
I think fundamentally sharing it with the research community at large has always been um a net positive for like the development of AI and we actually saw that with uh safety in particular like uh once like powerful open models were available to the researchers actually like so many insights and so many new safeguards were actually developed.
21:10
So I think that like fundamentally it isn't that positive.
21:11
I think like maybe you want to just be more careful on like uh who has like access to to the model like you might want to just like restrict it more but uh fundamentally sharing it with the research community and uh with uh you know the vast majority of developers out there who uh have like good intentions and actually want to help uh us and the world to just like make this these systems more u uh safe.
21:34
It's a it's a net positive and it's the right approach.
21:39
approach. Now this is actually true that uh uh we see that like models get uh more and more powerful uh you know they they hack their sandboxes they they they they form communities online >> they collude with each other I mean >> yeah they collude with each other with
21:55
each other exactly and uh this is uh something that uh uh you know we see more and more at the same time they also like become better at like policing themselves or like you can actually use AI uh to uh to control AI and to just like uh make it safer. You know, there's
22:09
You know, there's like a research that like goes into that.
22:13
Uh and you know, fundamentally, I think that it is important that uh we share many of those findings.
22:18
I think the the fact that uh even the closed labs went ahead and just like released the findings of uh how these different agents uh uh operated uh can only provide us more insight and more information so that we can protect ourselves and we develop the necessary defense mechanisms.
22:36
So this is to say that like in general being open, being transparent uh makes uh things safer and makes uh the our society better prepared for like uh the advantage of AI.
22:50
>> But is there is there a capability threshold you would hit where you'd go we can't put these weights out?
22:54
Like is have you have you guys thought about that?
22:57
that? uh we we've definitely thought about that and also obviously there's a lot of discussions now in DC around that notion and that is very important and independent whether a model is open or closed right if you hit a certain capability uh then you know we should as
23:15
a right as a community of both you know industry and government right there are policies around right what should be kind of released and not and under what you know gated measures you you'd allow that to happen uh So that's really around kind of a capability of intelligence. Um I wouldn't describe us
23:30
intelligence. Um I wouldn't describe us as you know hardliners of uh you know close is not safe or open you know it's really around an ecosystem is safe like having an ecosystem is safe and having um you know closed models out there u
23:47
with their points of view open models with their points of view and this and that right that is the the safer world and of course there's a threshold of capability um under which these things um need to be regulated uh but that is independent of whether they are open or not. And I think that there when it
24:03
not. And I think that there when it comes to openness specifically there are arguments uh that well look you can release an open model and maybe we need to like that some bad actor right could use it to do this and that um which by
24:16
the way that's exactly what people do with software and why there's cyber security right so uh but there's this other thing that I don't think uh people talk about as often which is there's the bad actor issue and then there's the unintended consequences issue and what
24:32
happened with the open AI hug face hack that you um just described is that was an unintended consequences issue that again because you have to like when we do these runs right um these large RL runs there can be like a billion agents over the life of a run right doing stuff
24:50
and it is impossible to imagine that you know you're like 100 researchers 300 researchers or whatever will be able to you know to really track each of those agents right so there will be when when there are a small number of blood cells right there. Um there's a large surface
25:03
Um there's a large surface area of things to protect against and there are unintended consequences.
25:06
And so what happened there was that um right these agents hacked out.
25:11
I think [snorts] Hugging Face tried to use closed models to protect them but because the safety guard rails prevented them from being used for cyber they were not able to protect themselves.
25:20
So they had to turn to an open model.
25:21
And that actually shows you an ecosystem working because okay there's some you know there's some principles on one side that say no we can't use it for that.
25:30
Then principles on the other side is like no you can you can use it for it and ultimately because the ecosystem existed that enabled hugging face to protect itself and so it's just these this technology has a lot of unintended consequences and the only way to really deal with it safely is to have um a community and an ecosystem around it.
25:50
>> I totally get that ecosystem approach makes sense.
25:52
I'm curious though technically how do you approach this technically does that differ when you're building an open model how you do safety alignment at all or is it the same?
26:02
>> Uh so actually like in terms of the technology that we're building it's quite similar to the technology and the recipes that are used in the closed labs.
26:08
The way that you actually do it is that uh there are two parts to it.
26:10
Like you need to first ensure that you have like all the right evaluations across like all the different risk profiles so that you can measure uh the uh the the the performance for the model also the response of the model and ensure that like uh it's not either a model that just like refuses everything because like a model that refuses to do anything is like super safe.
26:30
It's the safest model but it's not useful.
26:32
So there's like a you have like these two things this uh on one hand you have usefulness on the other hand you have like safety and it's like a trade-off that you need to ensure that like you're measuring it right by having the right evaluations and you also like hill climbing it in the right way.
26:46
Uh once you have your evaluations and you've actually collected the data and you've defined what is the behavior that you that you want your model to to to have and the way to do that is by collecting uh examples of like you know what good behavior and bad behavior looks like.
27:02
then you can actually apply the usual methods of uh you know imitation learning or SFT as it's called all reinforcement learning to ensure that the model always behaves uh uh is aligned and behaves in the way that you want it to behave.
27:15
Um this is uh you know a field of study that like uh you know has it has been developed in the past 5 years.
27:22
uh these are like methods that uh you know they've actually been developed in the in the closed labs and you know in in in the open research community and uh we're also like really putting a lot of effort.
27:32
We have a uh a team of like researchers who are only focusing on like safety and ensuring that like the models that we put out there are like uh really evaluated correctly and they are like safe and uh they really at the at the at the at the frontier of like uh safety and uh usefulness.
27:49
Even if you buy the argument that model capability will create intelligence, we cannot fathom.
27:55
We cannot comprehend what it will do in the world if it's fully diffused.
27:59
And therefore, it's not fair to compare to an operating system.
28:03
It's actually something much more almost like a a being in the sense of it's we don't know how to to control it.
28:09
We don't know its parameters.
28:11
I feel like that's starting to seep into the consciousness of of the public with how some AI leaders talk about things.
28:17
And I'm curious where you guys sit there.
28:20
>> I think that like the other thing that um it is important and it goes back to the diffusion of power that like Misha was like talking about is that uh every time you have like only one or two institutions controlling something that's like so powerful even if they have the best intentions it's like a single node of uh failure.
28:35
single node of uh failure. So you want to ensure that you have like a redundancy even in the system to ensure that like it's not you know if if this system just like fails then all systems have failed and the hagging face example is a prime example of that that if you
28:51
could actually use AI to protect against like a bad behaving AI then the world ends up being safer and if um other people are actually like enabled and other institutions are enabled to use like powerful models for cyber security and for protecting I think um yeah for protecting everyone uh then that's a net positive. >> One of the biggest pains I've had
29:13
>> One of the biggest pains I've had running sources has been managing my books.
29:16
I'm not an accountant and all of the software out there is unintuitive, clunky, and detached from where my transactions actually live until now.
29:25
Mercury recently launched Mercury Books.
29:27
It's AI powered accounting that works with your Mercury banking and credit card transactions, plus external cards and payroll systems as well.
29:33
You can choose between cash and acrruel accounting and quickly generate reports for your cash flow, P&L, and balance sheet.
29:41
Like everything Mercury does, the design of Mercury Books is super approachable and clean.
29:44
I finally don't feel overwhelmed when I'm trying to understand the big picture of my business.
29:50
I love that AI does most of the heavy lifting, including for things like autocategorizing transactions and that Mercury's Command AI agent can handle tasks end to end.
29:58
You don't need to be paying for separate accounting software anymore.
30:02
And you can even invite your accountant or CPA to work alongside your Mercury books with you.
30:08
This doesn't cost extra, and Mercury will even help you find a human accountant if you need one. Visit mercury.
30:13
com/books to learn more and apply online in minutes.
30:17
Mercury is a fintech company, not an FDIC insured bank.
30:19
Banking services provided through Choice Financial Group and column NA members FDIC.
30:24
AI is only as useful as the context it has.
30:26
But when that context is scattered across tools, threads, and DMs, your team and your AI agents are flying blind.
30:33
That's the problem Jira by Atlassian solves.
30:35
What's the goal tied to your project?
30:38
What got decided last week in Slack DMs?
30:40
Atlassian's teamwork graph pulls all of the valuable pieces together from Jira, Confluence, GitHub, Slack, and more.
30:47
So nothing falls through the cracks.
30:48
You get 44% more accurate results with 48% less token usage.
30:53
With Jira, you can easily share your work context with the AI agents you already love, like Claude, Cursor, and GitHub Copilot.
31:00
Assign them work directly or connect your tools through MCP.
31:04
All of this lets you spend less time digging through endless links and messages, chasing down what got decided and by who, and spend more time actually shipping. Learn more at jira. com. That's jir. com.
31:17
I spend a lot of time context switching between meetings, often with no time to process one before the next starts.
31:23
Thankfully, Granola runs in the background the whole time.
31:25
It's an easy to use AI notepad for meetings that works everywhere, even on phone calls.
31:29
I use Granola to recall what was said in meetings and create helpful summaries.
31:33
I use it every day to stay on top of what I need to get done with my team.
31:37
It connects to my email and suggests follow-ups for me to quickly review and send, saving me valuable time.
31:42
Granola isn't just a core part of my workflow.
31:46
It's basically my second brain. Try Granola at granola.
31:48
ai/sources and use the promo code sources for 3 months off.
31:54
Framer is the AI website builder that powers the sources podcast website at podcast. sources. news.
32:01
Framer brings AI agents into the same canvas where your website is designed, managed, and published so you can move faster without giving up your taste or control.
32:09
I use Framer to make the sources podcast website be the destination for everywhere you can find the show.
32:15
Plus, you can also see recent issues of my newsletter.
32:17
Start building with agents for free today at framer.
32:20
com/sources for 30% off a framer pro annual plan. framer. com/sources.
32:27
Rules and restrictions may apply.
32:29
>> I have other questions about the business.
32:30
But before we get into that, um, for people who don't know, what's the origin story of Reflection?
32:34
You guys work together at Deep Mind.
32:36
I actually heard that you reported to Giannis at one point.
32:41
>> Uh, yeah, Giannis was uh, basically my boss.
32:43
>> And now you're the CEO.
32:43
So that's that's a fun twist.
32:45
>> Well, we're actually equal partners like the way >> that's that was going to be my question.
32:49
How do you run the company?
32:49
You started it together leaving Deep Mind.
32:51
Um maybe there first like you bet on Open in what this 2024. Is that right?
33:00
>> Well, I guess it was a year ago.
33:00
So, you know, we we well actually when we started the company, we had two two bets.
33:06
And I would kind of bring it back to to the origin story.
33:08
At the time, Giannis and I had worked on Gemini 1 1. 5.
33:12
Uh Giannis was the RHF lead um and you know one of the most storied reinforcement learning researchers at DeepMind.
33:19
Um I actually switched from physics to uh AI uh when I saw Alph Go come out and then I later learned you know well I I knew that Giannis was was an author but um that was kind of a you know sort of a dare I say it a celebrity crush moment [laughter] uh when when I joined DeepMind and and had the pleasure to work under Giannis on his his RL team.
33:39
But at the time we shipped you know Gemini 1 1.
33:41
5 and uh those models were really good at chat.
33:47
You have to remember when chat GPD came out these models were not good at coding.
33:50
They were not good at egentic.
33:52
Uh they were really good at chat.
33:52
And we started the company with kind of two bets in mind.
33:57
A scientific bet and kind of a market bet.
33:59
One was that uh there would be that that reinforcement learning uh was the way to get these models to have agentic encoding capabilities.
34:09
Uh and the other one was that there'd be powerful western open models we can build on top of so that we can just do this research a lot more cost effectively.
34:18
Uh and the reinforcement learning within the first kind of year of the company uh became right across the industry.
34:25
right across the industry. um we were able to uh show some you know really strong reinforcement learning results at small scale uh and obviously the reasoning models started coming out uh at that time from other labs but at the same time the the western kind of open
34:41
frontier receded uh with the DeepC V3 moment and uh we actually waited around for a few months for something to come out that was as good and then uh saw that wasn't happening and so that's kind of how the company changed course but there I think even a better uh you know a better story in terms of how we actually uh started it. And when we were
34:58
And when we were building Gemini, there would be these sprints where uh teams from different offices would come together and I was in the New York office, Jiannis was in the London office and kind of an SF New York and London team descended on New York for, you know, one of these sprints.
35:13
And uh at the time, you know, it turned out that Giannis and I were thinking these ideas kind of independently.
35:19
I'd wanted to broach a subject with Yiannis, but he's an intense guy and like it's uh you know, like telling your boss like, "Hey, do you want to [laughter] like start something?"
35:28
uh was uh you know was a bit intimidating.
35:30
So there was an evening when we you know Giannis came to New York and I was kind of uh uh showing people around, right?
35:37
Uh and so we uh went to took a group to the comedy seller and uh it was like an outlier runchy set.
35:44
It was the runchiest set I've ever been to.
35:46
So it was a little awkward.
35:48
>> I've been there a lot.
35:48
I've seen some runchy sets. Yeah.
35:49
runchy sets. Yeah. Is this like uh you know do people find us funny or am I uh you know we went to that and then uh we took a long walk um and went and grabbed this dinner on um there's this kind of boat in in New York by uh you
36:04
know by the Google office that uh serves lobster rolls and uh actually Giannis broached the subject uh of hey you know maybe we we start something in this shape uh and honestly that for me at that time it was really you know whatever it is if it's kind of with Giannis were starting it. Um, but we
36:20
Um, but we ended up, you know, we're both reinforcement learning pled, both AGI pled.
36:26
I think had a very um, similar view of where we thought the world was going and uh, and and then decided to start something together.
36:33
I >> I heard a rumor, I don't know if it's true. Was it Jensen?
36:35
Nvidia is one of your large investors.
36:37
Was it Jensen who pushed you guys to go the open route when you were focusing on coding?
36:41
Is that is that accurate?
36:44
>> We uh, we came to that conclusion independently. Got it.
36:45
independently. Got it. uh and uh >> but he's I mean he's been a huge champion of open and I mean it's very clear that Nvidia sees that as strategically just important not just for them but the entire ecosystem and they're a large partner of yours right >> yeah I mean they're uh Nvidia is um a big investor across the whole ecosystem
37:02
because indeed right the ecosystem um is important and obviously uh open models are kind of um a core part of that that is an an enabler for an ecosystem so I think there's a lot of um vision alignment there um and obviously we're uh very grateful for uh for their support, but they're an extremely prolific investor across the whole industry. >> And you guys have raised over two
37:23
>> And you guys have raised over two billion, two and a half billion.
37:25
You have a lot of great investors.
37:27
I think the last known valuation was around 25 billion.
37:31
And and that brings me to my question about the business.
37:33
How do you make money on open models?
37:36
I think we were talking about that the Chinese labs are starting to but the margin is all at the frontier closed right now it seems and you guys have also been striking some pretty big compute deals.
37:47
Uh there's the 1 billion with Nebus, the over 6 billion with SpaceX.
37:51
So it implies you see the revenue ramp and you have an understanding of what you're going to do and I'd love to hear how you all will approach this.
37:59
>> Yeah, I'll I'll share some thoughts and obviously would love to hear Giannis's uh as well.
38:04
uh as well. uh at a high level when we think about open-source open models or any anything else that is open source right open source technology is a demand driver for something right and an example is Kubernetes demand driver for uh CPU clouds effectively and
38:27
open models right are really like when we think about where the market is going is that uh the largest buyers of open intelligence Well, they're definitely going to be well this the scaled up startups that became enterprises, large enterprises, public sector sovereign entities and allied countries. And the
38:42
And the thing is if you just give them an open model they don't know what to do with it because uh it turns out that when you're buying a token from let's say a closed model lab uh they've you know they've packaged up and abstracted a whole host of technology that went into that which is the cluster management software the serving software the inference software kind of application layer the whole system around it.
39:05
So what we see open models as is that it's a demand driver for that.
39:10
So when you go to an enterprise and you're their trusted open intelligence provider, well you need to serve right inference into them, right?
39:19
You need to help them set up agents and you set up that infrastructure for them and and serve that for them, right?
39:23
And so the business is really uh when when you sell a token for closab sells a token, they're not just selling a model, they're selling everything else around it.
39:34
>> And that's what you go and do.
39:34
>> And that's what you go and do. you go and help enterprises set up right in a package turnkey way all the other stuff around that they need to make the the open models useful and we see open models as trust and uh demand drivers um for you know that full stack
39:50
>> do you worry at all about the scalability of that model or in terms of how how highouch that is to do that forward deployed work >> I think like it's not just the forward deployed work I mean this is part of it but like uh if you actually take a step back and you just kind of like think of what the future will look like. I think
40:06
I think that uh people and especially like big enterprises uh allied nations like everyone realizes that uh AI is an extremely important technology that cannot be outsourced to a third party.
40:19
It needs to just be um controlled and owned which uh means also the data, it means the infrastructure, it means like the deployment of it and uh owning your own intelligence is actually like something that uh uh everyone realizes that's like extremely important.
40:31
you just like see that there's a lot of movement in terms of capital, in terms of investment, in terms of capex in in this direction.
40:39
Uh and this something that like we also observe.
40:41
that like we also observe. So what we're becoming and what uh what what what we're building is the um the company that's the intelligence providers for anyone who wants to own their own fate and own their own intelligence and uh that includes the software includes the
40:55
inference it includes the um applications that are actually like used internally and uh I think this is going to be like uh a significant part of the market because um the way that I always think about it is that uh AI in in a way it's like a form of labor. I feel like
41:09
I feel like this is this is why it's priced the way it's priced.
41:12
And if you think of uh who are the main consumers of labor in the market, it's actually like uh big enterprises.
41:18
It's uh it's you know government.
41:21
It's like big institutions.
41:23
It's exactly the institutions that want to consume their own intelligence.
41:26
And this is why um you know serving these particular customers is uh I think that like in the in the medium term even it's actually like the biggest chunk of like the AI uh economy.
41:36
the AI uh economy. Yeah, I think there's there's maybe a somewhat of a misconception that the this kind of business is somehow you know like a forward deployed you know engineer services only kind of business really it's like the what you're doing is
41:52
you're unlocking inference right and so just like closed labs are unlocking inference they're you know the open labs right in in in China have APIs are unlocking inference so you're you're serving inference right and the estimates are that right like there'll
42:08
tens of gigawatts like of um open model inference served uh right within the next few years which is which is a lot right like uh and and the way you think about everything else is how do you kind of uh remove frictions to unlocking more
42:24
inference right and what you do is you help enterprises and you help sovereigns and public sector agencies uh serve their own inference and unlock it at scale >> does this imply that you all need to be a fully vertically integrated open hyperscaler yourselves. You've got these
42:39
You've got these compute deals I talked about that your partners you're working with you Elon and Nebius but do you need to have your own data centers?
42:47
Well, I I think that the data center stuff, you know, in some kind of um asmtotic end state that probably makes sense, right?
42:53
Because what ownership allows you to do is that it well, it basically increases your margin because you can kind of think about any any part of the AI stack.
43:02
You'll go to an investor or something or or you know some someone uh and they'll say, "Oh yeah, that thing's getting commoditized, right?
43:08
Open models are commoditizing the closed models and all this competition at the application layer that's commoditized because the model is where it's at.
43:14
model is where it's at. Oh, and the and the bare metal that's also commoditized because you know there right everything is sort of uh commoditized is is sort of the argument and so I think that all of that is actually true everything over time right gets somewhat commoditized to
43:28
some kind of appropriate margin and the companies that are vertically integrated uh that accumulate margin across the stack um right will be the most durable ones but the question is uh when do you do what right and I think that Uh fundamentally we're like the market is early. Uh you're in the game of um
43:47
Uh you're in the game of um getting great open models out there and unlocking as much inference for them as possible and then optimizing your margin.
43:57
>> Yeah, because I mean Alibaba they can give away Quinn and make money on their cloud.
44:01
And I assume there's a situation in the future where you all could do something similar.
44:05
something similar. I mean because right now if if someone runs a reflection model and you're not helping them with the things you're talking about helping large companies and you aren't you all aren't making anything from that and I assume that would you imagine a future I guess where you're all's models are not distributed by all of the hyperscalers it's really like just the reflection compute that serves
44:29
the reflection models is that what you're describing >> yeah there I have kind of two uh views on in this one is that I think we keep underestimating how big these markets are and how first multi-polarity is a good thing that you know customers have choice and these markets are so big that they'll support multiple large companies and then the revenue potential of a company it follows a pretty simple equation. It's uh intelligence density
44:53
It's uh intelligence density times the amount of compute that you have times the trust that you have with an enterprise to work with you.
45:02
>> That last one's a little fuzzy. The the trust you have. How do you measure that?
45:05
That one is actually it's it's fuzzy but it's it's really important right and enterprises make decisions both on base capability and then how much do they trust you right and uh something that we've found uh is that building the models builds a lot of trust right we've
45:23
been able to uh make substantial progress on a number of you know commercial efforts before the model has been released because of the trust that building the model uh right and showing some of the milestones along the way has um has built up. So I think that it's a
45:37
um has built up. So I think that it's a it's a fuzzy thing but ultimately when you go and like uh that's been you know I'm a obviously former researcher uh and looking at enterprise buying decisions it's like yes kind of capability but it's also do I trust this team to help me in my intelligence journey to deploy it >> uh and it's a and it's a big differentiator because um you know we have to remember that
46:02
>> enterprises public sector sovereign entities they're not as technically you know sophisticated as you know, some of the companies in Silicon Valley and they're looking for not just, oh, who's my inference provider, they're looking
46:14
who's who's my intelligence partner and when you're going talking to a CEO of one of these companies and they're looking for an intelligence partner, uh the the trust is actually really important. So, uh I really think it's
46:24
So, uh I really think it's intelligence density times the amount of compute that you have, how much compute can you get under your umbrella times how much trust you can build.
46:32
Uh and that there will be multiple winners.
46:35
Is there any part of this growing um fear and push back from the administration even on companies of like you shouldn't be using these open Chinese models.
46:42
We want you using domestic AI, you know, US AI.
46:47
Is there any part of that that's like overblown?
46:49
Like is it really how important is it that a company is using a an open model made in the west versus an open model made in China?
46:57
an open model made in China? I think that um if you think of like what open software is and we saw that with uh you know the internet era uh it's actually the infrastructure upon everything was built right like the internet was actually built by open software that was developed here in the US and that actually like affected the position of
47:17
like the US in the world and uh uh kind of like how the internet was built and I think like people uh you know who we should should not underestimate the importance of just becoming the default infra infrastructure provider and AI and open models will actually be this like infrastructure layer upon which the new economy will be built. So people should
47:35
So people should not like underestimate the importance of actually owning this layer and this is something that uh you know the the US and the west in general I think like they've actually started to realize uh uh recently and so it is important that like we own our infrastructure because this is the infrastructure upon which the economy will just like sit.
47:51
So it's like really important it's actually owned.
47:54
At the same time you know as these models become more and more powerful you know there are like um there are security risks right like um associated with that uh the agents are getting to the point where um and you know maybe that was like more of a far-fetched sci scenario a couple of years ago but like now it isn't.
48:10
You can have an agent that you can inject it with a prompt and then it just like goes off and it does something that uh it is harmful.
48:18
Now I'm not saying that like they're doing that.
48:20
I'm saying that like it is possible that this happens and if you don't uh uh if you don't use models that uh you know you can inspect and you can ensure that like you know there's a a company that you know you trust to kind of like provide you with this intelligence then uh you're just like introducing unnecessary risk.
48:34
So it's both the the fact that you want to ensure that the infrastructure of the future is actually owned and uh secondly to ensure that like there are minimal security risks.
48:44
We've been talking about, you know, bigger picture, long-term stuff for you guys, but in the short term, you have had some commercial partnerships.
48:52
You've already announced you've got the work you're doing in South Korea, which I'm interested in also as a model for how do you work with other countries, and then also um the Dell partnership.
49:00
I'd be curious to touch on both those and then can you give us a sense of what's coming on the how how do you get these models out to the world beyond just putting on hugging face or something?
49:11
>> You know, there are kind of two things here.
49:13
One thing is ecosystem building and the other is you know the these sort of commercial constructs on getting the models out into the world.
49:18
Uh it's actually it's really important that the ecosystem doesn't build itself like you you need to be an active participant in building that ecosystem and working with partners to ensure models are natively integrated natively supported uh across the whole stack from you know some of the neoclouds that exist today to um you know other inference providers to hyperscalers to the Dells of the world.
49:46
uh and we have a whole open source team that whose job is we mean one build open- source software around the model.
49:53
So it's not just the model the ecosystem is the software that you build around it and we'll have uh beyond the model releases of uh open source tools that are meant to be coupled with the model um that help the community better understand uh well better understand and build safety guardrails um agents and so forth.
50:10
And then obviously there's an arm of the open source team around working closely with the ecosystem to to integrate things.
50:17
So what ecosystems don't get magically built.
50:19
You have to put uh put effort into it.
50:20
Uh I think that on the kind of on the commercial side, the way we think about these partnerships is this.
50:27
This again going back to the notion that uh there are certain things that enterprises, sovereigns, public sector entities are good at and there are certain things where they need partnership and uh typically you know a place where there's some expertise among these as well finding you know land power shell like these are real bottlenecks these are limiting factors to deployment of compute.
50:50
compute. uh right if you have a rack sitting somewhere that is not powered it's not you can't monetize it it's not useful and so uh partnerships like this allow you to basically get uh right power uh land shell um and even stand up the infrastructure right to to bare
51:08
metal so right Dell is one of the uh one of the uh best infrastructure providers in the world um then there's a question of like well what happens from there right so let's say you figured out and this is right In Korea, it's this 250 megawatt uh data center. So, you figured
51:22
So, you figured out how to set up the bare metal.
51:24
Uh well, what happens from there?
51:26
Today, the market in compute uh is largely in offtake.
51:32
So, it's I take that bare metal and I sell it to hyperscaler.
51:34
That market is moving very quickly into inference.
51:38
Like I'm no longer selling offtake.
51:41
I'm selling inference into other enterprises into my own enterprise.
51:44
enterprise. And at that point um well I need an open model because uh I can't get my hands on a closed one to run on my own stuff right maybe a hyperscaler can through some partnership or a closed model itself can by nature of you owning your own asset you need an open model to serve on it or right an ecosystem of open models and you need a partner that will set up the stack for you because you don't have that expertise uh that
52:10
will enable you to do that right and so that's where we come in at the software layer above the bare metal uh right with the model and the whole stack around it and and and maybe I actually think Yiannis had a great analogy to it and so I'll let you speak
52:22
to it but uh it's it's a you can kind of think about um frontier model builders um as kind of um right they you have to build all this software at scale for serving models in order to train models in the first place. So there's a really
52:35
So there's a really nice analogy that Giannis has to how right AWS came from dog fooding uh right web services at scale and how that translates to AI. Maybe you want to speak. >> Yeah.
52:47
It's uh the idea that you know the the clouds you have like three main hyperscalers.
52:52
You have like GCP, you have AWS, you have like Azure and the way they actually like started is because they want to solve like uh their internal problems, right?
52:59
Uh so AWS had to support the Amazon business.
53:02
So they required like these servers uh uh you know GCP was like built because like Google had like this massive search engine and had just uh really served like billions of customers.
53:12
So they had just like solve an internal problem and that helped them just develop the really dog food.
53:18
You know the dog fooding is the idea of like using your eating your own kind of like product or like using your own products.
53:23
So they had like this internal demand that kind of like led to the development of this technology and then realized that like this technology can be like sold more widely and it actually gave rise to the era of internet and like all the SAS companies that were actually built on over the cloud.
53:38
So I think like we will be entering like a a similar era where um you know you need to to to to have like the uh insight of like building a frontier model and just like really be at the frontier of like intelligence uh in order to ensure that like you design the system in a way that it's uh scalable and it's uh the right solution for the right problem. >> Yeah.
53:59
So to kind of Jiannis's point what that means is that by training these models right we serve trillions of tokens into our own runs.
54:04
tokens into our own runs. We serve billions of agents into our own runs and we actually have built all right the technology necessary for cluster managing very large clusters across many different GPU fleets uh serving the inference on top of them sandboxing into agents at scale and you know with a with
54:23
in a closed lab that's again abstracted and so you kind of just get the experience out of it but if you're an enterprise or company that wants to own your stuff we externalize all this software that we built and already battle tested at scale for them right and effectively make them um have the same capabilities as a frontier closed model, right? Through a partnership. Through a partnership.
54:43
>> As we're describing all this, and I do buy what you're saying that this is the shift that the world is going through. I see it as well.
54:49
What are the implications for the frontier closed labs?
54:53
Do you guys think in terms of how would they need to react or not?
54:56
Do you think there's room for them to continue pushing the way they are and for this open ecosystem we've been talking about?
55:04
I I really think that it's again like a this and that world and we keep underestimating the depth of um you know of revenue potential in the AI markets because it is artificial labor on tap and uh labor has actually been to to Jiannis's point like that's sort of it's
55:21
hard to get more labor for things right it's hard to like there's a shortage of all sorts of labor in this country and other countries and I think that uh it's it's hard to understand what a world looks like when you have kind of um uh artificial and human labor kind of on tap, right? Because it's kind of humans
55:36
Because it's kind of humans working with many agents to produce things.
55:39
Uh so the depth of these markets, we're still early on.
55:41
Uh right, the open model stuff has just started.
55:45
Uh the markets are starting to mature, but I think that we're underestimating the depths of of uh these markets.
55:48
And so when you go back to the equation of intelligence density times number of GPUs times trust, I think the closed model labs will do just fine because they have those three things.
56:01
they have those three things. But I think a much more positive and the way you know this is kind of where the world is going the momentum is uh towards a world of multipolarity where there are instead of two three companies that uh you know have this equation that open models enable um not just a model builder but other companies building on
56:22
top of that right to um to also thrive and I hope right we we move into a world with um a multi-polar world with a lot of intelligence providers Obviously there's certain only a certain amount of companies that can raise let's say billions of dollars or so forth but it's much better to have 10 20 than it is u to have two or three and I think that's kind of where where this world will head. >> Let's end on maybe teasing a little bit
56:46
>> Let's end on maybe teasing a little bit about what you guys have coming.
56:48
I imagine you're training a large model.
56:53
Give me a sense of what you want to be pushing on as you continue to build these models and the cadence we can expect from you.
57:00
expect from you. you've been in stealth now you've got this model it's very exciting but what can we expect in terms of cadence and sizing and and all of that >> yeah that's a that's a good question I think like uh the uh both the mission and the vision of the company is to close the gap between the open and the closed frontiers and ensure that like there is uh one frontier and you know reflection is at the frontier of
57:22
intelligence in order to do that like you need to both be scaling in terms of compute in terms of like the model size so we'll have like you know significantly bigger models coming out like um uh in in 2027 and also in terms
57:36
of like your scale of reinforcement learning uh and then the thing about reinforcement learning is that uh it can it can happen in like waves that's why you see that many of the other models like GLM 5.1 5.2 5.3 or like uh you know 1 5. 2 5.
57:46
3 or like uh you know fable 5. 1 5. 2 5.
57:50
5 and so on and so forth.
57:54
So uh you know there's going to be a similar uh structure where we'll just like have a a bigger base model and then we'll have like multiple releases like uh following uh after that.
58:03
Uh so there's a lot of like uh uh more work that's happening already.
58:08
We already started like building our next generation of models and uh uh we'll be able to release uh and share more with the world in 2007.
58:17
>> Yeah, there's uh there's something about you know Giannis Giannis's work actually in Alpha Go and Alpha Zero.
58:21
Um, one of the things that I thought was so inspiring for me and why I switched from physics to AI is that um, an algorithm was developed that never stopped learning.
58:31
And so by the time that AlphaGo beat Lisa Doll, uh, right, it became superhuman or super intelligent.
58:39
And uh, you at that point it's just well, you can put more compute in it and it can become even more super intelligent.
58:43
intelligent. uh it's just that it didn't make necessarily economic sense to maybe make it you know why why if if you already if if you're already that good what's the economic value of being even better at go now when we actually started the company we thought that this moment would come for uh uh general models general coding agentic models and one of the things the team really
59:05
cracked is reinforcement learning just keeps scaling right we've uh this is we believe the largest um reinforcement learning done uh reinforcement learning run done in open source that's been openly documented uh right it's around 10,000 GB300s that have been running for weeks and uh it never stopped learning right it's a it's completely compute bottleneck that you just put more compute and the thing keeps on learning
59:31
and so what we feel has happened and this is across the field and I think the closed labs have kind of figured this out a bit earlier uh but it's extremely important this kind of technology and science is also written about in the is uh the field has figured out at a scale models not just you know scale pre-training but also scale reinforcement learning and so now we're in this moment where you can really make
59:55
the model super intelligent at anything um and it's just an economic t like you know how much money do I want to put in to um and how do I get the data and so forth but I think some of the fundamental scientific breakthroughs
1:00:08
have been made and now it's about making them a lot more efficient and that's really exciting because it's kind of It's a full circle moment I think for both of us for why we got into this field. Um Giannis I think AGI believer
1:00:16
Um Giannis I think AGI believer in uh I think you joined DeFi in 2012. >> Yeah.
1:00:22
>> Uh and it's it's just incredible that I mean now it's 14 years after and uh it's hard for me to dismiss these systems as anything other than um AGI or something um on the clear kind of smooth uh path to it.
1:00:40
uh and now it's the question is how do we operationalize it and make these the benefits of this technology uh distributed throughout the world.
1:00:48
We believe multipolarity is very important.
1:00:51
How do we do this as safely as possible?
1:00:53
Again, we believe transparency, visibility and ecosystem of builders is important.
1:00:57
U but it's uh it's a really incredible time. >> It is.
1:01:00
I look forward to uh seeing what else you guys have in in the coming months as well.
1:01:04
Thank you for your time both of you. >> Yeah, of course. Thank you. >> Thank you. Thank you. >> Thank you.
1:01:10
Banking should feel like modern software.
1:01:11
Get everything you need in one place. Visit mercury.
1:01:14
com to learn more and apply online in minutes.
1:01:16
Mercury is a fintech, not a bank.
1:01:18
Check the show notes for details.
1:01:19
Granola is the best AI notepad I've tried.
1:01:22
It works everywhere on a video or phone call, in person, or an Apple Watch. Try it now at granola.
1:01:28
ai/sources and use the promo code sources at checkout for 3 months off.
1:01:32
Jerroad Lassin is where your team and your agents work from the same context. Try it free at jira. com. That's jir. com.
1:01:43
Framer is the AI native website builder that lets you build faster without giving up control. Visit framer. com/sources for 30% off.
1:01:50
Rules and restrictions may apply.