How DeepSeek leveraged Qwen and Llama to build its model in $5M
Open-source AI requires disaggregating components, moving beyond restricted licenses like Llama's to foster true community trust, and leveraging distillation and small language models for sustainable global innovation.
Highlights
4 clips · 6 min of 40 min-
- Disaggregate AI into components like models, algorithms, and datasets rather than defining AI monolithically.
- True open source requires that anyone can use the component for any purpose.
- Evaluate each layer independently against standard open-source licenses.
-
- Meta's Llama license includes acceptable use policies and commercial user caps.
- True open source allows unrestricted use without user thresholds or policy limits.
- Labeling restricted models as open source is a form of open-washing that erodes trust.
-
- Distillation techniques use existing models to train equivalents without building from scratch.
- This approach slashed training costs dramatically, dropping from hundreds of millions to millions.
- Small language models derived this way enable powerful on-device applications.
-
- Open ecosystems thrive on community engagement, maintainer support, and organic word-of-mouth.
- Building successful open projects requires mastering the art of collaboration, not just publishing code.
- De facto standards like MCP emerge because they solve real integration challenges across ecosystems.
That’s all the highlights.
Questions this video answers
9 · every answer cites the moment it was said01 How should we define open-source AI given its emerging components? 2 sources · 7:01
We should disaggregate AI into its components—like algorithms, weights, data sets, and agents—rather than trying to define the whole system at once. Each component must be evaluated independently to see if anyone can use it for any purpose under an OSI-compliant license.
02 What is the difference between open models and open weights? 1 source · 9:10
When using open weights or open models, you are only getting that specific model piece rather than the whole large language model ecosystem or training data. This means certain conditional parts and the underlying datasets remain closed and unavailable to the user.
03 Why is Meta's Llama license not considered true open source? 2 sources · 12:01
Meta's Llama license includes acceptable use policy restrictions and a commercialization provision requiring a commercial license after hitting a certain user threshold. True open source requires that anyone can use the code for any purpose without usage caps or restrictive conditions.
04 How can I check if a model's license is genuinely open source? 2 sources · 16:35
Developers should verify that the model relies on an OSI-approved, OSD-compliant license like the GPL, Apache, or MIT without any unauthorized amendments. Additionally, good documentation is critical, as demonstrated by DeepSeek R1, where clear documentation allowed Hugging Face to quickly reproduce and train equivalent models.
05 How does model distillation dramatically lower training costs? 2 sources · 17:59
Distillation involves using existing powerful models like Qwen and Llama to train a new model instead of building one entirely from scratch. This technique drastically reduces resource requirements, bringing training costs down from around $100 million to just $5 million.
06 When should I use a cheaper derived model over an expensive API? 2 sources · 23:06
Cheaper derived models or small language models are ideal during prototyping or rapid development tasks where you need fast, cost-effective outputs rather than slow, deep reasoning. As open-source models approach production-grade quality, they provide a natural, cost-efficient alternative for various workflows.
07 How do specialized training and agents improve enterprise ROI? 2 sources · 25:42
Industries are refining base models with sector-specific data, such as telecommunications networks collaborating to train domain-specific models. Combining these refined models with an agentic layer delivers truly productive outputs that solve real business problems and secure return on investment.
08 Why is community trust essential for open-source AI projects? 2 sources · 14:55
Open-source success relies heavily on community, collaboration, and contribution rather than just a legal definition. Practices like open-washing destroy trust, whereas frameworks anchored in foundations like the Linux Foundation ensure long-term security and ecosystem growth.
09 What blockers are currently holding back enterprise open-source AI adoption? 2 sources · 20:46
Adoption is currently hindered by risk management fears, a lack of understanding among legal and finance professionals, and general anxiety around job displacement. Over time, as general skills and confidence improve, open-source AI will naturally become the standard industry norm.
Section-by-section summary
8 parts
Defining Open Source Components in AI
Traditional open source principles focus on anyone using code for any purpose, but AI requires disaggregating components like models, weights, and datasets. Rather than defining a monolithic open AI, we must evaluate each component independently against open standards.
Open Models Drive Unprecedented Global Innovation
Access to open models like Llama and DeepSeek accelerates innovation by giving developers worldwide the building blocks they need. This democratizes technology, enabling countries outside the US and China to participate in AI development without prohibitive costs.
Understanding Open-Washing and Commercial Restrictions
Many models labeled as open source, such as Meta's Llama, include acceptable use policies and commercial user caps that violate true open-source freedom. This open-washing undermines trust and community collaboration, which are essential for long-term success.
Distillation Dramatically Lowers Model Training Costs
Techniques like distillation allow developers to leverage existing models like Qwen and Llama to create highly capable equivalent models at a fraction of the cost. This efficient approach enables smaller teams and nations to build powerful small language models.
Agents and Specialized Training Drive ROI
Agentic layers can help compensate for model inefficiencies while domain-specific training enables sectors like telecommunications to achieve real business value. Focusing on lean compute and abstention training further improves practical productivity.
Community Building and Standards Like MCP
Successful open ecosystems rely on deep community engagement, maintainer support, and organic developer word-of-mouth rather than mere code releases. De facto standards like the Model Context Protocol bridge different systems and maintain trust.
The Death of the Traditional Intellectual Property Moat
Intellectual property no longer acts as an impregnable moat as open technologies proliferate across global infrastructure. Collaborative open development offers a sustainable alternative to having a handful of tech giants control our AI future.
Overcoming Enterprise Skepticism and Fostering Sovereignty
Enterprise adoption of open models faces temporary hurdles from risk aversion, lack of understanding, and regulatory fears. Building national foundations and supportive local ecosystems will be key to unlocking sovereign AI adoption.
n next · p previous clip · t highlights/full
< > slower/faster · / search what was said
-
Open weights2× · first 2:25
Model parameter files provided freely to users without including the training datasets.
-
Open washing1× · first 14:02
Misleadingly implying that a deliverable benefits from open-source goodness when it does not truly meet open standards.
-
Model distillation1× · first 18:06
A technique that uses existing powerful models to train new equivalent models and significantly lower costs.
-
Small language models2× · first 18:34
Compact language models designed to run efficiently on phones and lean infrastructure.
-
Model Context Protocol3× · first 15:16
An open standard used to connect AI systems and tools together seamlessly.
-
Open innovation2× · first 11:52
A term used for models with usage restrictions and commercialization provisions that do not constitute true open source.
7,650 words · about 31 min to read · YouTube’s automatic captions
Defining Open Source Components in AI
There was a famous leaked memo from Google in May 23 and it says we have no moat. And what they meant was the intellectual property which will hold something safe, which will keep everybody off, allow you to charge revenue. And I think that's going to be the case with most things in AI. We've ended up with eight companies who control the digital infrastructure. And I don't think anybody wants to see that as the AI future with eight companies controlling our AI future. So, the value of open things up to something that is much more understood now than it was years ago when we began those big tech journeys. If you talk to people today about Lama and Meta, they'll say to you that they're moving away from openness.
And I think it's because they haven't done it right. The license has two things. It has an acceptable use for policy that puts restrictions in and it has commercialization provision that says when you hit X million users, you have to go back and get a commercial license from Meta. To get the real value from open source, it goes way beyond that legal definition of having a license and making it open. It goes to the heart of community and collaboration and contribution.
Given that there are significant number of open source options for people to pick from, why are most developers still defaulting to closed model APIs as their go-to?
It's interesting. I think
Back in November, we hosted the first ever in-person AI Native Dev Con in New York. This June 1st and 2nd, we're bringing it to London. It's two days built for AI Native developers and engineering teams. One day full of hands-on workshops and one day full of practical talks on agent skills, context engineering, agent orchestration and enablement platforms, and how teams are actually shipping AI in production.
Join us at the brewery in London near the Barbican for all of that plus networking, parties, giveaways, and a room full of people building the future of AI native development. You can also join us from anywhere in the world via the live stream. As you're listening to this podcast, you get 30% off your ticket with code pod30. Just head to ainativedevcon.io and we'll see you in London.
On today's episode, we're going to be talking about AI openness. What does an open-source model actually mean? What are open weights? And what is open training data? And what do we in the West need to learn from our Chinese counterparts in order to not fall further behind in our open-source journey? Hello and welcome to another episode of the AI Native Dev. And today we're going to be talking about all things open-source. Uh joining me today is the CEO of OpenUK, Amanda Brock.
Amanda, welcome to the session. How are you?
Thank you very much for having me along, Simon.
It's an absolute
Finally.
It's an absolute pleasure. We've been talking about this for quite some time now. And uh either you're traveling, I'm traveling, you're ill, I'm ill, but finally we found a we found a day where we could be here in person in London. And you're of course UK-based. Um OpenUK, tell us a little bit about OpenUK cuz I think a lot of our listeners will maybe have heard of OpenUK from some of the reports and and and research that has been created or maybe State of Open Con in the UK, a conference you run. But tell us a little bit about what it is.
So, we're just over 6 years old at OpenUK and we are an industry organization with a difference. We realized very quickly that the other industry organizations in other countries for open tech, so software, hardware, data, standards, and AI, that they focus on companies and by doing that, it's what you would expect from an industry body, right? By doing that, they end up with a small group of companies and they miss a whole part of the open-source ecosystem.
So, we focus on individuals, we bring people together, we've become a sort of convening point for the open tech sector. We then do a bunch of research and reporting, so we do legal and policy work, we respond to legislation, we do research into open tech in the UK and beyond. And then we do a bit of skills development, we have the annual conference that you've mentioned. We're doing that a bit differently this year, which we'll probably come back to. So we really bring people together to talk about the the stuff that's going on in open tech.
Yeah, it really is a truly wonderful organization. And of course And of course your background is is also in the kind of more on the on the legal and law side, right? You've been 25 years legal experience and you've done a lot of work there with open source legal frameworks as well as internet law.
Yeah, so I I guess I started in the mid to late '90s as a lawyer, went through the dot com boom. I actually thought that that was going to be the pinnacle of my tech career and I didn't think there was going to be anything else coming down the line that was going to be quite as um innovative and new and I really had enjoyed that part of my career because you're trying to fit the laws around the technologies or make, you know, rather than make the technology fit the law if that makes sense.
So in 2008 when I joined Canonical, I was just blown away at the stuff that was going on and Canonical back then was really cutting edge. I I can't really comment so much on what they're doing today. But it was really a moment in time in the industry, so I've worked in open source on and off since 2008. And I thankfully I will say gave up being a lawyer in what, 2018 19 and started the role with OpenUK in 2019.
Amazing.
So in theory I still use my legal skills in the work we do, but I don't have to be a lawyer anymore.
Amazing. And so when OpenUK started, of course it was probably we would say the more traditional open source which we think about open source maintainers building software components that is available for reuse under certain licenses and so forth. Then AI happened or as course AI's been happening for a long time. AI hit mainstream and it really caught everyone's attention and Open UK has recently actually done a lot in the in the world of AI as well.
Let's talk a little bit about what open what openness means to AI and to LLMs. When we think about an open model, which has been thrown around a lot, what is an open model?
Oh, you could get me into so much trouble with that very small simple question. There's been so much debate about it and we've seen people trying to come up with definitions. So, I'm going to give you what I think this is all about. And we all know what open source software is, right? We spent 30 years
1 Highlight · Disaggregating AI components to define openness
working on the open source definition, working on licensing based on that, relying on it as a a trust point. And that trust point has allowed us to build traction. And for me, key to it is that anyone can use the code for any purpose. Now, if we take that and sort of bring it to the next level for AI, to me it means that anyone can use what you're giving them for any purpose. Now, you can get into all the debates with models and things about whether you have to give the data, how much of it do you have to give. My preference is that we disaggregate the technology. We don't try and define something that's so emerging and that we look at AI as it evolves. So, couple of years ago when people were really trying to do definitions, all they were thinking about was language models. They weren't thinking about agentic or any of the stuff that's coming down the line now, the robotics, whatever else, embodiment. Um so, I think the best way to view it is to disaggregate the components, look at what the components of the AI are and
ask the question about the components. So, if we're looking at a model or an algorithm, if we're looking at a data set, if we're looking at an agent, is that something that anybody can use for any purpose? And that goes back to how it's licensed. So, I would look at each of the different components and then whether or not it's on a an OSD compliant license. So, the licenses that you know from open source software like the GPL, Apache, MIT, those kind of things.
And if it is, then I would say that that part of it is open source. It might not be that the whole of the AI is, but that part of it is.
And it's interesting when we say open source cuz open source just means, you know, the the the source code is there for you to see. Now, if I was to try and build an open source component, a piece of software, I could probably go ahead and well, it would be open. I could probably build it through through through GitHub, etc. And some CICD pipeline. Uh with a model, though, of course, just having the code isn't really enough to be able to build the model and use the model to the same level. So, there are other there are other kind of like terms that we've have been thrown around here. Open model, open weights, open training.
So, for me, I would say that you're looking at the component part, which is a recognizable part as the the algorithm, the weight, the model as being open and meeting that open source standard. So, it's something that's given to you freely in order that you can use and reuse. It doesn't mean that all the conditional parts are going to be available. So, particularly when we look at open weights and open models, I think the real differentiator there is that you're not getting the whole of the LLM or S small language model, whatever it is that you're using, that you're only getting that piece of it, the model piece. You're not getting the data. And that is almost universally the truth.
Open Models Drive Unprecedented Global Innovation
Not quite, but almost.
What's the value to the end user of something being open source as a as a model?
Yeah. I think it's enormous and I think it depends on where you're sitting in the ecosystem and who the user is. Value depends very much on who you are and where you are in the ecosystem. And if we're looking at innovators, that access to technology, and when we saw the the first LLM's being opened up in 2023, first of all with the Llama leak and then Llama opening up in the July, what we saw was a pace of innovation that was really unprecedented, right? And it was suddenly giving the ecosystem the access. And one of the big values of something being open is a trait of the development development, innovators being able to freely access and use technology they wouldn't otherwise have access to and to build on top of it. And you see that shift from Meta with Llama through to Deep Seek with R1, you know, and that wouldn't have happened one without the other. So, you get that kind of innovation. You also, when we look at the landscape globally, we see two
dominant players, China and the US, way ahead of everybody else. And then we have the middle countries, there's about 10 of us together. And really affording what China and the US has built without collaborative innovation would be impossible. So, we really need access either to what China and the US have done, or we need access collaboratively to further innovation to be able to just fund the scale of the the cost of AI innovation. And then for sort of end users and individuals, it's again, it comes back, as most things with open source do, to cost benefits, democratizing technology, allowing more access, you know, in in India the AI summit recently, we talked a lot about access for all. And when you look at something like the global south, the only way they're ever going to get that innovation is if we open it up and create access.
Yeah. Yeah, it's actually a lot and a lot of the the the things that you mentioned there are also still very true of traditional open source that you mentioned with libraries and people building upon that. It's a builder's world, right? Um, so when we you
Understanding Open-Washing and Commercial Restrictions
2 Highlight · Why Llama is open innovation, not open source
mentioned uh Meta and Llama there. Now now Meta stated that Llama's open source. The Linux Foundation backed them up. You called them out.
Yeah, that was all a misunderstanding, right?
The Linux Foundation didn't back them up. So, it it all got very confused. So, in July '23, we were actually Open UK was the only organization that supported the launch of Llama as open innovation.
Mhm.
And it was very carefully structured all the way through, I think it was 23rd of July off the top of my head, to that date as open innovation because the license has two things. It has an acceptable use for policy that puts restrictions in. And it has a commercialization provision that says when you hit X million users, you have to go back and get a commercial license from Meta.
Mhm.
Nobody knows what the the terms of that license is cuz I don't think anybody's hit that level of use that's been trackable
Mhm.
to date. So, what that means is when you go back to that basic principle of open source that anyone can use it for any purpose,
Right.
that just doesn't happen, right? You've always got the risk of this restriction. You don't have that same free flow, that ability to cascade the outputs and let somebody else use it, iterate it, and build on top of it. So, that that's kind of broken. And it was fine as open innovation and we supported it, and I still think it's the right thing to do because that shift opening up an LLM was critical for change in the industry.
Mhm.
And I think it will go down in history as one of the seminal moments. If you talk to people today about Llama and Meta, they'll say to you that the the word on the street is they're moving away from openness.
Mhm.
And I think it's because they haven't done it right. They haven't actually created true open source.
Mhm.
So, they they themselves weren't calling it until open source until Zuck posted on Facebook. And by that point, you know, nobody was going to go backwards from there. Whether it was intentional or not, we'll never know unless he tells us, of course. Um but, you know, I
recently met Yann LeCun. I had a really interesting conversation about some of his aspirations and he thinks we should be building a global model across countries, across borders, collaborating.
Mhm.
You know, and bringing money from states together to do that. So, I suspect they understood what open source was.
Mhm. Absolutely.
Yeah.
And for people who are building on something that they believe is open source, but actually it's not truly open source in in to your definition, what are the what are the real-world risks that uh I as a developer might have if I'm using something thinking it's open source, but actually it's not?
Well, we get into uh something called open washing which existed in open source software before. To be quite honest, it's such a nerdy thing that I never thought it would be in the mainstream. And there have been headlines in the last couple of years in the New York Times and The Economist with open washing on the the cover of the the magazines. So, open washing is where somebody takes that open source goodness and implies that the deliverable they're sharing with the world benefits from that open source goodness when it doesn't really.
So, if you're Meta and you're saying Llama is open source when it's not, you're misleading people. It's disingenuous at best. But, what it means is you can't have the same reliance in the ecosystem. So, you can't take, use, recycle, know that you know, those taking it from you can also use and recycle. And that becomes problematic, you know, at at its simplest term. And I actually think it's why Meta and now are looking at shifting away. Because to get the real value from open source, it goes way beyond that legal definition of having a license and making it, you know, open.
It goes to the heart of community and collaboration and contribution. And if people can't trust that openness is going to be there forever. So, it's like we we did a piece recently with one of the creators of MCP, and he was talking about why they've put MCP into the Linux Foundation's new Agentech AI Foundation. And it's all about keeping it open and knowing that you've got that security forever of it being open. And I think that that's the bit that's missing when you open wash, you take away the trust. And if you don't have trust, you will not be successful because you don't have the ecosystem around you and you won't grow that ecosystem the way you would with real open source.
Yeah, trust is a really important word here as well. It's like it's it's it's what you know, when we talk about open source and talk about the usages and the sharing of of open source code and open source models, trust is really what binds the whole community together in in terms of that. So, it's yeah, it's a term that I think we really need to think about when when talking when using this this language. Um so, for for me as a developer, if I'm using an open source or if I'm using a model rather, how how do I go about working deciding whether this model is open source, how open it is? Is there a good way easy way I can do it? Mhm.
That's quite difficult one. So, I suspect that if you're a developer doing that, you're more technical than me, so you understand the component parts of the model better than I do. And you are better placed than I am to to judge which bits you need. Uh from a licensing perspective, you want to make sure that it's on one of the OSI approved OSD compliant licenses that we all know, you know, the GPLs, Apaches, MITs, something that you've heard of and that that hasn't been amended. And that I think that's quite critical to it. So, it's really understanding what the licensing is and then which bits you're getting and you're not. Um I think we really see a benefit from documentation. And when we saw Deep Seek releasing R1, one of the things that really differentiated it was that the data that it was trained on wasn't provided. But the documentation was so good that within days Hugging Face had built R1 open, you know, and they're able to go and train the same thing themselves cuz they had that replicable data to help them understand.
Mhm.
So, I think you're looking at what do I need to use this if you're a developer.
Yeah, absolutely. Let's talk about um Well, actually one of the big rumors more recently is about DeepSeek-4.
Right.
Uh when that's going to come, I don't know. If that's going to come, I don't know. But, um back in the day, a year ago or so, when in fact I think I ran a panel at State of Open Con
You did.
in and around DeepSeek with some of my Chinese colleagues.
That's right. Yeah, and it was just as DeepSeek came out, I think a couple of weeks maybe after or a week after.
Um now, DeepSeek was a model that was I
Distillation Dramatically Lowers Model Training Costs
3 Highlight · Distillation slashes model building costs
think it's a distilled model from
Yeah, R1 open. R1 open R1 not R1 open, yes.
Yeah, and and so first of all, what is a distilled model?
Yeah, so they took Qwen and they took Llama and used what was in there through this distillation technique to create something equivalent without actually building a model from scratch. And that was super impactful because it took the price of building a model down apparently to 5 million instead of 100 million, right? So, 5% of the the cost at that point in time. I think that that distillation was really huge in terms of the innovation for the time. But, we look at models now, um India has just released through Sarvam open models which are small language models.
Mhm.
And it's quite interesting to see how that progression has worked in the last year. And that I think these small language models, you know, that can be used on phones and things that really that's where we're shifting to as a
market.
Mhm. And it's interesting when people are saying, "Oh, yeah, this is my model." How you actually sometimes don't even realize the models behind it. Um I think uh I I I was reading a thread. It was either on Hacker News or Reddit uh this week just gone about uh the new uh Cursor's is it the Composer-2 model or something
Oh, I don't know. So, I had quite an interesting experience in December. I went out to China. I several years ago now edited a book in on open source law. And it's been translated by Chinese open source community into Mandarin.
Mhm.
It's a 640-page long book. So, it took a bit of time to do it. And I was really lucky to get a call with DeepMind when I was there. And I'm not allowed to discuss the contents specifically, but it was an interesting experience. And it was interesting not just to talk to them, but to talk to people across the marketplace there. And you know, like DeepMind is obviously big, Quan and Kimmy, but Kimmy seemed to be the thing that absolutely every developer was using in China.
Interesting.
really seemed to be running ahead for the developers.
And that's interesting with Kimmy because there were rumors, I think, on either Reddit or or Hacker News that Kimmy was actually the the core model behind Curses Composer 2 model, which was just recently released. And those are I don't know if it was trolling. I don't know if how much proof there is, but it's very interesting that I think more and more of developers are becoming uh it's becoming it's becoming clearer to developers that not every model is built from scratch. And different models are learning from each other. Absolutely.
Yeah, it has to be. It really has to be.
So, so what time do you think it makes sense for a developer to use, you know, an expensive model, maybe an Opus or something from from OpenAI versus that, you know, the cheaper model, you know, these models that are derived from each other because the costs are so low. They're naturally going to be cheaper to the end user. So, what stage does it does it does a developer need to make that decision?
I think we're in an odd position right now where when we look at uptake and adoption, it is actually much lower now than you would expect it to be in the open source front. And it's very reminiscent of open source software sort of 20 years ago. And I think a lot of it is for similar reasons, which were about risk management, lack of understanding, you know, maybe unnecessary fears, and not really knowing what what the the consequence of of it being open were going to be. And I think what we saw then was risk professionals, lawyers, procurement, finance who had to sign off on contracts saying no.
And then there was sort of a shift and a momentum gathering around GitHub. And I think what we will see, it won't be a decision about a specific engineer developer making a decision to use it, is we'll see an industry shift where it just becomes the norm. And I think it's an absolute inevitability. Um I'm reassured by hearing people like Yann LeCun saying the same thing, you know, at events I've been at in the last few weeks. I I just think there is no way around it, particularly for anybody who is not from the US or China. And of course, if you're from China, you're already probably using open models.
Yeah, absolutely. Do do you feel like there's a group, maybe it's a a startup versus an enterprise or something? Do you think there's like a a style of user that is that finds open source models, you know, get
I think there's a I think it depends on where you are in the ecosystem. And there's a lot of concern about IP, managing your IP, leveraging your IP, and making money, right? And I I think you may be the wrong podcast to be saying this, but I think there is a piece where there's no return on investment right now, right? People are not seeing the returns that they expect to get from AI at this stage. And they're working out how to actually use it in a practical way when you get to that business level.
Um and I think we're a ways off people being able to say, "Here is how AI is working for me and here is how it's saving me costs, taking away risk, increasing my productivity." So I think there's, you know, different levels in that ecosystem. There's the core creators, the sort of development community, and then the end user. And it's a gradual shift that we're going to see. I don't think it's particularly about anybody being the right user for it. I think it's about building understanding. I think there's been a quality issue for a long time.
And we see that shifting now. I'm told that there's very little difference in the testing now. That you you're open source models are getting to a point where they are good enough.
And and there's probably like a a
Agents and Specialized Training Drive ROI
variation in the accuracy that is needed per task as well. Sometimes you might want to do some prototyping or something just to try think out an idea and and actually get some rapid development. It's not necessarily production code. In that case, you know, I don't want a slow model that's thinking very deeply.
Absolutely.
something that will create something fast and cheaply for me.
Yeah. And that was one of the things I took away from China, to be honest. They're absolutely obsessed with keeping everything lean and reducing compute. All right, because they don't have the same access that others might have. And I think we will increasingly, particularly in this sort of world of geopolitics, I think we will increasingly see people being more and more conservative about the access they have to compute and infrastructure. And looking at what's most productive. And I think some of the other models will start to come into play then. I think there'll be a a lot more around these small language models and opening up models, sharing more.
How much do you think an an an agent, an AI agent, can plaster over some of the uh some of the problems that a cheaper model, well not problems, but some of the inefficiencies maybe that a cheaper model will will will will show that a more expensive model would be able to just do out of the box.
That's an interesting way of looking at it. I think we're seeing such a shift to agentic because it really starts to deliver what we all thought AI was going to give us, right? It's a realer, more productive output that you're getting by applying an agent to what you're doing. I suspect that that is exactly as you're describing it. And a sort of sticking plaster that will hide a a multitude of sins with the models and I hadn't really thought about it like that before but yeah I think probably right.
sometimes you have to ask a model a hundred times to do something and it gets it right once. An agent can share that with me and I don't have to go through the pain of the 99 problems.
No, but I've also been talking to people recently about abstention and models being trained not to respond when they don't know the answer and it's not about hallucination per se it's just about some of the answers you get from models because they want to always please. And I think retraining them in that way is quite important. So I think we're seeing shifts there that will actually be the quite critical for the the business productivity of models and the agentic layer on top of it.
I also speaking to people increasingly see sectors looking at taking models and refining those models for the sectors and training it you know Telco is one great example at Mobile World Congress a couple of weeks ago the GSMA one of their industry bodies has pulled together all the big Telcos all the network providers and they are working collaboratively to train on specific data related to Telco to refine their models and you can imagine then how your agent's going to work with that. It's going to become super useful and I think as that begins to sort of bed in that's when we're going to really start to see return on investment and that's going to be critical to the landscape.
Absolutely yeah. We've talked about things like Lama. Lama everyone has heard of Lama I suspect. Is there is there a model that you think an open source model that you think oh do you know what this is a really great model it does so much but people aren't talking about it or people just don't haven't heard of it before.
I think Kimi is the the piece that we're missing in the West just from the reactions that I was seeing from engineers from developers around me in China you know it it's pace of of was so fast and so great, that I think Kimmy's probably being underutilized in the west.
And tell us a bit about Kimmy.
So, Kimmy's Chinese open source model created by Moonshot, which I think came out towards the end of last year. And I've not used it, but I know that every developer I spoke to said that it is just the best.
Yeah.
So, I can only recommend it through word of mouth and other people's recommendations rather than my own.
Community Building and Standards Like MCP
4 Highlight · Building community and organic developer adoption
Yeah, I know I've heard I've heard actually quite similar. So, yeah, I've heard it. It's funny how when It's like a drumbeat, it's like when Claude and and OpenAI were originally kind of like bringing out their models and everyone was saying, "Oh, yeah, Claude's the best thing. Claude's the best thing." And it's it's all about word of mouth a lot of the time, isn't it?
Well, this is the thing that people don't understand. So, when you are in that open ecosystem, or even an ecosystem, it's about community, right? And I think within engineering over the last 20, 30 years, it's something that's been refined so that it's now a dark art understand how to build community, how to get contribution, how to engage people, how to keep them engaged, how to be there for maintainers. You know, that stuff, if you're not from an open background, you don't understand it. You think you're going to build something and they will just come. You don't understand that you need to have a user base and people start and that that's sort of, you know, uh drum signals that everybody hears and everybody suddenly around. I mean, Maltbook.
Mhm.
How quickly did you know about it? Cuz I was sitting by Saturday trying to get myself all into it, you know? It's just one of those things that our whole ecosystem converses, whether it's through social media, whether it's through, you know, sharing tools amongst each other, but that there was a very definite way that it happens. And I think that art of collaboration is something that we're going to see more and more.
Mhm. Uh you know, there's a significant number of open source options for people to to pick from. Why are most developers still defaulting to closed model APIs as their as their go-to?
It's interesting. I think that's shifting. And when you say closed model APIs, do you mean closed models or closed model API? And if they even if they're using a closed model API, they're probably using MCP
Mhm.
as a protocol, which is open to get there.
That's a great question, actually. So, explain the difference there between a a closed model, a non-open source model,
Yeah.
a closed model API,
Yeah. the same. It's whether or not they the API is open or closed, and whether it's something that's freely available. Generally, you will get an open API because somebody wants to build an interface that everybody is able to access.
Mhm.
And then when we look at MCP, model context protocol, you have a an open standard effectively,
Mhm.
which hasn't gone through a standards process. It's a de facto standard like so many things in open tech. And it's just become the thing that everybody uses because it does what they needed to do, and it joins things up. It joins up between the old and the new world, and the new world and the new world. But, all of that being open means that in that ecosystem, even if you're going from a closed model to a closed API, you're still using something that's open in the middle.
Yeah, absolutely. Very interesting. Let's switch a little bit from models
Yeah.
to agents. We've touched on agents briefly.
Hey everyone. Hope you're enjoying the episode so far. Our team is working really hard behind the scenes to bring you the best guests so we can have the most informative conversations about agentic development. Whether that's talking about the latest tools, the most efficient workflows, or defining best practices. But, for whatever reason, many of you have yet to subscribe to the channel. If you're enjoying the podcast and want us to continue to bring you the very best content, please do us a favor and hit that subscribe button. It really does make a difference and lets us continue to improve the quality of our guests and build an even better product for you. All right, back to the episode.
How important is it that if someone uses a an open model?
Mhm.
The Death of the Traditional Intellectual Property Moat
They then use an open agent as well.
I think all of this really comes down to personal preference and choice, right? I'm going to advocate for you doing it openly because I think that you are going to build community around it. You're going to build adoption and engagement, and you are able to take technology innovation and iterate yourself on it and share it with others, right? So, I think if you open that up, you're going to build an ecosystem that you're not going to build otherwise, and that's important to most people. But it is like the the conversation we're having before. It really comes down to my IP and whether people feel that the only way they're going to be able to monetize, if that's their goal, is to to hold on to the IP and keep it closed. You go back to 2023, and there was a famous leaked memo from Google in May 23. It is only the opinion of one person. I have to caveat it with that.
And it says, "We have no moat." And what they meant was the the moat the water around the castle, which is normally intellectual property, which will hold something safe, which will keep everybody off, allow you to charge revenue, that that just wasn't there once things started to open up. And I think that's going to be the case with most things in AI. I I suspect we're going to see it really open up. Partly because we're learning from history, right? We've ended up with eight companies. I think eight's right. Um who control the digital infrastructure. And I don't think anybody wants to see that as the AI future with eight companies controlling our AI future. So, the value of opening things up is something that is much more understood now than it was 30 years ago when we began those big tech journeys.
Yeah. So, looking forward now, let's let's let's get the crystal ball out and start start gazing into into what we think is going to happen. Um what would what would you love to see going forward in either maybe adoption but more the the the the innovation around open source and AI?
I'm probably not going to give you the answer you're expecting.
Go for it.
So, about eight years ago China was very clear that it was adopting an open source first strategy as a nation. And that was embedded into policy. And then very specific activities and actions were taken across the ecosystem by the Chinese government, Chinese enterprises. And I think if you want to have a successful open source AI ecosystem future, if you want to be able to compete with what China's achieved on the open source front and AI, I think you really have to look at that ecosystem and landscape and what we need. I don't know if you saw it, but about three, four weeks ago Kanishka Narayanan, who's the UK's AI minister, made a sweeping statement, which I'm very pleased with, but it's quite sweeping, that the UK is going to become the home of open source AI.
Mhm.
So, how is he going to do that? How are we going to make that happen? And I think we have to learn lessons from China. We have to build on what they've done and do something more. We have to look at what everybody else is currently doing and how the world has changed in that eight-year period. And I think for me, putting those pieces of the picture in place, so it involves things like capacity and skills development and open source. It involves things like building a national foundation.
And I think we need a body that could hold something as standard like MCP, that could hold a gentic or you know, language model technology on behalf of UK enterprises.
Mhm.
And I think that's going to be a big shift and I'm hoping that we will move towards that. But I think that model is something we're going to see in every country. There was a a press piece this week that described the Linux Foundation as if it was the US National Foundation.
Mhm.
But we've all been contributing to that, right? And it was talking about a land grab around how the US is trying to own the standards in AI. You know, China will be doing the same thing. For us to do it, we need the home of that. And I think we'll also see other countries like Germany and France doing something similar. They've both taken steps in that direction without fully doing it already. So, I think building the the environment for a successful ecosystem at a national level for me is the the missing magic piece.
That's what I'd like to see us do.
And I think, you know, we were alluding to it during the episode, models getting better, people adopting it more. Is there anything in the way, do you think, in future for people to really grab hold of open-source models and use it, you know, in in whether it's in production or or more aggressively than they are today? Or do you think it's just a a timing thing that it will happen?
I think it will happen. I think it's inevitable.
Overcoming Enterprise Skepticism and Fostering Sovereignty
I think what's in the way of it is, again, going back to 20 years ago in software, it's understanding and the people who are able to stop that happening. The sort of naysayers, those with concerns, who maybe don't understand the risks associated. I think there's that piece where there are various blockers. You know, you probably saw about a year ago, I think it was, JP Morgan's CIO saying that my supply chain isn't user agentic. You know, so that there are pieces like that that will shift over time as we get more confident.
And I think we probably have had so much press in the last 30 months around AI that has caused a lot of concerns for a lot of people. I mean, things like mobile would terrify people. You know, if you explain to somebody who's not in tech that you've now got social media for agents and they're off creating their own religion, they think that's how, right? You know, and you can understand why. So, I think there has to be a sort of shift in general skills, general understanding. People are scared they're going to lose their jobs, so they're resistant to it. And I I think we have to look at the future of society, the future of work, and the impacts that AI can maybe will have, but it's perhaps not as quick as people are panicking around, and also building that understanding and giving them the training that they need.
Yeah, amazing. And let's pivot slightly onto back onto Open UK. Um what can people look forward to uh a in the next year with
There's a lot going on. Yeah, so we've got ongoing reporting. We did our first international reports at the end of last year with India, because we were there for the AI Impact Summit in February. We did quite a few events there. So, we've got an Africa report coming out this week. Um we are working with the Chinese open-source ecosystem to build one in China as well, particularly focused on not just open-source, but AI. We have a bunch of reports. We'll do our annual report, all the bits and pieces that you're used to seeing from us.
We have a a number of events in Parliament. We just had one last week with the Conservative Opposition Ministers talking to them about the the needs of open source. I think you'll see a lot more conversation around sovereignty. We have taken State of Open, which is our annual conference that I know you know well, but we've taken it on the road this year. So, we've got confirmed dates now. I think 5th of June in Edinburgh at NatWest offices, and then 8th of July in Cambridge at Pembroke College. We'll do probably three more of those in the autumn, and then we've got our awards launching after Easter.
Great opportunity to nominate someone you know, or yourself even, for an award. I think we've got about 10, 12 categories this year. And that will take place at the House of Commons on the 5th of November.
Amazing. And you actually
You didn't pick up on that. It will take place at the House of Commons on the 5th of November.
of November.
5th of November. Yeah, Guy Fawkes.
Yeah, absolutely.
So, for anybody who's not from the UK or hasn't done their UK history. 5th of November is a night upon which many years ago a gentleman called Guy Fawkes tried to blow up the House of Commons, which we're going to on its anniversary.
Amanda, it's been an absolute pleasure and and and sorry it was so long. We should have we should have done this earlier, but it was worth the wait.
Thank you.
Thank you very much, Amanda, and all the best with OpenUK. And and for people who want to learn more about OpenUK, what's the what's the best place?
So, uh website openuk.uk, stateofopencon.com for the conference, and then LinkedIn is generally where we hang out. Although, as we were saying, maybe we need to do more on X.
Yeah, absolutely. Yeah. Do check that out. That's uh a really great community, very very welcoming community, and actually very up-to-date with a lot of what's happening. I love the reports. There was a really interesting one uh that that I actually saw in the Houses of Commons or House of Lords, I think it was, uh with DORA, uh the DORA uh and and OpenUK report, which was which was excellent. So, thanks for tuning in, and we'll see you on the next episode. Bye for now.