Speaker A
In some businesses, I have the CFO talking to me because they're like, "Our token spend is out of control. We spent a million dollars this month alone on tokens. How can I get that under control? What are the trade-offs I can
Mozilla CTO Raffi Krikorian explains open-weight vs open-source AI models, the trade-offs of self-hosting, and why everyday AI tasks may favor owning cheaper models over renting frontier ones.
Key Takeaways
- Open-weight and open-source are distinct; open-weight is the more common and practical category today.
- Owning and self-hosting AI models gives control over versioning, data location, and costs.
- Match the model to the task: frontier models for research, cheaper owned or open models for routine work.
- Cost and reliability concerns are pushing organizations to reconsider exclusive reliance on hosted frontier APIs.
What the video covers
- Open-weight models are downloadable model files you can run yourself, while truly open-source models would expose the full data, pre-training, and fine-tuning pipeline.
- Most models discussed as 'open' are open-weight, and the terminology is often conflated.
- Open-weight models allow full control, version pinning, and data privacy, since data never leaves your own infrastructure.
- Self-hosting carries costs: hardware, power, and staff expertise to run GPUs and manage systems.
- Fine-tuning open models lets organizations add proprietary data and tailor model behavior, which is not possible with closed API models.
- 'Owners, not renters' is the Mozilla framing: hosted frontier models can change terms, behavior, or availability at any time.
- Open-weight models are reportedly approaching frontier performance on everyday business tasks at roughly 30-50x lower per-token cost.
- Frontier models remain valuable for cutting-edge scientific research, while everyday tasks like email summarization suit cheaper owned models.
- Businesses face rising token spend, prompting CFOs to seek cost controls and trade-offs.
- Regulatory and privacy concerns, such as HIPAA, strongly favor on-premise or self-controlled deployment.
Chapters
- 00:00Introduction and Token Cost Concerns
- 03:00Open-Weight vs Open-Source Definitions
- 05:02Advantages of Downloading and Running Models
- 09:00Owners Not Renters: Control Over AI Infrastructure
- 15:49Frontier Models vs Everyday Tasks
- 22:02Hardware Options and Running Models Yourself
- 31:45Cost Trade-offs and Business Decisions
- 44:40Audience Q&A and Closing Thoughts
Full Transcript — Download SRT & Markdown
Speaker A
make?" On some companies that talk to me, it hits me up. Rafie Creoran is CTO of Misilla, which makes Firefox and publishes [music] the state of open-source AI report. What are open-source and openweight AI models? If you're confused, you are not the only
Speaker A
one. So let me start there. Like there's been a lot of muddling around these terms open weight, open source, etc. The way that I like to think about it. So open weight models are what you what people are generally calling the open
Speaker A
models these days. And what they are is like think of it as like a compiled binary. Like it is a model you can download. It's a file or a series of files that you can run on your computer,
Speaker A
your laptop, your server closet, your GPU, and your data center. Think of these openw weight models as just like a downloadable file that you can execute.
Speaker A
The thing is you don't really know what's inside. That's why I kind of compared it to like a compiled a compiled executable. like you don't see the source code but you can download it and you can run it which is very
Speaker A
different than like you know an anthropics fable or opus or or an openis chat GPT which runs on their servers you have no way to run it on your own servers so that's the open weight models like you can see the weights but the
Speaker A
weights don't really mean much to most people and then an opensource model would be one that you can truly see everything that went into the model like the data the fine-tuning the post-training, the pre-training steps, everything that went in the the entire
Speaker A
thing that went into the model itself. So you can inspect it end to end and understand truly what's in the thing.
Speaker A
Now there aren't that many open source models out there, especially ones that can rival the frontier level models.
Speaker A
Everything that we mostly talk about as the open models are these open weight models. Now open weight models have a lot of advantages besides just downloading them. like you can then fine-tune them. You can then base another model and train another model
Speaker A
based upon one of these open models because you can download it and run it yourself. There's a lot of benefits, but the drawbacks are things like if you ask your model a question, say for example, you were to ask a model about Teneaman
Speaker A
Square and it refused to answer, you would have no idea why. You couldn't actually poke into it and see it. But with truly open source then you can see where in the fine-tuning or where in the post- train pre-training they like
Speaker A
censored that piece of data out or change that data. So that's the difference between the open source and open weight models. The the names are getting confusing like a lot of people conflate them together but if you want
Speaker A
to be really precise that's the difference. So with the openw weight models, what are the advantages to downloading the models with the weights?
Speaker A
I mean, you're not going to go in with a text editor and start changing the weights of the models.
Speaker A
The real advantage is that you get full control over what's happening. Again, think of it like I've just downloaded a program. I'm running it myself as opposed to a web service that I connect to. When you connect to a web service,
Speaker A
you have no full control. Like they might decide to take it down one day, they might decide to update it one day, the blue might become a different color blue. But if you download it and you run it yourself, then you know that you've
Speaker A
pinned to that version. Like it's not going to change. Like this is a thing you are running from here on out until you personally upgrade it. There are lots of detriments to that though, right? Like now I'm like running a
Speaker A
computer. Like I'm actually have to be like an S sur like I might be running a bunch of GPUs. I got to pay my power bill. Like there are downsides to it, but the upsides are the things like you
Speaker A
you can pin it. I know for sure where my data is going. Like if I send my data to Anthropic or OpenAI, I'm just going to name those two just because they're big platform providers. Like my data now
Speaker A
sits on their servers. There's lots of reasons you wouldn't want that, right? like if you're like in a HIPPA environment or if you're in a regulatory environment or if you just don't want anyone else to know stuff about you, if
Speaker A
you run it on your own computer, you know for sure the data is not leaving your computer. So there are lots of advantages. And then the last thing I'll mention is I and I kind of alluded to it
Speaker A
on your first point, Michael, is you can then what's called fine-tune the open model. So yeah, you can't go in there. I mean, you could, but like you just probably doing more damage than anything else. You're not going to go in there
Speaker A
and change the weights. But you could train like another layer of the network that'll that'll take the open model and you can put your next layer on it to fine-tune it or add your own data or add your own color to it. So when people
Speaker A
talk about fine-tuning a model, that's basically what they're doing. They're starting with the open model and then adding a few more layers which contains their proprietary data to bend the model and to answer the way that you want it
Speaker A
to answer. You can't do that with the closed models, but you could do that with something that you download.
Speaker A
Rafie, you have written that renting the frontier is fine, but everyday AI is something that you should own. What did you mean? The phrase that we like to use around Mozilla is something that we call it owners not renters. And what I really
Speaker A
mean by that is like do you have full control over the intelligence that's sort of like powering your day-to-day application environment? And in the current way that most AI systems work where you call open AI, you call anthropic, you call Grock, whatever, you
Speaker A
don't really own it in the sense that they could change the terms at any moment, right? Like on the one extreme you could have June 13th where the US government issued a letter to Anthropic which caused Anthropic to literally turn
Speaker A
its model off. On another end of the spectrum you know when GPT went from GPD 4 to 5 people protested because the model just changed the way it spoke like it just something fundamentally changed about it. So like you don't have those
Speaker A
you don't want to necessarily have those problems and like that's what I mean by renting it. like you're in someone else's environment. You're in someone else's data center. You're under someone else's terms. And I think for like most
Speaker A
day-to-day tasks, people want consistency. People want reliability. People want to know how much it's going to cost and not have it variably change on me every single day. People don't want the rules of the road yanked from underneath them. And so in a world where
Speaker A
these openweight models are actually starting to achieve similar levels of performance on the everyday tasks that the frontier models are doing and at a fraction of the cost usually something like 35 to 50x cheaper. It seems like a
Speaker A
no-brainer. Like it seems like if you're a business who's running on these intelligent systems that you're going to want to go to the cheaper model that you know you can control, that you know you can pin your version on, that you know
Speaker A
your data won't go anywhere. So that's kind of what I mean by everyday task.
Speaker A
And an everyday task to me is like the kind of the boring stuff that a business needs to do like summarize an email or like manage a calendar or like all that kind of stuff that it's not scientific
Speaker A
research. For scientific research, go to the frontier models like they're making mathematical hypotheses that no human has ever done before, but I'm not going to hire an MIT PhD mathematician to run my calendar. So like like use the right
Speaker A
tool for the right time kind of thing. Rafie, how do you define that line between what makes sense to run on anthropic or open AI versus what makes sense to run locally on an open weight model.
Speaker A
Let's think of it as like a big picture thing. on the big picture like they call this the jagged frontier in the sense that like on certain tasks the frontier is going to be way better. The frontier model is going to be way better but on
Speaker A
like boring everyday path tasks they're not as much better or not that much better than an open model that you can go do. And so like if you again if you think about like these 80% tasks like your general everyday like you know
Speaker A
administrative kind of things versus long horizon hard engineering scientific those type of problems. Now that's very unspecific. So what I would say is you have to measure this for yourself and in some ways the one thing I recommend that
Speaker A
almost every single CTO do um and I talk about this all the time to CTO's engineers etc is like when you're using these intelligent systems especially if you're using it via an API call or a command line or stuff like that just
Speaker A
record and log everything so record and log the prompt that you give it record and log the response that you get back from it because then Those traces become your own personal evaluation set. You can then run those against a variety of
Speaker A
different models and see like what's the cheapest possible model you could use against your real data and what got you to basically similar level of performance. So like the the the everyday use case is going to be different for every single person. And
Speaker A
so what we need to all do is just like build a habit or build a set of tools that allow us to measure it for ourselves so we can each individually figure that out. But it's generally going to come down to like the boring
Speaker A
mundane stuff versus the crazy high scientific stuff. What about accuracy some for example or or quality? Because sometimes cost is not the only defining factor. for example, I don't if I have it sending emails or writing, I want the best
Speaker A
possible writing I can get. So like in those situations, you know, I think that like you can a fine-tune a model to actually be tuned to your writing style. So then you can then output it in your in your um voice. So I
Speaker A
think that's one potential option. And also I think that like you it is possible that like on a what's called a single shot. So like if you were to give a single prompt to a model and have it
Speaker A
do work that the frontier models might do better on that single shot like on the one one attempt but the the other more commodity models might get to the same level of quality. in fact is documented in a lot of cases they do if
Speaker A
you allow them to iterate a few times and in the case that I just said when it's 35 to 50x cheaper that means you actually have 35 to 50x times to try to iterate and get closer and still
Speaker A
maintain underneath that cost budget so you know you're trading a little bit of like incre incremental quality on a single line basis but then it could be refined you're also maybe trading a little bit of time but you know
Speaker A
especially on things that are running on a schedule running overnight like for example I use open weight models all the time to prepare me with briefings for the for the meetings I'm going to have the next day. So if you go into my
Speaker A
calendar, every single calendar event for me has a summary of just like here are the pertinent emails, here are the pertinent documents, stuff like that that ran overnight. Like I didn't need it to be real time. So I can run those
Speaker A
on really cheap commodity models because they'll just like iterate and loop on themselves back and forth. So I think there are many engineering ways to get around. Yeah, fable will probably do it better on the first try, but you don't
Speaker A
need the first try like in practical purposes. So at some point, as you say, you need to balance one wants one in all aspects of life. You balance quality and results against the cost cost against time against like when
Speaker A
you're going to run it. Like I just I want to live in a world where people can choose the right model at the right time at the right price point and things like that. And so like you know we don't live
Speaker A
in that world right now. It's significantly easier just to go use Opus Fable etc. But there are real tradeoffs like we've been talking about of like your data now lives on their servers.
Speaker A
You're under their update regime all those things. So like I want you to be that that choice. If you have a particular piece of useful thing you want to go do on the fable on the fable model go do it. But like for everything
Speaker A
else I want you to have options that you can then choose from. Let's take a question from Twitter. This is from Gus Beckdash who says, "Rafy, we talk a lot about the benefits of open weight, which is great, but what about the risks which
Speaker A
are even more great? How do we balance?" The hypothetical that most people talk about when it comes to these open weights is that, you know, the open weight models are only trailing the Frontier models by a couple of months,
Speaker A
especially the really big openw weight models like a KIK3 or a GLM53 or things like that. They're only trailing by a couple of months. And the concern they have is like whether or not you can use these models to do malicious things in
Speaker A
the world. Um, and so like my my response to that is like, you know, we one we live in this world right now. We live in this world where these models have been released and so regulating or trying to trying to remove the ability
Speaker A
to use them doesn't actually make us safer. In fact, it might make us less safe because there are just less people working with this technology, less people building with this technology. I think the thing we've learned from open
Speaker A
source in general is that the more eyes there are on the problem, the more potential solutions we have to the problem. If we just acknowledge we live in a world where openw weight models are now out there, I actually think we can
Speaker A
get to a safer world then the alternative which is basically trust that anthropic open AI gro etc are going to keep models safe for all of us which there's demonstrated proof that they can't do. So therefore I would say we
Speaker A
need more people working on the security side not less and that for me biases toward the open models being more out there. It's that this is an intractable problem essentially.
Speaker A
I mean it's a problem we've been living with already with the internet. Like if you think about the questions of signal for example the secure endto-end messaging protocol. Signal is more secure because it is open. More people have been able to poke at it. More
Speaker A
people have been able to find bugs. More people have been able to fix it. More people have been able to understand it.
Speaker A
Yes, there are trade-offs. Like people do use signal in malicious ways but the system is more secure because more people can poke at it. So it's a problem that we deal with every single day when it comes to open source, open networks,
Speaker A
open protocols but I think history has told us that openness actually lend leads itself to more security not less.
Speaker A
Here's a question from Swami Vajonathan who says, "Do you see AI models becoming closer to a commodity like electricity with open models continually closing the gap or will the frontier models keep a lasting edge the way package software
Speaker A
held its own against open source? And if models do commoditize, where is the moat for Frontier Labs? Is it specialized use cases?" All these could be trueer at the same time. So I think that there will always be a place for these specialized
Speaker A
frontier labs. Like we're always going to be pushing forward scientific discovery. We're always going to be pushing forward these huge societal questions in front of us. I mean look I mean like I sit in marvel of what these
Speaker A
open lab these labs are doing right now to mathematics. Like the fact that we are just basically going through all these completely unsolved mathematical problems is kind of bonkers. And imagine that world when applied to drug discovery. Imagine that world applied to
Speaker A
astronomy where we can start analyzing signals in ways that we've never been to analyze before. Hopefully one day applied to climate change, applied to all these different things. So there will always be a place for these frontier labs. And frankly, I think
Speaker A
there will be always ways for them to make a lot of money doing this. Um, but that being said, I think if you also buy my argument that these commodity models, these open weight models are starting to get in the everyday use case looking
Speaker A
more like infrastructure. Like instead of being a product, which is what these big frontier labs are doing, these open models are becoming more like infrastructure that just power the everyday world that we just don't pay attention to anymore. Like someone
Speaker A
deploys it and they walk away because it just keeps on running. And so I think like that's going to be the world that we live in where like these open models are just going to be embedded into banks. They're going to be embedded into
Speaker A
office spaces. They're going to be embedded into all the boring software that's on your computer because it's just going to be the cheaper mo more cost-effective way to go do it. Now when it comes to your question about modes, I
Speaker A
think it just it's a matter of like switching your mindset. Like right now, everyone thinks about the AI moat being the model itself, but there's so much more to building an AI stack. So if you look at what people like Mera Morati are
Speaker A
doing with her with her work on Inkling, like she's releasing an open model, but her business is based around building the infrastructure to help you fine-tune it, deploy it, use it, monitor it, etc.
Speaker A
If you look at what Sarah Hooker is doing with Adaption Labs, it's kind of the same deal. like she's not making making money off the model. She's making money off your ability to have recipes that help you compose models together in
Speaker A
novel ways so that you can deploy them and you can use them. So I think the moat will be shifting from model creation to model usage, the infrastructure around them, the adoption of them, the diffusion of them. And I
Speaker A
think there's going to be a whole sets of software stacks that need to happen there. Like if you just think about like on the coding models themselves right now there actually isn't very much of a moat on the coding models and in
Speaker A
fact like like coders like me switch every single day between claude gro um you know codeex or open code there's nothing that makes it sticky like they haven't found the mode but the mode will be when someone builds like network
Speaker A
effects on how coders can work together using these tools which no one has figured out yet. There are lots of opportunities for moes if you just get out of the mindset of the moat is not the model. The model is just one piece
Speaker A
of this way larger system that needs to be deployed. By the way, I mentioned this since you are the CTO of Misilla which is one of the pioneers of open-source software. All of this is way too complicated for most business people
Speaker A
for many many many IT people. It needs to be far simpler. We are just in an evolutionary path.
Speaker A
Like I think we're there's two problems that are on top of each other. One is that this technology is new, right? Like we're still in this formative stage where like the nerds are still tinkering and getting excited and deploying stuff
Speaker A
every single day. And then the second is just moving absurdly quickly. Like the fact that like in the one on one day I can see that like the Jev model which is like a decision-based model is released and three days later there's an open
Speaker A
source version like it's absurd how fast things are moving. But the exact same thing happened again in the internet just on a slower speed of just like when it would first happen the lampstack showed up and the nerds would like
Speaker A
deploy it and we would tinker with it and stuff like that until it became usable deployable and then their bosses got excited because it was way cheaper and they had control. So we just have to go through this learning arc again with
Speaker A
this new set of technology is just it's just being compounded by the fact that it's changing so fast. But I'm confident that like we will get through this. We just got to get through this moment.
Speaker A
Let's take another question and this is from Wayne Anderson who has been a listener in the past. And Wayne, it's great to see you. It's been a very long time. And Wayne Anderson says, "How do organizations balance the balance and the concern of
Speaker A
maintainability, a lack of real skills and experience to run these models internally and so on.
Speaker A
Whether we agree or not, we see this as a concern used used in less complex systems like just adopting and running Linux as a platform instead of Windows Server. There are still companies out there that aren't adopting open source
Speaker A
because of that kind of question. It's the the complexity issue. No, it's totally the complexity issue and I and I acknowledge it like you know I am I am preaching a dream but like we do need to close this implementation gap
Speaker A
and I mean like we see this right now across every single industry for every single piece of software, right? Like there's always a better option that's marginally cheaper. The question really comes down to what can be like the TCO
Speaker A
the total cost of ownership for your organization which is a partial dependent on like what's your skill set who are your vendors what are the consultancies you know who live down the street from you like all those things
Speaker A
matter when it comes to to that kind of like complicated thing so I would say that for for open-source openweight models there's multiple options here there's one option which is like yeah I'm going to download run it myself I'm
Speaker A
going to try to do everything I can internally. So Pinterest did this something like one year ago. There was a blog post by the printer CTO where they talked about how they saved $10 million that quarter because they moved to open
Speaker A
models that they've self-hosted as opposed to relying to the model providers. Now Pinterest is a Silicon Valley company that has engineers coming out of every single orifice and SRRES that I can manage and do anything you want. So not everyone can be a
Speaker A
Pinterest, but there are still opportunities in other places. Thompson Reuters just fine-tuned their own model based on their own internal data. So they I don't think they're not a Silicon Valley company, but yes, they have really great engineers. But I think even
Speaker A
if we go down the stage, like you don't need to download an open model yourself.
Speaker A
There are now starting to be neoclouds and API providers who are running these open models for you under pretty good license terms. Similarly to how you give your credit card to open a chat or anthropic in order to get tokens, you
Speaker A
can now give it to one of these neocloud providers and for like a 50th to a 35th of the cost you can get open tokens back coming back and forth. So you could then take your your open code or your codeex
Speaker A
and just point it to that instead of pointing it to cloud. So there are lots of different options. The failure here is that we the nerds need to do a better job of explaining it, making runbooks, making like doubleclick installers and
Speaker A
stuff like that. But I will say we've come so far in the last year, I can just only imagine where we're going to be one year ago, one year from now on what the usability is going to look like. And on
Speaker A
this topic of practical implementation advice Axaren says he uh what are the best platforms to run good openweight models at home for coding everyday help agentic tasks and so on. It's hard for a regular user to measure and compare them. So we have
Speaker A
to rely on pre-built options the way we do with browsers. Sure. The tech is moving fast and closed systems give a polished readytouse experience.
Speaker A
No, and that's the thing. I mean, we keep on coming back to it. Like your last three questions, Michael, have all been about usability in some way. Look, I can I'll say that like when I'm on a when I'm flying United and on United
Speaker A
Wi-Fi, which basically means I'm not online. Like, I use like the Quen 30B model. I run it on my MacBook Pro. I have a MacBook M5, I think. Um, and so with unified memory, I can run it on my
Speaker A
MacBook and I can run my coding agent against it. So I can just point open code at my laptop. The trade-off is my lap gets a little bit warmer than not, but you know, I can run it all on my on
Speaker A
my own machine. So on the single machine deployment case, there's actually a pretty good ecosystem out there, whether it be Olama, BLLM, um you know, I I forgot the name of the framework that the MacBooks have done to
Speaker A
like really optimize it, but there are a lot of frameworks to run these things on your machine itself. You know, even Mozilla back in the back in the day, I was going to say like less than a year
Speaker A
ago, we we created something called Llama File, which is a universal binary that packages a model along with it. So you can literally down that download that onto almost any computer architecture, this one file, and it'll execute on the file and start up a model
Speaker A
server for you. So there are lots of options for running it yourself. If you're willing to spend a little bit more money, then the Mac Studios are a really good platform. Like their their memory architecture and and their GPUs
Speaker A
make them a really good option. I have two Mac Studios and like I'm just like churning away on what they're doing every single day. If you're willing to spend a little bit more money, Nvidia has their DGX Spark, which is a
Speaker A
similarish from a if you squint hard, you could think of it as a similarish architecture, a GPU with a crap ton of memory associated with it. So, you can actually load fairly large models in it.
Speaker A
And the DGX actually works really well if you want to train or fine-tune models. So, I think there's a few different options under the like $5,000 US price point in order to actually be running these things yourself. But like
Speaker A
like again like if you're not doing anything super complicated, then most most modern computers will be just fine and running some version of an open model to do stuff locally for you. And then I I don't even get me started on
Speaker A
what the opportunities look like on mobile devices these days. Like those models are actually like accelerating their capabilities pretty fast. It's kind of amazing to watch.
Speaker A
I have a Windows computer here. Here's here's where you know I ask for my for my own psychotherapy. Uh I have a Windows computer with a 5090 graphics card. Okay. What's the best model I can run and and I am one to me the quality
Speaker A
if I'm doing something the quality matters more than anything. That computer is going to be memory limited more than anything else. It's like so there's amount of there's the amount of memory that's associated with the GPU. So that's going to limit the
Speaker A
amount the the size of the model you can connect but I would start with the Quen family of models like the Quen family of models like I can fit a 30B on my MacBook but you know I got multiple
Speaker A
gigabytes of RAM on this thing. Um, you know, I can't even tell you how much if I pull open the about in section, but on something like that, I would start with the I would start with the biggest Quinn
Speaker A
model you can fit, which is probably the 9B or something like that, and just start seeing what it can go do. Now, I'm I'm going to be very honest, like once you get below the 30, like the language
Speaker A
abilities and stuff like that are going to start degrading. So, like RAM is king in this world. Like you you need as much RAM as you can to just fit the model into memory. Now there's a lot of tricks
Speaker A
people have written a lot of open source software to actually like you know load um load models into memory incrementally. So like you execute a few layers then you dump it then you execute another few layers and you dump it. So
Speaker A
you're trading up speed to fit it in. So you could do a really big model on that machine like maybe one of the bigger Quinn models but you're going to be trading speed at that point for it. So
Speaker A
but I think for your purposes my purposes I would start with the Quen family models. Those are pretty good.
Speaker A
Let's go to another question that shifts us away from the technical discussion we've just been having to an enterprise perspective. And this is from Arcelon Khan on Twitter and X and he says, "Who should really care about cost? Is it
Speaker A
business or it? Is it neither of them? Both of them? and or or what is the optimum responsibility by percentage for business versus IT? It's like who's responsible? Is it business, the line of business or the IT folks?
Speaker A
In all the enterprises I talk to, it varies who really cares the most. Like in some businesses, I have the CFO talking to me because they're like, "Our token spend is out of control. We spent a million dollars this month alone on
Speaker A
tokens. How can I get that under control? what are the trade-offs I can make on some companies that talk to me.
Speaker A
It hits me up because they want to talk about control and where PII is going. So like they're less about cost sensitivity and they're more about we don't want business confidential IP to leave the premises. They want that to stay inside
Speaker A
because you know we did this legal review a few months ago. It turns out it might be the case, not yet tested in court, that if you send a piece of of business IP to one of the cloud
Speaker A
providers, you might no longer be able to call it a trade secret anymore because you've now disclosed it to one of the cloud providers. So like the IT teams usually talk the IT teams, the security teams, the CIO usually talks to
Speaker A
me about that where they were really concerned about like how do we manage our internal business processes and our internal business secrets. The CEO's office usually talks to me about questions around choice and and questions about resiliency and
Speaker A
reliability. Like they all saw like I'll say it again, they all saw the June July 13th issue or where Anthropic was just turned off and they're like we can't have that in our company. And like in fact like that that June 13th is not a
Speaker A
bigger deal because so let me back up. Fable was just released around that time and then the US government sended an order to Anthropic that said no non- US citizen can access this. They were just banning it which and Anthropic had no
Speaker A
other choice except to turn it off because they had no way to identify all their users in that way very quickly and it was off for something like 13 days.
Speaker A
We don't think of it as the biggest deal only because Fable was only out for maybe 2 days at that point. No real businesses had switched to it. If this happened a month into Fable's deployment, three months into Fable's
Speaker A
deployment, you would have taken out swaths of companies with that one order all of a sudden. And so CEOs talk to me about that all the time of just like how can we be more resilient to that kind of
Speaker A
failure? Like like what does it mean to have something redundancy and choice in there? So maybe to answer your question like it's not usually any single department. It's usually everyone has their own reasons for why they care about, you know, choice, resiliency,
Speaker A
cost structures, etc. And it's a matter of just trying to figure out how to reconcile all of those at any given time. All of them point to, in my opinion, just having a good backup plan.
Speaker A
Like if you're going to start somewhere, just even have a good backup plan of what happens if one of the providers turned off or what happens if there's a blip or what happens if cost goes out of control. So just having that good backup
Speaker A
plan as a first shot seems like a good place to start at least for the frontier model suppliers do want to be seen as core infrastructure like electricity or yeah but like sir but Michael I so I used to run infrastructure at Twitter
Speaker A
that was one of my responsibilities when I ran engineering and we used to build data centers you would never just run a single power line into a data center cuz like the day a squirrel chews through that line the squirrel's gonna have a
Speaker A
bad day, but your data center's offline. Like, you run like redundant power into your data center. You run redundant internet. You should be thinking about it in the exact same way when it comes to these model providers. Like, if
Speaker A
something goes wrong, what's my redundancy plan? No, it's absolutely true. Uh there was a period of time when Claude was was down.
Speaker A
Uh they were just having problems, technical problems. And if you build a lot of your business around working with the models, the model is solving problems, you know, not high-end math problems, but just running your business type of problems. It's like
Speaker A
your internet goes down, you're out of business. You're set out of a luck. I mean, this is why like I love these the open models like like my practical advice for most CTOs is never switch today. That's crazy. Like no more like don't do that
Speaker A
but at least start evaluating what it would take to get this backup plan in place. Like what's your fall back for the day that Anthropic has an outage?
Speaker A
What's your fall back for the day Chat GPD decides to like lower their token limits because they're over capacity?
Speaker A
Like what are you going to do? Start evaluating a backup plan today. So Wayne Anderson says, "When you think about openweight models at CIO scale, what are the cues you see in business invaluable things to solve that should cause an
Speaker A
executive to immediately say, I need to look at open weight." We were facing this even at Mozilla, right? So like at Mozilla like when our you know finance team is starting to use Fable all of a sudden to do financial
Speaker A
workflow analysis, that seems crazy to me. Like there is no reason that you need something that that's like basically trying to drive a Ferrari to Whole Foods. Like like you don't need that level of like power in order to
Speaker A
pull that off. So I think like there's general workflow that your common sense can tell you that shouldn't be that complicated. We should be able to try something simpler. And like you know the first thing I would try is like move off
Speaker A
Fable and move to Haiku and just see what it does before you even try moving to an open weight model. But like if you think about the general innovation curve that most like technical organizations go through first you're going to like
Speaker A
you know you know dick around for a bit to understand a problem and then you're going to start building a solution and then you're going to optimize it. And so like in that first like explore and just hack around. Frontier models are great
Speaker A
for that because you don't exactly know what you're doing. you just want to try all the different parameters and then you're going to start formalizing and writing some code and then once you get that code stable then you start
Speaker A
optimizing it right like engineers talk about this all the time it's like don't prematurely optimize so like in that optimization phase that's when you should start thinking like is there a cheaper thing I could do here can I run
Speaker A
this in a faster or slower way is there a different price point or a different model I can be using to answer this question so I like to think about it as part of that innovation and deployment ment curve that when you get to
Speaker A
production and before you hit the scale button, that's when you're going to be trying to eek out every single scent of your product, every single scent of your workflow, that's the time you should be really evaluating like, do I need a
Speaker A
frontier level model to do this work or can I just run a set of evaluations across all the other models? And if you did what I said to do earlier, which is like I record every single prompt, record every single output. You have
Speaker A
your own personal training and evaluation set you can use to make that choice. So I think like that's the way I would think about it. I wouldn't go into it off the bat unless you're like you're like a superpowered engineering
Speaker A
organization. I wouldn't go off the bat being like I need to find the right open model for this. I would start with the frontier and then just whittle my way down as I get toward production. I'm sorry to keep coming back to this, but
Speaker A
it's so it's just so complicated. It is complicated, you know, for normal for normal people who are not in the business of evaluating models. Sorry for interrupting you. This is not where enterprise folks want to spend their time.
Speaker A
I think it's different for every single enterprise. I think it's unfair to paint every single enterprise the same way.
Speaker A
Thompson Reuters is an enterprise company. I wouldn't say they're a technology company by any means, but they have figured out how to fine-tune a model and save a lot of money by going and doing it. So, I think, you know, we
Speaker A
last year we were talking about Pinterest. This year we're talking about Thompson Reuters. Next month, next year we'll be talk I don't want to name a company because that makes them sound bad, but like next month we'll be
Speaker A
talking about the next tier of companies down. So, I think we're just in this evolutionary curve and I'm trying to read what the future looks like. I agree with you, Michael. Today it's hard but is easier than yesterday and tomorrow
Speaker A
will be easier than today. Can you talk about the patterns that make open model deployments uh succeed or fail for for enterprises who want to go through this adoption process you're describing? What works? What doesn't work?
Speaker A
Most of it comes down to a question of talent on your team, whether time is on your side in order to try to figure this out. And then also a question of just like do you have access to your own data
Speaker A
sets that you can use to evaluate how these models are going to perform. And so like most companies that you know when we did the state of open source AI report the first version we asked all these enterprises did you try open-
Speaker A
source models and a v more than a majority of them said yes 60 70% you should look in the report state of opensource.ai AI. I don't remember the exact number, but a vast majority of them actually had tried open models, but
Speaker A
only something like 20% of them got into production. And when we when we drilled into why, they said a few things. One, too complicated, like you said. Two, we didn't have the talent on the team to go do it. Three, it's a variant or too
Speaker A
complicated, but like we couldn't figure out the right hardware software combinations to pull it off at scale.
Speaker A
And so like all those things are still at play right now. So if you are a midsized organization and you have like a good engineering team, if you're a midsize organization and you have a good vendor or consulting agreement with a
Speaker A
good vendor, then I think you are in the realm of people that can pull this off.
Speaker A
If you're a small medium-sized business or your business has nothing to do with technology, I agree with you. This is not the year to go try the open models yet. And in fact, but I would also say this is this is maybe the year to try
Speaker A
AI, but it's not the year to try open models. Next year it will get easier and then those we can keep on coming down the curve. But if you were on on the Fortune 500, Fortune 1000, like the big
Speaker A
companies who do have well staffed IT teams, they have relationships with really good vendors or any one of you can just call Mozilla. We would be interested in talking to you too. Like I think those companies are ready set
Speaker A
right now to go do it for everyone else. I hear you. It's complicated, not impossible, but it's again similar to the internet. Like when the internet first happened, everyone ran III servers and these days the majority of people do
Speaker A
not. But it took a few years to get to that point. Do you want to say a few words about tokconomics and the wild explosion of token costs as AI agents proliferate? That must be a factor here 100%. Like I mean this is actually one
Speaker A
of the reasons why people talk to me about switching away from the closed model providers is because the per token pricing is unpredictable, super variable. It's moving every single day.
Speaker A
Like in some ways it is getting cheaper but the the agents themselves are starting to explode on the number of tokens they're using. So it's kind of washing itself out. So for all those who are listening what's really what's
Speaker A
really happening is that per token basis that the models are emitting you know they charge you for every single token the model emits. It's actually even more complicated. They sometimes charge you for the tokens you put in and the tokens
Speaker A
you get out whether you hit the cash or not. But you can just maybe at a high level think about it as they charge you for a token that comes out of the system. And the problem the good thing
Speaker A
is those token prices have been falling. Every time a new model comes out though those token prices are fairly high. Um the coding agents all these agents that we're using though consume a crap ton of tokens. Like it's not it wouldn't be
Speaker A
unsurprising for me. In fact, I'm looking at one of my dashboards. Like on a regular task, it used something like I don't know 300,000 tokens to do a regular task. So like these tokens are being consumed at a fairly crazy rate.
Speaker A
Most people don't see this problem though because you you have a clawed subscription or you have a chat GPD subscription. So the tokens are kind of behind the scenes like you're they claim it's all you can eat, but it's not
Speaker A
really true. But like you're getting a lot of stuff coming at you. But the problem is once you exceed that boundary or if you're writing software that's calling these API calls directly and you're paying on a per token basis like
Speaker A
it's getting to be kind of nuts. The open- source models are again 30 to 50x cheaper on a per token basis. Um and they're fairly stable in their cost because once you deploy them you just know that this is how much it costs to
Speaker A
run it. Um it's not going to change depending on the state of the business.
Speaker A
And the other problem that we're we're going to see is that these big model providers, these big labs are going to IPO sometime in the next 12 to 18 months. They keep on saying they're IPO, delay the IPO, whatever. But, you know,
Speaker A
we saw this in ride sharing that when ride sharing first came out, it was fairly cheap for everyone to ride share.
Speaker A
Post ride sharing IPOs, the ride sharing cost went up because now these companies are accountable to their shareholders for profits. So right now a lot of these token prices when it comes to the big labs are being subsidized by the labs
Speaker A
themselves as a way to try to get you to use them more. And so like I'm I'm really worried about this entire dynamics of this system right now which again lead me to open weight models because then I can have predictability.
Speaker A
I can do real math around it. I can actually run a financial model around it. Whereas if I'm building my system on variable pricing of models like that it just seems crazy to me. It seems too hard.
Speaker A
I have the $200 Claude Anthropic Claude subscription. Yeah. And if I had to pay for the API cost, the actual token use, my my life with AI would be very different than it is right now.
Speaker A
One of the engineers at MAI, one of our subsidiary companies, measures this every single day because he just finds it so amusing. We all have at MAI to $200 subscription so we can write so we can use it for our coding purposes and
Speaker A
then he has a dashboard where he actually measures exactly how many tokens and then does the math and it comes out to like $6,000 a month if you had to pay for the tokens themselves. So like that difference between your $200
Speaker A
subscription and $6,000 a month post IPO is going to disappear. Like they're going to want that money to make to make their books look profitable. So like that's the risk that everyone should be staring at. Yeah, I actually have done
Speaker A
those measurements, too. It's pretty striking. It reminds me of the dot I'm old enough to remember this, the dot days where VCs were funding all of these companies. Like I remember I bought a table from furniture.com that was so underpriced what they paid
Speaker A
for it. It was, you know, and I thought it's so amusing, you know, this I have a VC funded table desk. Okay, so we have another question here. Uh this is from Chris Peterson who says, "What are the
Speaker A
business guard rails around how we use Genai in general and open-source openweight models in particular to make money? What can be copyright protected or sold on for profit?" If you look at the licensing terms of most of the
Speaker A
closed providers, they usually say something like we give you access to we give you a license to the to the content that we create for every single purpose we can think of or something crazy like that. Like the language is actually like
Speaker A
intricately worded because the law is a little bit unclear of exactly how this work. I'm not a lawyer my but my read of it is because it's unclear of how the how the law works. This is the reason
Speaker A
again why I like the open models because just the rules are a little bit clearer of exactly how this gets generated and work because you're not generating it on someone else's computer. You're generating them on your computer. You're
Speaker A
generating with software that you control the software that you are running in that ways again at the engineers view of the law. But I wrote a blog post about this. You can see it at owners not renters. um that the locally
Speaker A
run models gives you better IP protections in most cases and allows you to have more flexibility and the kinds of things you can do from their output than it does from the big closed providers. Now the big closed providers,
Speaker A
it also depends on the exactly which plan you're on and exactly which boxes you've ticked at every single given time. So there's a fairly complicated thing math that you have to do in your head. But, you know, I say this now to
Speaker A
almost every single one of my investor friends that if you're investing in a company, ask them like, are you doing all your work on Claude and which which tier of plan are you on? Like, you got to be on the highest tier plan with all
Speaker A
the right boxes checked. So, these are all the variables you got to be thinking about. So, again, not a lawyer, but I encourage everyone to really read these policies because it's wild how they've had to phrase it to make it all work.
Speaker A
Wayne Anderson comes back and he uh says inspection at scale for AI is a critical headache for the CIO.
Speaker A
And he says BCG published a report saying um the summary is this is important and how so the question therefore is how does open weight deployed at scale in multiple places get inspected and tracked at scale? Do we do we even know
Speaker A
yet? I think this is a per enterprise business question of just like what are the guardrails that you're putting in?
Speaker A
What's the logging procedures you're putting in? How are you auditing the system? How did you evaluate the models before you even deployed it to make sure you've guard rail correctly? And I do think it matters on the location by
Speaker A
location basis because like the data laws might be different if you're deploying in Europe versus if you're deploying in North America in United States specifically versus Canada. I think this is actually a very important business question that every single
Speaker A
enterprise needs to look at. But I don't think it is pertinent to just open weight models. I think this matters no matter which model provider you're using. Because if you're using one of the big model providers, then you're
Speaker A
almost certainly hitting an American data center which is going to have its own set of rules, its own set of guard rails, and then like if you're just relying on what Anthropic is doing for their evaluation, I think you're not
Speaker A
thinking about it correctly. Like you need to be running that as well so you can understand what your exposure could look like depending on what the model response looks like. So I think that this is a great question that you should
Speaker A
be asking regardless of open versus closed. like you just need to understand that as you in your business. The next question again from Gus Beckdash who says OpenAI etc are building data centers like crazy because they know the moat is not the
Speaker A
model it's running models as you said so they agree with you Rafie are they wrong and are they underestimating the openweight threat they're not underestimating open wave threat in fact I would say they are they see the openweight threat right in front
Speaker A
of them and so like you know people like Mosilla that has a billion dollars in the bank and we are trying to figure out how to deploy that money to make this open weight open source ecosystem work better. That's the amount of money that
Speaker A
those companies are going to spend in marketing this year alone. And so like they do they do not underestimate this threat. They see this threat headon which is why they're also in policy conversations, regulatory conversations, etc. to try to lock themselves in and
Speaker A
try to edge out the open providers. So I I think they understand the threat and they are behaving as if they understand the threat.
Speaker A
Many of the best open models come from Chinese labs. What does that mean for a company in the US or in Europe?
Speaker A
We're not pro or against Chinese labs making models, but I am concerned about geographic concentration risk in the same way that I'm concerned that both OpenAI and Anthropic are in the US and therefore are under the regulatory regime of the US. If all the open
Speaker A
providers are in China and under one regulatory regime, it's just a risk for the entire industry overall. So, you know, Mozilla, we've been starting to talk with and have conversations with open model providers all around the world, whether it be, you know, like the
Speaker A
ALF alphas in Germany or the coheres in in Canada, the reflections in the US, like there are the apparatuses in Switzerland. there are starting to be a a broader ecosystem of people providing openwave models which I think is a good
Speaker A
thing because it means that we are not under a single regulatory regime. Now that being said you you can't undownload a model. So yeah it should be concerning about moving forward on Chinesebased models but all the ones that are already
Speaker A
out there on hugging face etc. you can't und them. So like you know maybe download the one that you're going to run on your business and keep it on like keep it on your own discs. don't rely on
Speaker A
someone else to keep it for you. So they can always have it. Now I think a common fallacy that a lot of people have is like okay but I'm running a Chinese model does that mean the Chinese government etc etc etc. And the answer
Speaker A
is generally no right think of it as like I'm now running this on my machine is under my control you can then even inspect like is anything leaving your network boundary that you don't know about? Like it's just not the way that
Speaker A
these things work. Now, I wouldn't go and say run your business on a Chinese model that's hosted on Chinese infrastructure without fully understanding the risks of what that might mean to your business. But those are the risks you should be concerned
Speaker A
about, not the model itself that you might download and run. All right. Critics say that once a model is open, it can never be recalled and its safety controls can be stripped out. Where do you draw the line on what should be open
Speaker A
or not? I firmly believe that having open models in the world and allowing more and more people to use, tinker, understand, etc.
Speaker A
means that we'll build better open guard rails, better open monitoring systems, better open safety system to detect, to react, to manage, etc. all these things. I think if we get in a world where we don't allow open models
Speaker A
to be proliferated in some way that will actually be less secure because then we're only going to be relying on the big labs to to effectively protect us all. And as demonstrated by the hugging face stuff, the Australian hack, the
Speaker A
like all this stuff like they're not capable of doing it. So like I actually want more security researchers. I want more academics. I want more people working on these security problems. I think the only way to do that is
Speaker A
actually to have these open models out there. A Sargisian comes back and he said, "Do you really think that when these companies go public that tokens will become much more expensive?" 100%.
Speaker A
There is nothing which leads me to believe that that will not be the case.
Speaker A
Now, the only thing that might be counterveiling to that is that as these models as we develop more and more models, they are getting more efficient.
Speaker A
So we're able to do more with less hardware, but then you know honestly but then we'll just drive up what the capability of those models are and those token costs will go up again. So, I 100% believe we live in a subsidized regime
Speaker A
right now. Like you were saying, Michael, on VC subsidizing your car, your your tables, what I'm saying about ride sharing in the preIPO world of ride sharing after IPO, it's not going to happen on day one, but you're going to
Speaker A
see it in quarterly reports of just like we figured out a way to ek out more profit this way and it's going to come down to the cost of those tokens at some point.
Speaker A
Uber has become quite expensive. That's exactly my point. Like I I mean I having been one of the people who worked at Uber preip like I remember the cost of Uber back then like we subsidized an entire generation of kids effectively to
Speaker A
use Uber and now Uber is like you know is it expensive but like Uber you're now seeing the true cost of that system effectively. Final thoughts, advice, uh maybe comment on lock in.
Speaker A
Your report says that lockin has moved from the model to the context that surrounds it. How can companies overcome this form of lock in?
Speaker A
The biggest theme of the latest version of our report, so state of opensource.ai, we we rev it every couple of months. is a fairly live document to help you understand how open source AI is sort of evolving in the world. But
Speaker A
the latest version basically says that open models have moved from a thing that you can just download to something now that you can inspect, fine-tune and make your own in some way. And that means that you can then actually own and
Speaker A
control this infrastructure in ways that just wasn't possible even just a few months ago. Like the infrastructure to do fine-tuning has super evolved just in the last 3 months. The ability to deploy a model that started off as one of the
Speaker A
Quen models like we were talking about earlier and now can speak with my voice or can make be a decision-based model like a Jev model. It's like it's becoming significantly more powerful.
Speaker A
And so I think because of that arc we can get to a place that you're not locked into any one of the big providers anymore. the big provider's business model is for you to get less locked in as possible. You can see this right now
Speaker A
if you spend a week using Claude and then you switch to chat GPT. It's like talking to an amnesiac all of a sudden because all your memory got sucked into the clawed system. And so in a world where we can now download, fine-tune,
Speaker A
deploy ourselves, it means also our memory, our data, etc. can stay on our side which provides us a leverage again.
Speaker A
we can now have choice about what we want to deploy where we want to deploy it and prevent that lock in from happening. So, I think there's a lot of work to still be done. Like you we've been talking about like it's not easy
Speaker A
yet. It's easier than last year. It will be easier next year, but like we have a lot of work to do, but we're on that verge that if you are like an AI native company and you're not evaluating openw
Speaker A
weight models, you're doing something wrong. And if you're a company who's starting to think about AI, great. In maybe two or three months, you should start thinking about the open weight models. like get through that adoption curve first and then start thinking
Speaker A
about like we're at that point in the conversation. I think Rafie Creoran is CTO of Misilla and I must say a very eloquent advocate for open-source AI openweight models. Rafie, thank you so much for being a guest again on CXO
Speaker A
Talk. Thank you, Mike. It's been such a pleasure. Everybody, thank you for watching especially you folks who ask such great questions. Before you go, subscribe to the CXO Talk newsletter and we'll keep you up to date on we have unbelievable
Speaker A
guests and shows coming up. [music] So check it out and thank you so much everybody and we'll see you again soon.
Topics:open-weight modelsopen-source AIMozillaself-hostingfine-tuningfrontier modelsAI costsdata privacyLLM deploymentRaffi Krikorian
