Mozilla CTO:
Open Source and Open Weight AI, Rent or Own?
Companies are deciding which AI work to keep on frontier models and which to move to open weight models they control. Mozilla's CTO explains the trade-offs and how to test models and plan a fallback.
Companies are starting to run everyday AI work on open weight models they can download, fine-tune and deploy themselves. Raffi Krikorian is Chief Technology Officer of Mozilla, which makes Firefox and publishes the State of Open Source AI report. The conversation examines what separates open weight from open source models, where renting the frontier ends and owning everyday AI begins, and the different worries finance, IT and CEOs bring him.
Key points:
- Open weight models handle everyday tasks at 35 to 50 times lower cost, and a company's own logged prompts and responses show which model fits.
- He tells CTOs to keep today's provider and start evaluating a backup plan for an outage, a price jump or a lower usage cap.
- Big labs subsidize token prices today, he expects them to rise after the IPOs, and the cost of a deployed open weight model holds steady.
Krikorian says which companies are ready: large firms with well-staffed IT and strong vendor relationships can start now, and smaller or non-technical ones should wait a year. Mozilla's own survey found that a large majority of enterprises had tried open models and far fewer ran them in production. He argues that openness makes AI safer because more people can inspect and fix models, and that keeping memory and data in-house prevents the next lock-in.
Episode Participants
Raffi Krikorian is the Chief Technology Officer at Mozilla, where he leads the organization's open-source AI strategy. His career spans some of the most consequential technology deployments of the past two decades: VP of Platform Engineering at Twitter, Director of Uber's Advanced Technologies Center, where he launched the first commercial self-driving fleet, and first-ever CTO of the Democratic National Committee. He also writes Owners Not Renters, a Substack on open source AI, and contributes regularly to The Atlantic.
Michael Krigsman is a globally recognized analyst, strategic advisor, and industry commentator known for his deep business transformation, innovation, and AI leadership expertise. He has presented at industry events worldwide and written extensively on the reasons for IT failures. His work has been referenced in the media over 1,000 times and in more than 50 books and journal articles; his commentary on technology trends and business strategy reaches a global audience.
Chapter summaries
What is the difference between open weight and open source models?
Krikorian likens an open weight model to “a compiled binary”: you can download and run it without knowing what is inside, while open source shows everything that went into it. There are not many open source models, he says, especially ones that rival frontier models. If a model refuses a question, about Tiananmen Square for example, open weights cannot show you why; true open source shows where that data was censored or changed. Running an open weight model yourself still gives you control: it will not change until you upgrade it, and your data does not leave your computer. Watch this part 0:14
Why should everyday AI run on models a company controls?
Because a company that calls a provider’s model does not really own it, he says: “They could change the terms at any moment.” Mozilla’s phrase for the idea is “owners, not renters”. When GPT went from GPT-4 to 5, people protested because “something fundamentally changed about it.” For most everyday tasks, such as summarizing an email or managing a calendar, he thinks people want consistency, reliability and a known cost; scientific research belongs on frontier models. Asked by Michael Krigsman about quality, he says cheaper models might reach the same level if allowed to iterate a few times. His own overnight meeting briefings run on open weight models. Watch this part 5:11
Do open weight models make the world less safe?
He argues the opposite: the models are already out, so regulating them or trying to remove access “doesn’t actually make us safer”. Restrictions might even make the world less safe, he says, with fewer people working on the technology. The worry most people raise, in his telling, is that big open weight models trail the frontier by only a couple of months and could be used for malicious things. Open source has taught, he thinks, that more eyes bring more potential solutions, and Signal is “more secure because it is open”, though some people use it maliciously. He wants more people working on security. Watch this part 12:21
If models become commodities, where does the advantage move?
Krikorian thinks “the moat will be shifting from model creation to model usage”. He expects frontier labs always to have a place in scientific discovery, and ways to make a lot of money. Open models, he thinks, are starting to look more like infrastructure that will sit inside banks, offices and everyday software. Mira Murati releases Inkling as an open model, but her business is infrastructure that helps customers fine-tune, deploy and monitor it. Sara Hooker’s Adaption Labs makes money from recipes for composing models. Coding models have little moat, since coders like him switch daily; the moat will come from network effects that “no one has figured out yet.” Watch this part 14:45
How hard is it to run open models, and on what hardware?
Running them is hard for now, he acknowledges, though for simple work most modern computers will do. In his words, “I am preaching a dream, but we do need to close this implementation gap.” For a little more money, he points to Mac Studios, of which he has two, and Nvidia’s DGX Spark, among options under five thousand dollars. Pinterest moved to self-hosted open models something like a year ago, saving ten million dollars that quarter by its CTO’s account, though not every company has Pinterest’s engineers. A company need not download a model at all, since NeoClouds and API providers are starting to host open models for you. Watch this part 18:28
Who cares most about AI model costs, business or IT?
It varies by company, he finds: some CFOs come to him over token spend, one at $1 million in a month. At some companies, IT, security teams and the CIO come to him about where personal data and confidential business IP go. It might be the case, he says, “not yet tested in court”, that IP sent to a cloud provider can no longer be called a trade secret. CEOs ask about resilience, having seen June 13th, when, in his account, a US government order led Anthropic to turn Fable off for something like 13 days. In his opinion, all of these point to “just having a good backup plan”. Watch this part 27:42
When should an enterprise move work to open weight models?
Near production, Krikorian says, “before you hit the scale button”; unless a company is a superpowered engineering organization, he would start with frontier models. At Mozilla, using Fable for the finance team’s workflow analysis struck him as “trying to drive a Ferrari to Whole Foods”; he would try Haiku first. Only something like 20 percent of the enterprises surveyed for Mozilla’s first report got open models into production; they cited complexity, missing talent and hardware-software combinations. Fortune 500 and Fortune 1000 companies, he thinks, are ready now, while for small and medium-sized businesses this is maybe the year to try AI, not open models. Watch this part 32:51
Is per-token pricing a risk for AI budgets?
He thinks so: per-token pricing is “unpredictable, super variable”, and while token prices have been falling, agents are starting to use so many more tokens that “it’s kind of washing itself out.” One regular task on his dashboard used something like 300,000 tokens. Subscriptions hide this, since “the tokens are kind of behind the scenes”, but heavy users and API callers pay per token. At Moz AI, an engineer’s dashboard shows the $200 subscription works out to about $6,000 a month if paid by the token. That difference, he says, “post-IPO is going to disappear”. Michael Krigsman, who has made the same measurement, likened the gap to the dot-com days. Watch this part 38:51
Can companies own what AI creates, and inspect models at scale?
In his read, offered with “I’m not a lawyer”, locally run models give better IP protection in most cases, and inspection at scale is a question for each enterprise. Most of the closed providers word their licenses intricately because the law is “a little bit unclear”. The rules are a little clearer, he says, when you generate the output on software you control. He tells investor friends to ask which tier of Claude plan a company is on. Answering Wayne Anderson, he lists guardrails, logging, auditing and model evaluation, under data laws that might differ by location. In his view, “This matters no matter which model provider you’re using”. Watch this part 43:13
Should a company run open models from Chinese labs?
Krikorian sees the risk in Chinese hosting, not the model: on your own machines it is under your control, and on government access his answer is “generally no”. Geographic concentration also worries him: if all open providers were in China and under one regulatory regime, “it’s just a risk for the entire industry overall”. Since future Chinese-based models are a concern and “you can’t undownload a model”, he suggests a business might keep the one it runs on its own disks. To critics who fear open models cannot be recalled, he replies that openness builds better guardrails. The big labs, he says, see the open weight threat “head on”. Watch this part 47:00
Will token prices rise, and how can a company avoid lock-in?
He is 100 percent sure they will, and his way out of lock-in is open models a company can fine-tune and run itself, though that is not easy yet. Efficiency gains might offset some of the price rise, but he expects better capability to push token costs up again. After a week on Claude, switching to ChatGPT is “like talking to an amnesiac”, because the memory stays with Claude. An AI-native company not evaluating open weight models is “doing something wrong”, he says. A company just starting with AI should get through adoption first and consider them in maybe two or three months. Watch this part 51:08
Chapters
Audience questions
From LinkedIn
- 19:53 Wayne Anderson: how organizations handle maintainability and the lack of skills to run models internally.
- 23:02 Aik Sargsian: the best platforms to run open weight models at home for coding and agentic tasks.
- 32:51 Wayne Anderson: the cues that should make an executive look at open weight models.
- 45:22 Wayne Anderson: how open weight models deployed at scale get inspected and tracked.
- 51:08 Aik Sargsian: whether tokens will become much more expensive once these companies go public.
From X
- 12:21 Gus Bekdash: how to balance the risks of open weight models against their benefits.
- 27:55 Arsalan Khan: the split of responsibility for AI cost between the business and IT.
- 47:00 Gus Bekdash: whether labs building data centers are underestimating the open weight threat.
From the audience
Mentioned in this episode
Mozilla
- 0:14 State of Open Source AI report: about 20 percent of the enterprises in its first survey had open models in production.
- 24:18 llamafile: one of many ways to run a model yourself: one file, almost any computer.
- 42:16 Moz AI: where an engineer’s dashboard prices the $200 subscription at about $6,000 a month in tokens.
- 44:17 Owners, not renters: his post arguing locally run models give better IP protection in most cases.
Other companies and tools
- 1:10 Fable, Opus, ChatGPT: Anthropic’s closed models and OpenAI’s product, all running only on their own servers.
- 5:43 Grok: one of the closed models you rent; the terms can change at any moment.
- 5:55 GPT-4 and GPT-5: his example of how a provider’s model update can upset its users.
- 12:38 Kimi K3, GLM-5.3: the big open models people worry about, a couple of months behind the frontier.
- 14:05 Signal: the secure messaging protocol, his example of openness making a system more secure.
- 17:01 Inkling: Mira Murati’s open model, his example of a business built on infrastructure around it.
- 17:18 Adaption Labs: Sara Hooker’s work, earning from customers’ ability to compose models in novel ways.
- 17:56 Claude, Codex, OpenCode: coding tools he switches among daily, along with Grok; nothing makes them sticky yet.
- 19:18 LAMP stack: his internet-era example: nerds deployed it first, then bosses saw it was cheaper.
- 20:19 Linux, Windows Server: the same skills worry; he says total cost of ownership, skills included, decides.
- 21:21 Pinterest: saved $10 million in one quarter on self-hosted open models, its CTO wrote.
- 21:55 Thomson Reuters: not a technology company, in his view, yet it saved a lot by fine-tuning.
- 23:34 United: on its Wi-Fi he runs a local model and codes without the cloud.
- 23:34 Qwen: the open model family he suggests starting with, the biggest that fits the machine.
- 23:51 MacBook Pro: runs Qwen 30B with unified memory; a laptop serves his coding agent.
- 24:04 Ollama, vLLM: frameworks in what he calls a pretty good ecosystem for running models on one machine.
- 24:50 Mac Studio: a really good option, he says, because of their memory architecture and GPUs.
- 25:04 Nvidia DGX Spark: he loosely likens it to a Mac Studio and says it suits training and fine-tuning.
- 26:01 5090 graphics card: in Michael Krigsman’s Windows PC; its memory limits the size of model it can run.
- 33:35 Haiku: the simpler first move off Fable for routine work, before any open weight model.
- 38:42 IIS: the early-internet default most dropped over a few years: complicated, not impossible.
- 43:05 Furniture.com: Michael Krigsman’s dot-com memory of VC-subsidized prices, his parallel to today’s token prices.
- 45:22 BCG: its report, cited by Wayne Anderson, says inspection at scale for AI is important.
- 48:28 Aleph Alpha, Cohere, Reflection, Apertus: open model providers outside China, so the ecosystem is not under one regulatory regime.
- 49:00 Hugging Face: where released models already sit; a model cannot be undownloaded, so maybe keep your copy.
- 52:06 Uber: subsidized before its IPO, true cost now: the path he expects for tokens.
In this episode
Raffi Krikorian: In some businesses, I have the CFO talking to me because they're like, our token spend is out of control. We spent $1 million this month alone on tokens. How can I get that under control? What are the trade-offs I can make? On some companies that talk to me, IT hits me up.
Open weight versus open source models
Michael Krigsman: Raffi Krikorian is CTO of Mozilla, which makes Firefox and publishes the State of Open Source AI report. What are open-source and open-weight AI models?
Raffi Krikorian: If you're confused, you are not the only one. So let me start there. 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 models these days. And what they are is think of it as a compiled binary. It is a model you can download.
It's a file or a series of files that you can run on your computer, your laptop, your server closet, your GPU, and your data center. Think of these open weight models as just a downloadable file that you can execute. The thing is, you don't really know what's inside. That's why I kind of compared it to a compiled executable.
You don't see the source code, but you can download it and you can run it, which is very different than, you know, an Anthropic's Fable or Opus, or an OpenAI's ChatGPT, which runs on their servers. You have no way to run it on your own servers. So that's the open weight models. You can see the weights, but the weights don't really mean much to most people.
And then an open source 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 entire thing that went into the model itself. So you can inspect it end to end and understand truly what's in the thing. Now, there aren't that many open source models out there, especially ones that can rival the frontier-level models. 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. You can then fine-tune them. You can then base another model and train another model 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 asked your model a question, say, for example, you were to ask a model about Tiananmen Square and it refused to answer, you 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-training, pre-training they censored that piece of data out or changed that data. So that's the difference between the open source and open weight models. The names are getting confusing. A lot of people conflate them together, but if you want to be really precise, that's the difference.
Michael Krigsman: So with the open weight models, what are the advantages to downloading the with the weights. I mean, you're not going to go in with a text editor and start changing the weights of the models.
Raffi Krikorian: 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, you have no full control. They might decide to take it down one day, they might decide to update it one day, the balloon might become a different color blue.
But if you download it and you run it yourself, then you know that you've pinned to that version, it's not going to change. This is the thing you are running from here on out until you personally upgrade it. There are lots of detriments to that though. Now I'm running a computer, I'm actually have to be like an SRE. I might be running a bunch of GPUs. I got paying my power bill. There are downsides to it, but the upsides are the things like you can pin it.
I know for sure where my data is going. If I send my data to Anthropic or OpenAI. I'm just going to name those two just because they're big platform providers. My data now sits on their servers. There's lots of reasons you wouldn't want that. If you are in a HIPAA environment, or if you're in a regulatory environment, or if you just don't want anyone else to know stuff about you, if 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 on your 4th point, Michael, is you can and then what's called fine-tune the open model. So yeah, you can't go in there. I mean, you could, but you just probably do more damage than anything else.
You're not going to go in there and change the weights, but you could train another layer of the network that'll take the open model and you could 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 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 contain their proprietary data to bend the model into answering the way that you want it to answer.
You can't do that with the closed models, but you could do that with something that you download.
Rent the frontier, own everyday AI
Michael Krigsman: Raffi, you have written that renting the frontier is fine, but everyday AI is something that you should own. What did you mean?
Raffi Krikorian: The phrase that we like to use around Mozilla is something that we call owners, not renters. What I really mean by that is, do you have full control over the intelligence that powers your day-to-day application environment, and in the current way that most AI systems work, where you call OpenAI, you call Anthropic, you call Grok, whatever, you don't really own it in the sense that they could change the terms at any moment.
On the one extreme, you could have June 13th, where the U.S. government issued a letter to Anthropic, which caused Anthropic to literally turn its model off. On another end of the spectrum, you know, when GPT went from GPT-4 to 5, people protested because the, it just, something fundamentally changed about it. So you don't have those, you don't want to necessarily have those problems. And that's what I mean by renting it. 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 most 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. Yeah, I think that's a really good point.
And so in a world where these open weight 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 no-brainer. 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 your data won't go anywhere.
So that's kind of what I mean by everyday tasks. And an everyday task to me is the kind of the boring stuff that a business needs to do, like summarize an email or like manage a calendar or all that kind of stuff that it's not scientific research. For scientific research, go to the frontier models. 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 use the right tool for the right time kind of thing.
Michael Krigsman: Raffi, how do you define that line between what makes sense to run on Anthropic or OpenAI versus what makes sense to run locally on an open weight model?
Raffi Krikorian: Let's think of it as a big picture thing. On the big picture, they call this the jagged frontier in the sense that on certain tasks, the frontier is going to be way better. The frontier model is going to be way better. But on boring everyday tasks, they're not as much better or not that much better than an open model that you can go do.
And so, again, if you think about these 80% tasks, like your general everyday, you know, administrative kinds of things versus long horizon, hard engineering, scientific, those type of problems. Now, that's very unspecific. So what I would say is do you have to measure this for yourself?
And in some ways, the one thing I recommend that almost every single CTO do, and I talk about this all the time to CTOs, engineers, etc., is 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 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 different models and see what's the cheapest possible model you could use against your real data and what got you to basically a similar level of performance. So the everyday use case is going to be different for every single person. And so what we need to all do is just 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 the boring, mundane stuff versus the crazy-high scientific stuff.
Michael Krigsman: What about accuracy, for example, or quality? Because sometimes cost is not the only defining factor. For example, if I have it sending emails or writing, I want the best possible writing I can get.
Raffi Krikorian: In those situations, I think that you can, A, fine-tune a model to actually be tuned to your writing style so then you can output it in your voice. I think that's one potential option. And also, I think that, you know, it is possible that on what's called a single shot. So if you were to give a single prompt to a model and have it do work, that the frontier models might do better on that single shot, like on the one attempt.
But the other more commodity models might get to the same level of quality. In fact, it's documented in a lot of cases they do if 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 maintain underneath that cost budget. So, you know, you're trading a little bit of 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, especially on things that are running on a schedule, running overnight, for example, I use open weight models all the time to prepare me with briefings for the meetings I'm going to have the next day. So if you go into my 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. I didn't need it to be real time.
So I can run those on really cheap commodity models because they'll just 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 need the first try, in practical purposes.
Michael Krigsman: At some point, as you say, you need to balance. One wants, in all aspects of life, you balance quality and results against the cost.
Raffi Krikorian: The cost, against time, against when you're going to run it. 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. We don't live in that world right now. It's significantly easier just to go use Opus, Stable, et cetera, but there are real trade-offs, like we've been talking about, of your data. Your data now lives on their servers. You're under their update regime, all those things. I want you to be able to have that choice.
If you have a particular piece of useful thing you want to go do on the Fable model, go do it. But, for everything else, I want you to have options that you can then choose from.
Openness leads to more security, not less
Michael Krigsman: Let's take a question from Twitter. This is from Gus Bekdash who says, Raffi, we talk a lot about the benefits of open weight, which is great. But what about the risks, which are even more great? How do we balance?
Raffi Krikorian: 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, especially the really big open weight models like a Kimi K3 or a GLM-5.3 or things like that. They're only trailing by a couple of months. And the concern they have is whether or not you can use these models to do malicious things in the world.
And so my response to that is you know 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 remove the ability 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 that we've learned from open 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 open weight models are now out there. I actually think we can get to a safer world than the alternative, which is basically trust that Anthropic, OpenAI, Grok, 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 need more people working on the security side, not less. And that, for me, biases toward the open models being more out there.
Michael Krigsman: It's a this is an intractable problem, essentially?
Raffi Krikorian: I mean, it's a problem we've been living with already with the internet. If you think about the questions of Signal, for example, the secure end-to-end messaging protocol, Signal is more secure because it is open. More people have been able to poke at it. More people have been able to find bugs. More people have been able to fix it. More people have been able to understand it. Yes, there are trade-offs. 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've dealt with every single day when it comes to open source, open networks, open protocols. But, I think history has told us that openness actually lends, leads itself to more security, not less.
The moat shifts from model to usage
Michael Krigsman: Here's a question from Swami Vaidyanathan 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 packaged software held its own against open source? And if models do commoditize, where is the moat for frontier labs? Is it specialized use cases?
Raffi Krikorian: All of these could be true at the same time. So I think that there will always be a place for these specialized frontier labs. 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, I sit in marvel of what these open labs, these labs are doing right now to mathematics. 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 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 there will be always ways for them to make a lot of money doing this.
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 to looking more like infrastructure, 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.
Someone deploys it and they walk away because it just keeps on running, and so I think that's going to be the world that we live in, where these open models are just going to be embedded into banks. They're going to be embedded into 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, more cost-effective way to go do it.
Now, when it comes to your question about moats, I think it just, it's a matter of switching your mindset. 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 Mira Murati are doing with her work on Inkling, 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.
If you look at what Sara Hooker is doing with Adaption Labs, it's kind of the same deal. She's not making money off the model. She's making money off your ability to have recipes that help you compose models together in 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 think there's going to be a whole set of software stacks that need to happen there. If you just think about on the coding models themselves, right now there actually isn't very much of a moat on the coding models. And in fact, coders like me switch every single day between Claude, Grok, you know, Codex or OpenCode.
There's nothing that makes it sticky like they haven't found the moat, but the moat will be when someone builds network effects on how coders can work together using these tools, which no one has figured out yet. So there are lots of opportunities for moats if you just get out of the mindset of the moat is not the model. The model is just one piece of this way larger system that needs to be deployed.
Open models are getting easier to run
Michael Krigsman: By the way, I mentioned this since you are the CTO of Mozilla, which is one of the pioneers of open-source software. All of this is way too complicated for most businesspeople, for many, many, many, many IT people. It needs to be far simpler.
Raffi Krikorian: We are just in an evolutionary path. I think there are two problems that are on top of each other. One is that this technology is new. We're still in this formative stage where the nerds are still tinkering and getting excited and deploying stuff every single day. And then the second is just moving absurdly quickly. The fact that in the one, on one day I can see that the Jev model, which is like a decision-based model, is released, and three days later there's an open-source version.
It's absurd how fast things are moving. But the exact same thing happened again in the internet, just on the slower speed of just like when it first happened, the LAMP stack showed up, and the nerds would deploy it, and we would tinker with it and stuff like that until it became usable, deployable. Then their bosses got excited because it was way cheaper and they had control. We just have to go through this learning arc again with this new set of technologies.
It's just being compounded by the fact that it's changing so fast. I'm confident that we will get through this. We just got to get through this moment.
Michael Krigsman: Let's take another question. 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 and the concern of maintainability, a lack of real skills and experience to run these models internally, and so on? Whether we agree or not, we see this as a concern 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 because of that kind of question. It's the complexity issue.
Raffi Krikorian: No, it's totally the complexity issue, and I acknowledge it. I am preaching a dream, but we do need to close this implementation gap. We see this right now across every single industry for every single piece of software. There's always a better option that's marginally cheaper. The question really comes down to what can be the TCO, the total cost of ownership for your organization, which is a partial dependent on what's your skill set? Who are your vendors? What are the consultancies? You know, who lives down the street from you?
All of those things matter when it comes to that kind of complicated things, so I would say that for open source, open weight models, there's multiple options here. There's one option which is like yeah, I'm going to download, run it myself. I'm 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 Pinterest CTO where they talked about how they saved ten million dollars that quarter because they moved to open models that they've self-hosted as opposed to relying to the model provider. Now, Pinterest is a Silicon Valley company that has engineers coming out of every single orifice and SREs that can manage and do anything you want. So not everyone can be a Pinterest, but there are still opportunities in other places. Thomson Reuters just fine-tuned their own model based on their own internal data.
So they, I don't think they're not Silicon Valley company, but yes, they have really great engineers. But I think even if we go down the stage, you don't need to download an open model yourself. 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 OpenAI or Anthropic in order to get tokens, you 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 OpenCode or your Codex and just point it to that instead of pointing it to Claude. 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 double-click installers and 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.
Michael Krigsman: And, on this topic of practical implementation advice, Aik Sargsian says, what are the best platforms to run good open-weight 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 to rely on prebuilt options the way we do with browsers. Sure. The tech is moving fast and closed systems give a polished, ready-to-use experience.
Raffi Krikorian: No, and that's the thing. I mean, we keep on coming back to it. Your last three questions, Michael, have all been about usability in some way. Look, I can, I'll say that when I'm on a, when I'm flying United and on United Wi-Fi, which basically means I'm not online, I use the Qwen 30B model. I run it on my MacBook Pro. I have a MacBook M5, I think. And so with unified memory, I can run it on my MacBook and I can run my coding agent against it.
So I can just point OpenCode 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 own machine. So on the single machine deployment case, there's actually a pretty good ecosystem out there, whether it be Ollama, vLLM. You know, I forgot the name of the framework that the MacBooks have done to 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 day, I was going to say less than a year ago, we created something called llamafile, which is a universal binary that packages a model along with it. So you can literally download that onto almost any computer architecture, this one file, and it'll execute on the file and start up a model 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.
Their memory architecture and their GPUs make them a really good option. I have two Mac Studios, and I'm just 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 similar-ish. From a if you squint hard, you could think of it as a similar-ish architecture: a GPU with a crap ton of memory associated with it, so you can actually load fairly large models in it.
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 five thousand dollar U.S. dollar price point in order to actually be running these things yourself. But again, if you're not doing anything super complicated, then most modern computers will be just fine in running some version of an open model to do stuff locally for you. And then, don't even get me started on what the opportunities look like on mobile devices these days.
Those models are actually accelerating their capabilities pretty fast. It's kind of amazing to watch.
Michael Krigsman: I have a Windows computer. Here's where I ask for my own psychotherapy. I have a Windows computer with a 5090 graphics card.
Raffi Krikorian: Sure.
Michael Krigsman: Okay. What's the best model I can run? And I am one, to me, the quality, if I'm doing something, the quality matters more than anything.
Raffi Krikorian: That computer is going to be memory-limited more than anything else. So there's the amount of memory that's associated with the GPU. So that's going to limit the amount, the size of the model you can connect. But I would start with the Qwen family of models. The Qwen family models, I can fit a 30B on my MacBook, but you know, I got multiple gigabytes of RAM on this thing. You know, I can even tell you how much if I pull open the About section.
But something like that, I would start with the biggest Qwen 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 going to be very honest, once you get below the 30, the language abilities and stuff like that are going to start degrading. So RAM is king in this world. You will, you need as much RAM as you can to just fit the model into memory. Now there's a lot of tricks.
People have written a lot of open source software to actually, you know, load models into memory incrementally. So you execute a few layers and you dump it, then you execute another few layers and you dump it. So you're trading off speed to fit it in. So you could do a really big model on that machine. Maybe one of the bigger Qwen models, but you're going to be trading speed at that point for it. But, I think, for your purposes, my purposes, I would start with the Qwen family models.
Those are pretty good.
Start evaluating a model backup plan today
Michael Krigsman: Let's go to another question that shifts us away from the technical discussion we've just been having to an enterprise perspective. This is from Arsalan Khan on Twitter. And he says, who should really care about cost? Is it business or IT? Is it neither of them, both of them? 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?
Raffi Krikorian: In all the enterprises I talk to, it varies who really cares the most. In some businesses, says, I have the CFO talking to me because they're like, our token spend is out of control. We spent $1 million this month alone on tokens. How can I get that under control? What are the trade-offs I can make? On some companies that talk to me, IT hits me up because they want to talk about control and where PII is going.
So 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 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 business IP to one of the cloud 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 the IT teams usually talk, the IT teams, the security teams, the CIO usually talks to me about that where they were really concerned about 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 questions about resiliency and reliability. They all saw, I'll say it again, they all saw the June to July 13th issue where Anthropic was just turned off and they're like, we can't have that in our company.
And in fact, that June 13th is not a bigger deal because, so let me back up. Fable was just released around that time, and then the US government sent an order to Anthropic that said no non-US citizen can access this. They were just banning it, which, and Anthropic had no 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.
We don't think of it as the biggest deal only because Fable was only out for maybe two 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 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 failure? What does it mean to have some redundancy and choice in there?
So maybe to answer your question, it's not usually any single department. It's usually everyone has their own reasons for why they care about you know choice, resiliency, 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.
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? Just having that good backup plan as a first shot seems like a good place to start, at least.
Michael Krigsman: For the frontier, model suppliers do want to be seen as core infrastructure, like electricity.
Raffi Krikorian: Sure. Yeah, but Michael, I used to run infrastructure at Twitter. That was one of my responsibilities when I ran engineering, and we used to build data. You would never just run a single power line into a data center because the day a squirrel chews through that line, the squirrel's going to have a bad day, but your data center's offline. You run 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. If something goes wrong, what's my redundancy plan?
Michael Krigsman: No, it's absolutely true. There was a period of time when Claude was down, 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 your internet goes down, you're out of business.
Raffi Krikorian: Yes. You're sadly out of luck. I mean, this is why I love the open models. My practical advice for most CTOs is never switch today. That's crazy. No one, don't do that. But at least start evaluating what it would take to get this backup plan in place. What's your fallback for the day that Anthropic has an outage? What's your fallback for the day ChatGPT decides to lower their token limits because they're overcapacity? What are you going to do? Start evaluating a backup plan today.
When enterprises should move to open models
Michael Krigsman: So Wayne Anderson says, when you think about. Let's talk about open weight models at CIO scale. What are the cues you see in business, in valuable things to solve, that should cause an executive to immediately say, I need to look at open weight?
Raffi Krikorian: We were facing this even at Mozilla. At Mozilla, when our finance team is starting to use Fable all of a sudden to do financial workflow analysis, that seems crazy to me. There is no reason that you need something that's basically trying to drive a Ferrari to Whole Foods. You don't need that level of power in order to pull that off. So I think 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, you know, the first thing I would try is move off Fable and move to Haiku and just see what it does before you even try moving to an open weight model. But if you think about the general innovation curve that most technical organizations go through, first you're going to, 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, in that first, explore and just hack around frontier models are great 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 optimizing it. Engineers talk about this all the time. It's like, don't prematurely optimize. So in that optimization phase, that's when you should start thinking like: Is there a cheaper thing I could do here?
Can I run 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 curve that when you get to production and before you hit the scale button, that's when you're going to be trying to eke out every single cent of your product, every single cent of your workflow, that's the time you should be really evaluating.
Do I need 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 your own personal training and evaluation set you can use to make that choice. So I think that's the way I would think about it. I wouldn't go into it off the bat unless you're a superpowered engineering organization.
I wouldn't go in 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.
Michael Krigsman: I'm sorry to keep coming back to this, but it's just so complicated.
Raffi Krikorian: It is complicated.
Michael Krigsman: 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.
Raffi Krikorian: I think it's different for every single enterprise. I think it's unfair to paint every single enterprise the same way. Thomson 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, last year we were talking about Pinterest. This year we're talking about Thomson Reuters.
Next month, next year, we'll be talking, well, I don't want to name a company because that makes them sound bad. But next month, we'll be 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 it's easier than yesterday, and tomorrow will be easier than today.
Michael Krigsman: Can you talk about the patterns that make open model deployments succeed or fail? For enterprises who want to go through this adoption process you're describing, what works? What doesn't work?
Raffi Krikorian: 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 datasets that you can use to evaluate how these models are going to perform? So most companies that, you know, when we did the State of Open Source AI report, the first version, we asked all of these enterprises, did you try open source models?
And more than a majority of them said yes, 60, 70%. You should look in the report, stateofopensource.ai. I don't remember the exact number, but a vast majority of them actually had tried open models, but only something like 20% of them got into production. And 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 of too complicated, but we couldn't figure out the right hardware-software combinations to pull it off at scale. And so all those things are still at play right now. So if you are a mid-sized organization and you have a good engineering. If you're a mid-sized organization and you have a good vendor or consulting agreement with a good vendor, then I think you are in the realm of people that can pull this off.
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 maybe the year to try 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're on the Fortune 500, Fortune 1000, like the big 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. I think those companies are ready, set 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 IIS servers.
And these days, the majority of people do not, but it took a few years to get to that point.
Subsidized token prices will not last
Michael Krigsman: Do you want to say a few words about tokenomics and the wild explosion of token costs as AI agents proliferate? That must be a factor here.
Raffi Krikorian: 100%. I mean, this is actually one 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. In some ways it is getting cheaper, but 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 happening is a per-token basis that the models are emitting.
You know, they charge you for every single token the model. It's actually even more complicated. They sometimes charge you for the tokens you put in and the tokens 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 is those token prices have been falling. Every time a new model comes out, though, those token prices are fairly high.
The coding agents, all these agents that we're using, though, consume a crap ton of tokens. It's not it wouldn't be unsurprising for me. In fact, I'm looking at one of my dashboards. On a regular task, it used something like 300,000 tokens to do a regular task. So these tokens are being consumed at a fairly crazy rate. Most people don't see this problem though because you have a Claude subscription or you have a ChatGPT subscription. So the tokens are kind of behind the scenes.
You're they claim it's all you can eat, but it's not really true. But 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, it's getting to be kind of nuts. The open source models are, again, 30 to 50x cheaper on a per-token basis.
And they're fairly stable in their cost, because once you deploy them, you just know that this is how much it costs to run it. It's not going to change depending on the state to the business. And the other problem that 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 going to IPO, delay the IPO, whatever.
But, you know, we saw this in ridesharing that when ridesharing first came out, it was fairly cheap for everyone to rideshare. Post-ridesharing IPOs, the ridesharing cost went up because now these companies are accountable to their shareholders from profits. So right now, a lot of these token prices when it comes to the big labs are being subsidized by the labs themselves as a way to try to get you to use them more.
And so 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, 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, that just seems crazy to me. That seems too hard.
Michael Krigsman: I have the $200 Claude, Anthropic Claude subscription.
Raffi Krikorian: Yeah.
Michael Krigsman: And if I had to pay for the API costs, the actual token use, my life with AI would be very different than it is right now.
Raffi Krikorian: One of the engineers at Moz AI, one of our subsidiary companies, measures this every single day because he just finds it so amusing. We all have at Moz AI the $200 subscription so we can write So we can use it for our coding purposes. And 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 that difference between your $200 subscription and $6,000 a month post-IPO is going to disappear. They're going to want that money to make their books look profitable. So that's the risk that everyone should be staring at.
Michael Krigsman: Yeah, I actually have done those measurements too. It's pretty striking. It reminds me of the dot-com, I'm old enough to remember this, the dot-com days where VCs were funding all of these companies. I remember I bought a table from Furniture.com that was so underpriced, what they paid for it.
Raffi Krikorian: Yeah.
IP rights and inspection at scale
Michael Krigsman: And I thought it was so amusing. I have a VC-funded table desk. Okay. So, we have another question here. This is from Chris Petersen, who says, what are the business guardrails around how we use GenAI in general and open-source, open-weight models in particular to make money? What can be copyright protected or sold on for profit?
Raffi Krikorian: If you look at the licensing terms of most of the closed providers, they usually say something like, we give you access to, we give you a license to the content that we create, for every single purpose we can think of, or something crazy like that. The language is actually intricately worded because the law is a little bit unclear of exactly how this works. I'm not a lawyer, but my read of it is because it's unclear of how the world, how the law works.
This is the reason again why I like the open models, because just the rules are a little bit clearer of exactly how this gets generated and works, because you're not generating it on someone else's computer. You're generating them on your computer. You're generating with software that you control, the software that you are running.
In that way, again, at the engineer's view of the law, but I wrote a blog post about this, you can see it at Owners Not Renters, that the locally run models give you better IP protections in most cases and allow you to have more flexibility in the kinds of things you can do from their output than it does from the big closed providers. Now, the big closed providers, 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 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 tier of plan are you on?" You got to be on the highest tier plan with all the right boxes checked. These are all the variables you've got to be thinking about.
Again, I'm 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.
Michael Krigsman: Wayne Anderson comes back and he says, inspection at scale for AI is a critical headache for the CIO. He says BCG published a report saying, 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 even know yet?
Raffi Krikorian: I think this is a per-enterprise business question of just like, what are the guardrails that you're putting in? What's the logging procedures you're putting in? How are you auditing the system? How did you evaluate the model? Models before you even deploy it to make sure you've guard-railed them correctly. And I do think it matters on a location-by-location basis because the data laws might be different if you're deploying in Europe versus if you're deploying in North America and the United States specifically versus Canada.
I think this is actually a very important business question that every single 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 almost certainly hitting an American data center, which is going to have its own set of rules, its own set of guardrails.
And then if you're just relying on what Anthropic is doing for their evaluation, I think you're not thinking about it correctly. 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 be asking regardless of open versus closed.
Open weight competition, China, and safety
Michael Krigsman: Like, you just need to understand. The next question, again from Gus Bekdash, who says, OpenAI, et cetera, are building data centers like crazy because they know the moat is not the model.
Raffi Krikorian: Sure.
Michael Krigsman: It's running models, as you said. So, they agree with you, Raffi. Are they wrong, and are they underestimating the open weight threat?
Raffi Krikorian: They're not underestimating the open weight threat. In fact, I would say they are They see the open weight threats right in front of them. And so, you know, people like Mozilla 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 those companies are going to spend in marketing this year alone. And so they do not underestimate this threat.
They see this threat head on, which is why they're also in policy conversations, regulatory conversations, etc. to try to lock themselves in and try to edge out the open providers. I think they understand the threat, and they are behaving as if they understand the threat.
Michael Krigsman: Many of the best open models come from Chinese labs. What does that mean for a company in the U.S. or in Europe?
Raffi Krikorian: 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 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 Aleph Alphas in Germany or the Cohere in Canada, the Reflections in the US. There are the Apertus in Switzerland. There is starting to be a broader ecosystem of people providing open weight models, which I think is a good thing because it means that we are not under a single regulatory regime. Now, that being said, you can't undownload a model.
So yeah, it should be concerning about moving forward on Chinese-based models, but all the ones that are already out there on Hugging Face, etc., you can't undownload them. So, you know, maybe download the one that you're going to run on your business and keep it on. Keep it on your own disks. Don't rely on 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 is generally no. Think of it as like, I'm now running this on my machine. It's under my control. You could then even inspect, is anything leaving your network boundary that you don't know about? It's just not the way that these things work now. 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 about, not the model itself that you might download and run.
Michael Krigsman: All right. Critics say that once a model is open, it can never be recalled and its safety controls can be stripped out.
Raffi Krikorian: Where do you draw the line? I firmly believe that having open models in the world and allowing more and more people to use, tinker, understand, etc., means that we'll build better open guardrails, better open monitoring systems, better open safety systems to detect, to react, to manage, etc., all these things.
I think if we get in a world where we have open models, we'll have a better way of managing where we don't allow open models 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 effectively protect us all. And as demonstrated by the Hugging Face stuff, the Australian hack, the, all this stuff, they're not capable of doing it. So 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 actually to have these open models out there.
Token price hikes and context lock-in
Michael Krigsman: Aik Sargsian comes back and he said, do you really think that when these companies go public that tokens will become much more expensive?
Raffi Krikorian: 100%. There is nothing which leads me to believe that that will not be the case. Now, the only thing that might be countervailing to that is that as these models, as we develop more and more models, they are getting more efficient. So we're able to do more with less hardware. But, you know, honestly, that will just drive up what the capability of those models is, and those token costs will go up again. So I 100% believe we live in a subsidized regime right now.
Like you were saying, Michael, on VC subsidizing your car, your tables, what I'm saying about ridesharing in the pre-IPO world of ridesharing. After IPO, it's not going to happen on day one, but you're going to see it in quarterly reports of just like, we figured out a way to eke out more profit this way, and it's going to come down to the cost of those tokens at some point.
Michael Krigsman: Uber has become quite expensive.
Raffi Krikorian: That's exactly my point. I mean, I, having been one of the people who worked at Uber pre-IPO, I remember the cost of Uber back then. We subsidized an entire generation of kids effectively to use Uber, and now Uber is like, you know, is it expensive? But Uber, you're now seeing the true cost of that system effectively.
Michael Krigsman: Final thoughts, advice, maybe comment on lock-in. Your report says that lock-in has moved from the model to the context that surrounds it. How can companies overcome this form of lock-in?
Raffi Krikorian: The biggest theme of the latest version of our report, so stateofopensource.ai, we rev it every couple of months. It's a fairly live document to help you understand how open source AI is sort of evolving in the world, but 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 control this infrastructure in ways that just wasn't possible even just a few months ago. The infrastructure to do fine tuning has super evolved just in the last three months. The ability to deploy a model that started off as one of the Qwen 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.
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 providers' business model It's for you to get less locked in as possible. You can see this right now. If you spend a week using Claude and then you switch to ChatGPT, it's like talking to an amnesiac all of a sudden because all your memory got sucked into the Claude system.
And so in a world where we can now download, fine-tune, deploy ourselves, it means also our memory, our data, etc., can stay on our side, which provides us leverage. Again, 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, it's not easy yet. It's easier than last year. It will be easier next year.
But we have a lot of work to do, but we're on that verge that if you are an AI-native company and you're not evaluating open 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. Get through that adoption curve first and then start thinking about it. We're at that point in the conversation, I think.
Michael Krigsman: Raffi Krikorian is CTO of Mozilla and, I must say, a very eloquent advocate for open-source AI, open weight models. Raffi, thank you so much for being a guest again on CXOTalk.
Raffi Krikorian: Thank you, Michael. It's been such a pleasure.
Michael Krigsman: Everybody, thank you for watching, especially you folks who asked such great questions. Before you go, subscribe to the CXOTalk newsletter, and we'll keep you up to date on, we have unbelievable guests and shows coming up, so check it out. Thank you so much, everybody, and we'll see you again soon.

