This Is a Fight Worth Having: The Case for Open Source AI | Raffi Krikorian, Mozilla CTO

Summary

Raffi Krikorian, CTO of Mozilla, discusses the transition of open-source AI from an ideological aspiration to a strategic business choice. He highlights how the economics of AI are starting to favor open models, citing examples like Pinterest saving millions by moving away from closed providers. Krikorian argues that as companies move past experimentation and into production, the maturity of their applications allows for more specialized and cost-effective open-source implementations.

The conversation also delves into the concept of a “LAMP stack for AI” — a standardized, interchangeable set of layers for compute, models, data, and developer experience. Krikorian emphasizes Mozilla’s role in driving these standards and providing “choice suites” that allow developers to switch between providers easily. He specifically underscores the importance of data sovereignty, privacy, and the need for a “user agent” in the age of intelligence to protect individual context.

Krikorian concludes by advocating for the “Rebel Alliance” of open-source contributors who, by building credible alternatives and maintaining standards, can prevent the consolidation of AI power among a few large corporations. He suggests that even small investments in “instrumenting choice” early in the development cycle can provide companies with the flexibility to adapt and scale as the open-source ecosystem continues to vibrate with innovation.

Highlights

”Economics and Values Starting to Align”

yt-dlp --download-sections "*1:33-2:33" "https://www.youtube.com/watch?v=3ohQX-oYndI" --force-keyframes-at-cuts --merge-output-format mp4 -o "3ohQX-oYndI-1m33s.mp4"

“I think that it is an unfair thing to assume that open source wins because of values. I think it needs to be this combination of values and economics need to really work out in order for it to make sense… We’re starting to see a bunch of people have very similar comments that they’ve passed through the prototyping stage. They’re getting into like the steady-state utilization phase.” — Raffi Krikorian, 1:33

”The LAMP Stack for AI”

yt-dlp --download-sections "*7:11-7:57" "https://www.youtube.com/watch?v=3ohQX-oYndI" --force-keyframes-at-cuts --merge-output-format mp4 -o "3ohQX-oYndI-7m11s.mp4"

“What I would propose is that we should look at on like four different layers. So like we look at it as like the compute layer, we look at it at the model layer, we look at at the data layer… and then we look at the developer experience layer. And I think this replicates itself both on training and on the inference side. And so I think that like what the conversation looks like that we need to have is like what do APIs look like or what do standards look like between those layers so that we as an open-source ecosystem can start coalescing around this lamp-like stack metaphor.” — Raffi Krikorian, 7:11

”I Own My Context”

yt-dlp --download-sections "*17:34-18:20" "https://www.youtube.com/watch?v=3ohQX-oYndI" --force-keyframes-at-cuts --merge-output-format mp4 -o "3ohQX-oYndI-17m34s.mp4"

“How do we make user context portable across a different sets of model providers so that we can invert that ownership paradigm that it’s not that the model provider owns my context. I own my context and I can make it portable and I can expose different aspects of it depending on what we’re trying to do at any given moment.” — Raffi Krikorian, 17:34

”Instrument Choice as Early as You Can”

yt-dlp --download-sections "*22:02-23:02" "https://www.youtube.com/watch?v=3ohQX-oYndI" --force-keyframes-at-cuts --merge-output-format mp4 -o "3ohQX-oYndI-22m02s.mp4"

“Try to instrument choice into your engineering stack as early as you can… Give yourself the exit door as early as you can so that you can then do it later when the time is right. Don’t make it a massive engineering, a herculean engineering effort later.” — Raffi Krikorian, 22:02

Key Points

• Economics vs. Values (1:33) - Open source needs to win on economics, not just ideological values.
• The Pinterest Benchmark (1:44) - Pinterest saved $10M by switching to open-source models for production.
• Mozilla’s Choice Suite (2:42) - A toolset to enable enterprise flexibility, similar to how Terraform works for cloud infrastructure.
• The Maturity Curve (5:12) - Companies often start with closed APIs (like OpenAI) but switch to open source as their application matures.
• Mozilla’s North Star (6:15) - To do for the AI ecosystem what Mozilla did for the web by bringing user choice.
• The Four Layers of AI (7:13) - Proposal for a LAMP stack for AI: Compute, Model, Data, and Developer Experience.
• Connective Glue (11:41) - The missing piece in the open-source ecosystem is the layer that makes components easy to integrate.
• Smaller vs. Larger Models (10:30) - Fine-tuned smaller language models (SLMs) are becoming a more efficient path than massive transformers.
• Mozilla Data Collective (10:50) - Creating a marketplace for provenanced data, like licensing radio audio for model training.
• Federated Learning (12:00) - Using technologies like Flower AI to shift compute from scarcity to abundance.
• Hardware Agnosticism (14:06) - Choosing the right model to run on the right device (e.g., local vs. cloud).
• Latency & Battery Constraints (15:01) - Why autocomplete and other interactive features should run locally on device.
• Data Sovereignty (17:15) - Preventing giant model providers from owning user context.
• The New “User Agent” (18:03) - The role of the browser/agent to protect users in an intelligence-mediated world.
• The Signal Strategy (18:55) - How a small, privacy-focused player can force the entire industry to adopt encryption standards.
• The Firefox Legacy (20:05) - Even with lower market share, Firefox won by forcing open web standards.
• The Rebel Alliance (21:14) - A distributed ecosystem using open source to provide credible alternatives.
• ** Instrumenting Choice** (22:05) - Developers should add shims early to allow switching providers without heavy refactoring.
• Accessibility of Open Source (24:11) - Tools like ‘Ollama’ make installing open models simple for anyone comfortable with a terminal.
• Einstein’s Dreams (24:48) - Raffi’s book recommendation exploring different vignettes of how time flows.

Mentions

Companies

• Mozilla (0:50) - Building the open AI ecosystem.
• Pinterest (1:46) - Saved $10M using open source.
• OpenAI (4:43) - Standard for fast API experimentation.
• AWS / Azure (3:01) - Cloud provider lock-in examples.
• OpenRouter (9:09) - Unified API for multiple LLMs.
• AI2 (Allen Institute for AI) (10:03) - Producers of open models.
• Databricks (10:48) - Major player in the data layer.
• Flower AI (12:01) - Federated learning infrastructure.
• Signal (18:53) - Influence on mobile privacy standards.

Products & Technologies

• Choice Suite (2:47) - Mozilla’s toolkit for AI provider choice.
• Terraform (3:03) - Metaphor for infrastructure abstraction.
• MCP (Model Context Protocol) (9:12) - Protocol for model interactions.
• LAMP Stack for AI (7:13) - Proposed standardized AI architecture.
• Ollama (24:11) - Tool for running open models locally.
• Federated Learning (12:03) - Distributed AI training technology.

People

• Raffi Krikorian (0:50) - CTO of Mozilla.
• Alan Lightman (24:48) - Author of Einstein’s Dreams.

Surprising Quotes

“It is an unfair thing to assume that open source wins because of values. I think it needs to be this combination of values and economics need to really work out in order for it to make sense.” — 1:33

“No one builds raw to AWS anymore… they build on Terraform. Those type of moments are going to start to show up this year (for AI).” — 3:01

“I want to invert that ownership paradigm that it’s not that the model provider owns my context. I own my context and I can make it portable.” — 17:43

“Signal… forces people like WhatsApp, Telegram, iMessage… to think about having encryption because it’s being driven by that conversation. If Mozilla… can start building what these interfaces could look like where my context stays for me… that could drag the ecosystem.” — 18:55

Transcript

0:00 An unfair thing to assume that open-source wins because of values. They’ve saved $10 million by moving to open source. Actually, there’s a lot of vibrancy in the open-source ecosystem. I think what’s missing is a connective glue layer to just make it easier for developers.

0:28 Hello everyone. We’re doing something different with Inference show by Turing Post this year. Starting with this episode, we’re dedicating each quarter to one topic explored through multiple conversations. And this quarter’s topic is open-source AI, and I’m starting with someone who is reorganizing the whole company around this bet: Raffi Krikorian, CTO at Mozilla.

0:53 Welcome Raffi. “It is such a pleasure. Thanks for having me.” Absolutely. But let me be honest with you about open source. I had one problem recently with that: all the conversations sound very aspirational. It’s about freedom, transparency, and ownership. But that’s not how companies do business, right? When does open-source AI become an operational choice rather than an ideological one?

1:19 “Well, on one hand, I’m actually kind of glad that you only have one problem with it, because I think there’s like a billion problems with it in the grand scheme of things. But no, you’re exactly right. I think that it is an unfair thing to assume that open source wins because of values. It needs to be this combination of values and economics that really work out in order for it to make sense.”

1:41 “One concrete example is at the end of last year, Pinterest put out that report that they’ve saved $10 million by moving to open source rather than relying on the closed search providers. We’re starting to see a bunch of people have very similar comments. They’ve passed through the prototyping stage, they’re getting into the steady-state utilization phase, and therefore they now know how to provision properly.”

2:06 “They’ve figured out which models actually perform quite well, so they’re no longer paying a per-token cost and instead absorbing that into the infrastructure. If I were to make a prediction for 2026, we’re going to start to see a lot more of those type of stories showing up: that the economics and the values are starting to align.”

2:36 “Another thing I’ve been thinking about is this notion of how to enable that choice to happen in the future. What we’re building at Mozilla, on the Mozilla AI team, is this thing they call the ‘choice suite.’ It can actually enable enterprises to be more flexible in how they even negotiate.”

2:56 “I remember the days when everyone actually built raw to AWS. No one does that anymore; they build on Terraform and all these other things. What that provides them is negotiating opportunities. If AWS says it’s going to cost 50,000, you can say Azure says it’s only 30,000, and I have a Terraform script so I can just move over quickly. Those type of moments are going to start to show up this year.”

3:27 I heard this case of Pinterest. If you think about it, it’s a huge company with a lot of machine learning engineers who can be dedicated to the switch. Does the company size make sense for this? Technical capabilities vs. spend level? When is open source actually cheaper for them, not just principled?

3:45 “I think it’s even before the dollar sign/euro sign conversation. I think it really goes to how fast can people move. In the Pinterest case, when you read between the lines, you’d implicitly see that they went past experimentation. They actually now know what their use case looks like, the kind of application they want to build and deploy.”

4:13 “Now they can actually codify it, and I’ll use fine-tune as an overloaded term, but they build the exact infrastructure they need in order to achieve their goals. A version of that plays out on every single level. The reason why I balk at a dollar sign is that small companies, by definition, are still in iteration. They’re trying to experiment to see what works best for them.”

4:40 “Right now, I can’t with an honest face say they should be experimenting with open-source AI. The OpenAI API is kind of lovely; it’s one API call that abstracts away all the stuff behind the scenes. Whereas if you try to use open-source software, you’re collaging together all these different pieces. You’re trying to figure out which evaluation provider, which traceability provider.”

5:03 “Once they get past exactly experimenting on what they want to go build and they want to achieve, to get into production or some form of cost optimization, then I think open source is getting to be positioned that it could be a valid part of that conversation. So I don’t think it’s just a dollar sign. I actually think it’s the maturity of the application that you want to build.”

5:23 When did it happen at Mozilla? I’m sure you were iterating and experimenting a lot. When was the moment you said: “We’ve always been an open-source company, although it’s simpler to use OpenAI, we don’t want to lock in”?

5:40 “We’re actually still going through this moment right now. For example, I wouldn’t say Firefox is fully locked into exactly what their AI-powered user experience really looks like. Teams at Mozilla are very deep into using things like Cloud Code, etc. Only a few of them have even started switching to things like Open Code.”

6:09 “Part of my job at Mozilla is to think about the broader AI ecosystem. How can we do for the AI ecosystem what Mozilla did for the web? How can we actually bring user choice? How can we give credible alternatives that force openness in different parts of the stack? My goal is to make it so that the open-source choice can happen earlier and earlier in that experimentation process.”

6:39 “For me, I think about Mozilla as even a testing ground. What are the different things we can be experimenting with as early as possible? We’re not quite there yet, but that’s maybe our north star for what we’re trying to get done this year.” Well, let’s talk more about the open-source stack then. In practice, it’s multi-layered. What does the stack look like?

7:10 “Stepping back, what we need is a LAMP-stack definition in a way that we’re all familiar with. Right now, there isn’t a good one. I would propose four layers: compute, model, data, and developer experience. This replicates itself both on training and on the inference side.”

7:41 “What we need to have is conversations about what APIs or standards look like between those layers so that we, as an open-source ecosystem, can start coalescing around this lamp-like stack metaphor that can make it really easy to deploy. With our choice suite, we’re trying to provide that top-of-the-stack opening.”

8:06 “If you use the any-LLM API, which is part of the choice suite, you can rapidly switch between different providers—local, remote, open, closed. We’re trying to embed choice into that conversation. We need interchangeability so people can choose what their stack looks like and form it rapidly. Whether for small businesses, enterprises, or governments, I want to figure out how to make that LAMP stack coalesce.”

9:04 Is anyone building something like this? “Lots of people are building parts of it. OpenRouter is a good example. I think MCP, which allows calling from the model layer, is part of this. Mozilla doesn’t need to build this all ourselves. My goal is to articulate what developers are looking for and drive the community towards standardization opportunities.”

9:38 Who do you think are the companies that serve as your ideal models currently? “OpenRouter and Mozilla AI at the top. I’m excited about the work that AI2 is doing at the model layer, or ETH Zurich. There’s going to be a lot of innovation in that model layer, especially as we come to grips with the fact that it’s impossible for anyone to play in it (financially).”

10:21 “The transformer architecture makes it computationally and financially impossible. But there’s a lot of evidence that smaller language models or fine-tuned models from the pre-transformer days might be the path forward. I’m intrigued by companies building infrastructure for faster fine-tuning or model composability.”

10:45 “On the data side, I’m interested in what Databricks is up to. Mozilla has the Mozilla Data Collective where we’re creating a two-sided marketplace for provenance-related data that you could license. Our folks are in conversation with every radio station on the planet to list audio data with transcriptions for model trainers to use.”

11:19 “And of course, at the top of the stack, there are so many companies. I’m a big fan of LangChain and the orchestration layers. There’s a lot of vibrancy in the open-source ecosystem. What’s missing is a connective glue layer to make it easier for developers. I want a world where it’s just as easy to use an OpenStack as it is to use the OpenAI API.”

12:13 “The compute layer will be the hardest to make a dent in. Companies like Flower AI are doing interesting work in federated learning across heterogeneous hardware. Imagine a world where compute was an abundance mindset instead of a scarcity mindset. That would be an interesting AI world to live in.”

12:42 I’m a big fan of federated learning. If we think about small models and move to specialization—one for coding, one for this—how can companies build that decision-making process for using open source without added complexity?

13:18 “I look at it like my previous jobs at Twitter. A decade and a half ago, Twitter was Ruby on Rails. Once we understood the application stack, we moved parts to the JVM to harden it and make it faster. The same thing is playing out here. Even I use the OpenAI API to rapidly form an idea.”

13:56 “But when I’m ready to deploy, I looks at whether to do a part in Qwen or something else. You start with a general-purpose tool to sort out your idea, and as you get solidity, you optimize. There’s also constraints like running in a disconnected environment or on-device. I do work in Africa where connectivity is poor, and you want it to run locally.”

14:59 “Another example is autocomplete on a web browser. There’s no reason to send that information to a large model remotely; you add latency. You want to do that as close to the user as possible, being sensitive to keystrokes, timing, and battery power. If it’s local, why hit the network?”

15:28 “Those constraints are the driving factors. As you go from experimentation to production, that’s the right time to think about those transitions. My job is to move that line as close as I can to experimentation.” Why does Mozilla focus so much on user data and privacy?

15:54 “Data in the AI world is one of the biggest flipping points. Most providers think data is free; they just scraped it. But there are datasets NOT on the web that are really useful, like indigenous languages or genetic data. People deserve recognition or compensation for these. That’s the human-centered version of data collection.”

16:39 “We want to enable more interesting experiences on the web that respect the people in the value chain. Then there’s the question of big model providers coming after yours and my private data. I’m interested in how we start thinking about protections for user context. How do we make user context portable across model providers?”

17:39 “I want to invert that ownership paradigm: the model provider doesn’t own my context; I own it. This is compatible with how Mozilla thinks about the web browser as a ‘user agent’ that protects me. In this new agentic world, intelligence is starting to mediate our interactions, and we need our data to be secure and private.”

18:22 This is a race so hard to win, considering how everything in an iPhone becomes connected to Google or Apple. “I don’t know the answer, but this is a fight worth having. It starts with having credible experiences. My canonical example is Signal.”

19:13 “Signal drove privacy as part of the communications conversation. It didn’t need every user to use Signal; its existence forced WhatsApp, Telegram, and iMessage (and now RCS) to think about encryption. If Mozilla can build interfaces where context stays with the user, that can drag the ecosystem towards user privacy.”

19:55 Conversations matter. Tricky question: Firefox challenged Microsoft, but then Chrome brought Firefox share down to 5%. How can we make sure AI doesn’t just consolidate again? “These are pendulum swings. Even if Firefox usage went down, Google had to release an open-source browser. We maintained web standards.”

20:41 “The battle we tried to win by releasing Firefox, we won. Because of us, the web is open and distributed. Consolidation occurred, but web standards and open source are part of the conversation. That’s not the direction AI is currently going. That’s why I’m dragging Mozilla into this fight—to drive the conversation through credible alternatives.”

21:22 “We’re not going to be the model provider; we can’t compete there financially. But if we can get an entire ecosystem speaking the same language, it’s like a ‘Rebel Alliance.’ If it means user data is secure, choice in the ecosystem, and I can surf the web without worry, then Mozilla will have won that fight.”

22:04 Advice for a company in experimentation mode: “Instrument choice into your engineering stack as early as you can. Use something like the choice suite. It’s just a small shim. Give yourself the exit door early so you don’t face a Herculean engineering effort later. It’s a five-line code investment at the start.”

23:02 “Think about refactoring and testing early. Cursor and others take these prompts well. Flexibility is the biggest piece of advice. Second, be open to what’s going on in the ecosystem—read blogs, listen to podcasts to understand when to swap components.”

23:43 I would like to see how a non-technical person can just install an open-source model on their phone. “That world is not here yet, but we’ll get there. Right now, on a Mac, anyone can do a ‘brew install ollama’ and have an open-source model. It’s a one-liner, but you need to know the terminal. We’ll eliminate that terminal requirement this year.”

24:26 To finish, what’s a book that recently or long ago influenced your work and values? “I’m not prepared for this! One book is Einstein’s Dreams by Alan Lightman. A set of vignettes about how time could flow differently—higher altitudes, circles, etc. It was the first book I gave my now-wife. It’s the book that changed my life.”