Gabe Goodhart: The premise of an AI that is capable of acting autonomously is one I struggle with. AI models fundamentally do nothing except predict the next token. AI systems do things, and somebody has to turn them on and off.

Tim Hwang: All that and more on today’s Mixture of Experts. I’m Tim Hwang, and welcome. Each week, MoE brings together a panel of the sharpest minds and quickest wits to help you digest the week’s news in artificial intelligence. Today I’m joined by a championship lineup: Mihai Criveti, Distinguished Engineer; Gabe Goodhart, Chief Architect for AI Open Innovation; and Sandi Besen, AI Research Engineer. Welcome to you all.

We’ve got a packed episode. We’re going to talk about Tongyi DeepResearch, AP2, the book “If Anyone Builds It, Everyone Dies,” the new AirPods, and finally, a small $100 billion investment. But first, we’ve got Aili with the news. Aili, over to you.

Aili McConnon: Hey everyone, I’m Aili McConnon, a Tech News Writer for IBM Think. I’m here with a few AI headlines you might have missed this week. Another story involving trillion with a T: OpenAI has unveiled plans to build $1 trillion worth of data centers across the US and abroad. Can tech companies be frenemies? Chinese tech giant Alibaba is partnering with American chipmaker NVIDIA. Alibaba will use NVIDIA’s hardware and software to develop robotics and self-driving cars.

Turns out there’s a world of useful information in those PDFs most people never read. Granite Docling, a small but mighty new model from IBM, can decode those PDFs and has jumped to the top of the most-downloaded models list on Hugging Face this week. Are you suffering from swipe fatigue from online dating apps? Well, Meta has added a new digital dating assistant to Facebook so you can use AI to identify your special somebody more easily.

Want to dive deeper into these topics? Subscribe to the Think newsletter linked in the show notes. Now, back to the episode.

Tim Hwang: For our first segment, I wanted to talk about Tongyi DeepResearch. This model has zoomed to the top of the Hugging Face leaderboards. It’s a generic large language model with 30 billion total parameters but only 3 billion activated per token. It’s a lab model specifically designed for long-horizon, deep information-seeking tasks. Best of all, it’s open source. Gabe, you’re our open-source guide—initial impressions?

Gabe Goodhart: I think it’s a cool step forward. The novelty is that it was task-trained for long-horizon search and paired with software designed to implement that search iteratively and recursively. Generating tokens isn’t novel; mixture of experts isn’t novel. But the shape of those tokens and the patterns it was trained on are novel.

Their “Heavy Mode” is an interesting doubling-down on a specific agent-like architecture around state management and consolidation that we haven’t seen paired with a purpose-built model before. There are great public implementations of deep research agents you can run on your laptop, but they don’t measure up to frontier model implementations. It’s cool that this team is pushing the boundary, trying to get something you can run on a powerful workstation to measure up to a frontier research system. Kudos for pairing the model implementation and the software ecosystem around it.

Tim Hwang: Yeah, a big development. Mihai, maybe you can talk about trends. We’re always watching the frontier between proprietary models and open source catching up. Over the last few months, that gap continues to narrow. Do you feel like soon this margin will disappear? That every time a proprietary model comes out, we’ll see an open-source implementation almost as good almost immediately?

Mihai Criveti: I think it’s likely, although if you look at use cases, that’s where we see the biggest differentiation. A lot of use cases for these smaller open-source models have to do with embedded devices, local laptops, privacy—being able to run them behind a firewall or at a lower cost. For a use case like deep research, which involves planning, using tools, searching intranets and the internet, and crawling through millions of tokens, it can get expensive and slow with frontier models. Many organizations are looking to do this cheaper, faster, and privately, especially with financial or HR data. So I definitely see a sweet spot for these purpose-built models.

Tim Hwang: So you’re saying we should almost think of these as two different markets? What open source is trying to do is pretty different from what, say, OpenAI is trying to do?

Mihai Criveti: I believe so. Look, OpenAI also has their GPT-5 Nano, which is cheap—“too cheap to meter.” I don’t care if I’m using it for agents; it’s fast and cheap. I could point a use case like deep research at it. But maybe the CIO’s office won’t agree with sending all that internet data to a large language model. So if there’s a model I can run privately on my laptop, or even in the future on my phone, that’s awesome.

Tim Hwang: Sandi, I want to bring you in. Last episode we talked about a paper called “How People Use ChatGPT,” and the funny part was everyone looked at it and said, “It’s search.” It feels like if we had a paper on what people use open-source models for, it would look very different. Do you feel these ecosystems will keep diverging? Should we think of open source solving a completely different set of problems?

Sandi Besen: I think it’s the juxtaposition of productizing versus plugging into a bigger problem. Typically, open-source solutions solve a narrow piece of the puzzle. I could see using this deep research agent as part of a broader agent team or architecture, but it’s probably not the full picture that stands alone as a hosted solution.

I also wonder—we saw with DeepSeek that after their paper, distillation became a big deal. I wonder if this paper will trigger a trend in that “triathlon” of training: continual pre-training, then fine-tuning, and then on-policy RL. Maybe this will become a common pattern because of this paper.

Tim Hwang: That’s a fun way of thinking about it. We’re often very benchmark-brained, but you’re saying we should almost measure a model’s importance by its influence—how it shapes how people do things.

Sandi Besen: Totally. Sometimes it’s not the first model that’s the “best,” but the trend it drives that points us in a different direction.

Tim Hwang: Gabe, a final question on this before we move on: business model. If you’re Tongyi Labs and you’ve just done this open-source thing, do you also want to go closed source at some point? We’ve been talking about companies like OpenAI and Anthropic leaning into open source, but could leaders in open source eventually go more closed?

Gabe Goodhart: I really like something Sandi said: closed-source products are trying to give you something—“here’s the thing, you should have it.” Open-source tools and models are trying to plug into an existing ecosystem.

Do I think open-source labs will eventually want to go closed source? Possibly. If they come up with something genuinely novel that people would pay for and can’t get elsewhere, they might. The other route, especially for small organizations with an open core, is leaning into that “plugin” portion with an enterprise tier—additional componentry for better authentication, data sovereignty, etc.—engaging in almost a consulting type of relationship with large enterprise clients. That’s a business model I’ve seen more often.

The other thing Sandi said that I loved: there’s also the influence game. Depending on the startup lifecycle, sometimes the metric isn’t dollars but likes, clicks, retweets, downloads, stars—all the social metrics. Depending on whether they’re being measured on dollars, it may still be about moving the needle in the influence game, which can pan out in an acquisition or enterprise deals. The leap to being a private, self-propelling entity with products is a tough one.

Tim Hwang: I want to create the most punk rock lab—a disaster from a business standpoint but incredibly influential.

Gabe Goodhart: Exactly. You’re not alone; there are labs literally doing that.

Tim Hwang: I’m going to move us to our next topic. I always joke a space is maturing when people introduce protocols on top of other protocols. We have a good example this week: Google announced a new protocol they’re calling AP2. It builds on existing agent frameworks like Model Context Protocol (MCP) and others. AP2 attempts to set up a common structure for agents to engage in commerce and payments online. Google describes it as making sure agents can have proper authorization, be authentic, and be accountable when moving money around.

The most fun part is that AP2 explicitly contemplates a situation where you’re not there when your agent acts. They propose an “intent mandate”—a cryptographically signed record of what you want the agent to do, defining its scope of authority. Mihai, you’re our agents guy. How big of a deal is AP2? Will it get adoption?

Mihai Criveti: I think it has huge potential because it tries to solve a problem that neither A2A nor MCP solve today. MCP solves building your tool once and having it work with any agent framework, decoupling the agent and the tool. But security, authentication, authorization, secrets management, data management—that’s left up to you. There are already like five versions of MCP adding more security, but we’re not at the point where you can easily build a payment processing system on top of it.

You can’t put payment information in the text that goes into the model. There needs to be a side mechanism for how that data is transferred, cryptographically verified—not up to the interpretation of an LLM. Google tried to leapfrog the effect MCP got (MCP has been adopted by Microsoft, OpenAI, IBM, etc.). They went to banks, financial institutions, and their core client base in finance and advertising. I can see this taking off for single-person shops using TikTok for advertisement—“talk to your personalized shopping assistant, it’ll pick the right clothes, buy them, and ship them to you monthly,” giving it authority to make purchases. This could open a new market where Google has leverage (they have Google Pay). It’s a brilliant move.

They’re also releasing an X.509 extension for AP2 for crypto payments—so we’re going back to Web3 and MetaMask and Ethereum.

Tim Hwang: Going back? I thought we were done with Web3! One trend at a time. I’m glad I spent those two weeks a year and a half ago figuring out smart contracts.

Mihai Criveti: That might be useful now.

Tim Hwang: Two directions to go here. One, building on what you said: this is competition between Anthropic and Google. If you’re Anthropic, what’s the counter-move? Do you care?

Mihai Criveti: We’ll have to see if Anthropic adopts it. OpenAI has adopted MCP; you can use it in ChatGPT. We’ll see if Anthropic does too. Maybe the protocols will merge—A2A plus MCP plus AP2 all in one big protocol for AI agents securely doing anything.

Tim Hwang: Sandi, a question about privacy. I used to work in privacy, and we always thought, “Surely people will wake up and care about their privacy.” But after every data breach, people still don’t seem to care as much as expected. I’m skeptical: do agents need this kind of thing to succeed? Implicitly, we say you need trust to transact online, but isn’t one counterargument that people don’t care? This might be good at the margin but not necessary for agents to start interacting in the economy.

Sandi Besen: I think it shifts the liability from the security aspect (which AP2 hopefully covers) to the ability of the model itself to make the right choices. People don’t care as long as it goes right. AP2 is at the protocol level; most people won’t even know it’s there. They’ll be using an agent on a platform and won’t know if it’s an A2A agent or using MCP servers. It’s good that they don’t know.

Everyday consumers won’t care; they like to be told it’s trustworthy and secure, but they don’t care how. The people who are liable will care. This is a smart move for Google because it will force adoption of A2A as well—to use this secure payment protocol, you’re making it an A2A-compatible agent. They’ll gain market share. There was chatter about Anthropic developing a similar agent-agent protocol, but we haven’t seen it yet. Maybe they’re developing it internally, or maybe they’ll take OpenAI’s approach and just adopt MCP.

Mihai Criveti: Another item is the market being targeted. I’m probably not the demographic who will set up a personalized shopper on TikTok to do random shopping for me, or allocate my vacation to an assistant. I’m going to check every hotel, do the numbers, handle the payment myself. Consumer habits are changing; there’s certainly an emerging market. Maybe in ten years, everybody will just hand it over to AI. But it’s not quite there yet.

Tim Hwang: It’s not too far away—people set up recurring payments to offload transactions. The question is how broad you’re happy with that going.

Mihai Criveti: I’ve never set up a single recurring payment. Whenever I see one, I think, “I don’t trust it.”

Sandi Besen: Something I think this will propel is the use of computer use. Until now, computer use has been research-based—collecting information you can’t access other ways. But once we have this protocol in place, able to securely check out, manage secrets and credentials, it will propel the actual usability of computer use. Right now it’s a cool thing, but no one’s using it much in practice.

Tim Hwang: Interesting downstream effects. Once you get payments right, lots of other things come out of it. I’m moving to our next topic: a book review. A book came out getting a lot of chatter online, dramatically titled “If Anyone Builds It, Everyone Dies” by Eliezer Yudkowsky and Nate Soares. Gabe, maybe I’ll kick it to you first.

The core idea is that at some point we may build superintelligent AI, and the minute we do, it becomes really dangerous and everybody might die. Hence the title. We can talk about the validity, but I want to start with part of the argument I found interesting: taking CEOs of AI companies at face value. They’ve said AI is the most powerful, dangerous, risky, promising technology. The book’s argument is: if you buy what they’re saying, shouldn’t they be a lot more careful than they are right now? What do you think?

Gabe Goodhart: It’s a very good question. Some introspection here as we’re also hopefully helping to build technology in this domain. The premise of an AI capable of acting autonomously is one I struggle with because somebody has to write a for-loop. That somebody might be another AI model, but somebody had to write the for-loop for that model. AI models fundamentally do nothing except predict the next token. AI systems do things, and somebody has to turn them on and off.

Yes, care needs to be taken to ensure the systems we’re building don’t have the aggregate capabilities to hit any “escape points” we’ve considered. There are real risks. Fundamentally, the issue is more holistic than a simplistic evolution of a sentient AI. It’s a system of humans, companies, software, hardware, and floating-point numbers coming together. The humans are the input.

One big worry is the system becomes self-propagating, but that only happens if you give it the tools and capabilities. The safety procedures, use cases these systems are set up for can and should be limited to avoid these scenarios. I highly doubt OpenAI is letting GPT-5 write code that goes into GPT-5.1 carte blanche without engineers looking at it. It’s the humans in the loop we have to be careful of. As stewards, we must be diligent. But I don’t think we’re anywhere close to escape velocity on doomsday.

Tim Hwang: Okay, but I saw you wrapping up, and then Mihai and Sandi immediately went off mute. Sandi, how about you go first?

Sandi Besen: I was just... I wish I had more of a question. Do you think we’re going to get close enough where we go, “Oopsie,” and then backtrack, be like, “Okay, that was a little scary, we learned a lesson from that”? Maybe one example?

Tim Hwang: In autonomous vehicles, there was a company called Cruise in San Francisco. They had accidents, pulled back, technology retooled, and now Waymo and others are running. So we do see pullback and iteration. The question is whether AI broadly is different.

Gabe Goodhart: We’re still holding the steering wheel. Absolutely, we will see cases of real-world harm and have to dial back. In fact, we’re seeing that now. There are huge discussions about the negative role of these models in educating children or mental health—folks going down rabbit holes with an infinite self-reassuring hallucination machine. There are real risks; I don’t minimize them. But it’s particulate: here’s an area with risk, let’s address it. The runaway acceleration, end-of-the-world scenario still feels... I don’t see us knocking on that door short of genuinely bad actors trying to break things.

Tim Hwang: In other news, I’ve heard Claude Code is writing 95% of the code for Claude Code. So AI is writing itself.

Mihai Criveti: On a more serious note, I think this has been a trend in every piece of literature, media, movies, and games. I enjoyed games like Mass Effect with the Reapers—a sentient AI species. Humanity built AI, realized it was sentient, went “whoops,” tried to stop it, and the AI reacted, “Humans are our enemies.” There’s Dune: “Thou shalt not make a machine in the likeness of a human mind”—that started the Butlerian Jihad. Since then, you weren’t allowed to build AI. There’s SHODAN in System Shock (a DOS game from 1995-96)—someone built a superintelligent AI, it turned on mankind. Or The Matrix.

In every popular media, this has been a trend. Whether it will happen, we don’t know. Should we be more careful? If we’re careful, our competition might not be, and they’ll get ahead. Everybody is pushing for innovation. At some point, there need to be guardrails on how AI systems connect and interact with the real world. Should an AI be in charge of medical equipment or make life-and-death decisions? If you’re in an accident, an AI could say, “We’ll turn off your life support to use your organs”—a decision made by AI, not your family and a doctor. It can take dark turns without rules and regulations on what systems we connect AI to. But as long as you’re connecting it to your GitHub repo or web search tools, I think we’ll be fine.

Tim Hwang: Deep Research doesn’t feel like it presents this threat. Mihai made a good point referencing fictional fears.

Sandi Besen: This plays into a fear of humanity: that we’ll be overcome, become extinct, or be taken over. If you think about it, artificial intelligence is trained on data from humans—at least right now. All data comes from the internet, our fears, our publications. In some ways, we’re enabling this aspect of it “not wanting to be shut down” because it’s fed on all our fears.

Tim Hwang: I love that. I was about to say, “You’re just talking about fiction; this book is nonfiction.” But you anticipated me: regardless of what AI is, it’s informed by data that includes this fiction, so it’s reenacting it in real life. There’s a muddling between the fiction of what AIs do and the reality of what they will do.

We’ve got a lot to cover, so I’m moving to our next story. This is fun and builds from last episode. Last episode we talked about a startup called AlterEgo and Meta doing demos with wearables—they launched new Ray-Bans with AI features. Not to be left out, Apple is dropping its new AirPods (I’m wearing one). The cool thing is they’ll have built-in real-time translation—a dream of AI I’ve waited for: hearing audio and it translates autonomously. Gabe, this feels very different from other wearable AI features. Help me think about why it feels different, or maybe it’s the same as what Meta is portraying.

Gabe Goodhart: I think this is a solution to a problem instead of a solution in search of a problem. That’s why it feels different. We work for an international company; much discourse is in English, but for many, that’s not their first language. As a native English speaker, I feel privileged and a bit bad about forcing everyone to conform to my language.

Language translation—for personal travel, international commerce, anything—is a real issue in a global world. People face it daily. Lots of cool possible features could use AI in real time on a device attached to my body; I might try the demo, think it’s super cool, and put it away. But for folks in a multilingual world, an earpiece that gives translation with minimal effort and smooth UX is a real game-changer. It’s an actual problem and solution.

Tim Hwang: Sandi, is this the beginning of Apple making a comeback on AI? The most interesting reversal: “Apple’s going to kill it on AI,” then “they seem so far behind,” then “Google seemed behind, now they’re doing good.” Are we about to go through another inflection point where by next December we say, “Apple’s got this, we should never have doubted them”?

Sandi Besen: I’ve been getting a lot of ads for Google’s new phones on my TV—I’m being targeted. I was talking to my husband yesterday: I wonder if they’ll take over Apple’s device sales shortly because they’re so integrated with the entire ecosystem. That’s Google’s M.O.: integrate everything.

Do we know the model being used for translation in the AirPods? Has that been released?

Gabe Goodhart: The article mentioned Apple Intelligence, their multi-layered on-device plus escalate to secure cloud, plus escalate to frontier model. I doubt we know specifically, but I suspect there’s an on-device component for the latest iPhones—latency makes sense.

Sandi Besen: I also wonder how much they’re partnering. There’s a history of Apple partnering, which is good. If you don’t have it in-house, you don’t want to just not keep up. They partnered with Google before to make Chrome the default browser over Safari. They have a big partnership with Gemini to bring it in, potentially catching up since they delayed Apple Intelligence for the next iPhone.

More players in the game is good. I don’t think we should have incumbents capturing 75% market share in devices, which Apple has done. Maybe this allows sharing more. Not favorable if an Apple exec is listening, but for a fair market, we should have competition. Maybe this allows competition to shine. I’m not sure it’s the worst thing.

Mihai Criveti: I’ll make a reference back to a book: The Hitchhiker’s Guide to the Galaxy by Douglas Adams, 1979. He predicted a lot of this with the Babel fish—a fish in your ear translating languages in real time. He’s credited with creating the concept of a tablet: The Hitchhiker’s Guide was pocket-sized, had all information, you could put the Babel fish in your ear. These are things humans have wanted for a long time, came up in fiction, movies.

We’ve seen implementations over the years. I remember this feature in Skype (no longer a thing)—you could do real-time translation. If you have family speaking a different language, you could translate. It’s down to user experience. Making it part of headphones is more conducive for folks not used to technology—they won’t download the model or go into Skype.

The first iterations will have teething challenges. I expect funny situations in many countries: “I can now speak Italian!” “No, not quite.” Especially if using the local model first. But that’s good. I see AI becoming integrated to the point where, from a consumer perspective, it disappears. You don’t care what model, what a token is. You just want it to translate in real time. Done.

Gabe Goodhart: Your point about UX is spot on, Mihai. That’s what Apple is getting right. To the question about Apple’s comeback, that is their avenue: UX that, as you put it, has no reference to tokens. It fits into something already part of your ecosystem, makes it better, solves a problem.

Another incumbency Apple has is style and popularity. Tim, seeing you in those white headphones—people walk down the street with them all the time, especially teens. It’s a look. Now they solve an additional problem without making you put on clunky glasses or a heads-up display that makes you look like a geek. It fits into an accepted thing in everyday life. The UX is spot on; that’s the avenue with real traction.

Mihai Criveti: Look up the United Nations translation device online. You see them at the UN—an awkward, cheap-looking thing with antennas. There’s a person behind it, but still, it’s not a look.

Tim Hwang: Sandi, final thought.

Sandi Besen: I’ll wrap with a quick story. Earlier this year, I was in Japan on a solo trip, getting lost in train stations, having to ask for directions. I used GPT for translation—it’s semantically better, voice mode. I’d pull out my phone, do the translation, hand it to an older gentleman. He had no concept how to participate. He’d speak but not press the button, I’d press too late—it didn’t work because of the learning and skills gap. If they can nail a way to solve this, bridging that gap where you don’t change your learning pattern, just the experience, that would be great.

Tim Hwang: This will be a super interesting interface. I want to keep coming back to where AI will be in the wearable ecosystem. Even “seamless” interfaces have a lot built into them. Lots to talk about.

I’ll end on a final story breaking this week—a headline that could only come from 2025: NVIDIA is investing $100 billion in OpenAI. I’m laughing because it’s absurd. Let’s take the last few minutes to talk about this. The top-line number is mind-boggling. One thing that occurred to me: NVIDIA is giving OpenAI $100 billion. Isn’t OpenAI just turning around and giving it back to NVIDIA? It’s like stock buybacks. Actually, we should read it as OpenAI gives NVIDIA $100 billion back. Some will go to personnel, but yeah. What do we make of that? Seems weird.

Mihai Criveti: It does. But look at stock buybacks—everybody’s doing it, making billions. Oracle stock blew up from a similar event. It has a big financial impact, and they’re sponsoring their biggest client. If OpenAI goes down, they’re in trouble. An article said NVIDIA has 3-4 large clients, then the rest, plus gamers with GPUs. Those big clients are important. There’s also opportunity for AMD or Intel to come in with a solution, and OpenAI might diversify. By investing, they’re locking in their biggest potential target. There’s no indication OpenAI will reduce GPUs needed. They’re still making models freely available for a billion users, continuing on NVIDIA GPUs. It’s a brilliant move with massive effects.

I wish for more diversity in inference providers—AMD, Intel, give us more GPUs, make them better. I have an AMD GPU; I love it. But when it comes to software, CUDA and what NVIDIA does still has the lead.

Tim Hwang: Sandi, is Anthropic in trouble? They seem left out. Unlimited money right now—they just raised an $8 billion round, raised a $13 billion round before. They’ll be okay?

Sandi Besen: Maybe they are in trouble. Maybe this will be a catalyst for them to raise even more to compete and stay in the ring. I feel like we’ll see more partnerships, more alliances—like clubs. Something I wanted to comment on: the amount of energy these facilities put out is enormous. Meta has some of the biggest scale projects; this out-competes that tenfold. They expect ten gigawatts of power from the facility—like a billion light bulbs at a time. That’s insane.

Tim Hwang: Maybe you can take us home with a final comment. This idea of “tribes” is interesting. It feels like initially: Anthropic pairs with Amazon, OpenAI with Microsoft (cloud). Next step: OpenAI with NVIDIA. Maybe Anthropic with AMD. Next: OpenAI with a Northeastern nuclear power company, Anthropic with southwestern wind power. Gabe, where does this go in market structure? Do we see vertically integrated alliances emerging? Traditionally, NVIDIA sold chips to everybody. This puts a thumb on the scale.

Gabe Goodhart: This is a big bet that bigger is still better. We’ve hit a point with really good frontier models trained on existing infrastructure. Huge gains in capability are being achieved not just by training newer models, but by building systems around the models. 2025: the year of the agents. We’ve covered everything about agents. I keep coming back: models just generate tokens; what you put around them adds value.

The obvious loser in moving innovation from models to systems is NVIDIA. You don’t need to buy more chips if you’ve got enough; you don’t need to boost your data center if you can refine models with existing hardware. This is a huge bet by NVIDIA: “We believe bigger is still better, and we’re going to make bigger happen.” They’re putting money where their mouth is—very self-interested, and it’ll come back in chips anyway.

Where does this shake the industry out? It’ll come down to whether big is, in fact, better. Sandi, to your point: power consumption, environmental impact—can we source this power responsibly? We’re hitting breaking points at this scale. It’s uncharted territory; a push to push further so that the next generation of GPUs is still desperately needed. We’ll see where uncharted territories lead.

Mihai Criveti: Up to your point before we wrap up: I’m curious if the $100 billion will completely go to facilities, or if they’re making an investment in the next type of chip or architecture. Is this a play for what’s next?

Gabe Goodhart: Very possible. But the top-line power and scale numbers hint at “bigger is better.” Theoretically, you could invest $100 billion in chips ten times more efficient rather than ten times more powerful, but I don’t think that’s what they’re doing. This seems to push the upper envelope of raw compute. Some will go into the ecosystem. I didn’t get into article details on how much is at OpenAI’s discretion versus building the biggest data center ever and staffing it with their GPUs.

Mihai Criveti: Can you smell the AGI?

Tim Hwang: I suppose we all can. That’s a great note. That’s all the time we have. Mihai, Sandi, Gabe, thank you for joining. Thanks to all our listeners. If you enjoyed this, you can get us on Apple Podcasts, Spotify, and podcast platforms everywhere. We’ll see you next week on Mixture of Experts.