← Back to search

Can 150,000 people save NVIDIA? - This Week In AI

Agents Hour · 2026-11-17 · 19 min
relevance 62 3636 words Episode page ↗ Audio ↗
Show full episode description
Jensen Huang resigned as NVIDIA's CEO on November 12, and a petition to reinstate him has passed 150,000 signatures. Shane and Abhi talk through why the board may have pushed him out, what prediction markets say about a comeback, and what NVIDIA without Jensen could mean for the AI race. Google's Gemini 4 Hydrogen and Alibaba's Qwen4-Max are out, along with a debate about Google naming its models after the periodic table. A new benchmark called Continual-ARC finds that OpenClaw, Hermes Agent and Prime Agent don't really learn from experience, and Abhi explains how it works. Plus Anthropic's pretraining efficiency slide, LangChain's $300M raise, shadcn's UI SDK and Eve 1.0. AI Agents Hour is a weekly livestream by Mastra CPO Shane Thomas and CTO Abhi Aiyer. Mondays 12PM Pacific. 📚 READ MORE Slides for this episode: https://mastra-agents-hour-slides.storage.googleapis.com/s/episode-2026-11-16 Jensen Huang resigns: https://x.com/WSJ/status/2120879589144572204 Petition to reinstate Jensen: https://x.com/Kalshi/status/2122028180441588141 Jensen discusses returning: https://x.com/WatcherGuru/status/2122043721307025915 Gemini 4 Hydrogen: https://x.com/GoogleDeepMind/status/2120940338869055788 Sol vs Hydrogen: https://x.com/CalvinHolloway6/status/2120955650960150614 Qwen4-Max: https://x.com/Alibaba_Qwen/status/2122148127520289151 Continual-ARC: https://x.com/DanielLenton1/status/2120535187587268777 Anthropic's compute efficiency slide: https://x.com/hilbertspaess/status/2121331333654086124 LangChain raises $300M: https://x.com/hwchase17/status/2121279222912352574 UI SDK: https://x.com/shadcn/status/2120951116682682646 Eve 1.0: https://github.com/vercel/eve/releases/tag/eve%401.0.0 📚 MASTRA RESOURCES Mastra: https://mastra.ai Mastra on X: https://x.com/mastra_ai Mastra Discord: https://mastra.ai/community/discord Mastra GitHub: https://github.com/mastra-ai Learn Mastra: https://mastra.ai/learn Principles of Building AI Agents (Book): https://mastra.ai/books/principles-of-building-ai-agents Patterns for Building AI Agents (New Book): https://mastra.ai/books/patterns-of-building-ai-agents WHAT IS MASTRA? Mastra is an open-source TypeScript framework designed for building and shipping AI-powered applications and agents with minimal friction. It supports the full lifecycle of agent development—from prototype to production. You can integrate it with frontend and backend stacks (e.g., React, Next.js, Node) or run agents as standalone services. If you're a JavaScript or TypeScript developer looking to build an agentic or AI-powered product without starting from first principles, Mastra provides the scaffolding, tools, and integrations to accelerate that process. ⏱️ CHAPTERS
0:00 / 0:00
📑 Chapters — tap a time to jump there
00:00
Cold open
00:19
Intro
00:29
Jensen resigns as NVIDIA CEO
02:16
A petition to bring Jensen back
03:38
What NVIDIA without Jensen means
05:03
Is Jensen coming back?
07:14
Gemini 4 Hydrogen
08:06
Models named after the periodic table
08:56
Qwen4-Max
10:46
Can agents actually learn?
11:30
Professor Abhi explains Continual-ARC
16:27
Anthropic's compute efficiency slide
17:20
LangChain raises $300M
17:30
shadcn's UI SDK
19:13
Eve 1.0
19:27
Outro
When you're in a position of power like that, you really can't slip up, dude. You can't get caught. If you're gonna do stuff, just don't get caught, man. It's much easier if you just take the ethical and moral high ground. The day after we did our last show, Jensen has officially resigned as CEO of NVIDIA. For those of you wondering why this is important, it was the largest company in the world at some points. If the CEO resigns, this has a big impact. It started with the Chinese spy and then there were some other things coming out over the weekend, essentially. Yeah, and I think we don't have all the answers, of course. I think there's investigations ongoing. One, someone illegally leaked this information about the spy because it was classified. So someone broke a law to leak the information. Two, was it originally a New York Times or Wall Street Journal? Whoever did the initial post, right? Then there's obviously some speculation on if the individual is a spy or not. I don't think that's actually been confirmed. I think there's a lot of skepticism, but it also could be. And obviously that wouldn't be good. And then I did hear a lot of talk on why Jensen would resign if the investigation hasn't been completed by the board. And some of it would be, if you're a public company, you have to not have conflicts of interest in potentially sleeping with a customer. Which was, whether a spy or not, is a conflict of interest. And if you're on the board, you need to kind of protect yourself. You have insurance if you're on the board of these public companies, you need to protect yourself. So there's some speculation that the board is kind of forced to ask Jensen to step aside because they're potentially at risk if Jensen was having some kind of a relationship with a customer that purchased a lot of NVIDIA products. And they did. And they did. And so then this came out over the weekend. There was a petition to reinstate Jensen. The argument here is that the board is trying to protect themselves. They want to limit their liability. So by getting rid of Jensen, they're doing the right thing for the board, protecting themselves. Not necessarily doing the right thing in the eyes of the shareholders. And so the goal here, I guess, is to get enough people to where it can go to a shareholder vote to bring Jensen back. And they got 150,000 signatures. A lot of those are shareholders. Not all of them, I imagine. And, of course, then you've got the prediction markets trying to predict if Jensen will return. And I've seen some different markets go all the way up to 15%. Right now I think it's like 8% to 9% that Jensen will return as CEO. Dude, I signed that petition. There's a bunch of employees on there too. Huge percentage of employees. I was also hearing that President Trump was saying, hey, we need this. Yeah, I heard that 90%, and I don't know if this is accurate or where this data came from, so take it with a grain of salt. 90% of employees favored or were in support of Jensen returning. So I think he was relatively well-liked as a leader. Whether you agree with the mistakes he made or not, I think people generally that worked for him I think he was a good leader and seemed to really want to continue working for him. Yeah. There's a shock though because of how he runs the company where it's very through him, let's say. It's going to hurt a bit if they don't figure it out. Yeah, and I think it kind of hurts the United States a bit. Because if NVIDIA slows down, let's say they have all this momentum, they slow down. Well, of course, other US companies could step up, and I think some will. OpenAI, Anthropic, Google, all working on their own chips. But also, maybe China starts to catch up a little bit too. They're also working on their own chips. So that's the risk. And I think if you look at NVIDIA, they were basically kind of propping up the economy a bit because they were funding all these deals to allow Anthropic or whatever to purchase these chips and basically kind of like moving money around to make sure that these data centers can be built, these chips can be allocated. And the question is, will the next CEO be as bold as Jensen and take as many risks? It is very challenging if you're not the founder to take big risks when you're trying to look at what's in the best interest of shareholders. Oftentimes, if you think about going from Apple, Steve Jobs to Tim Cook, there's kind of a little bit more risk-averse, but still did a great job, arguably, but didn't maybe innovate the same way that if you had the founder at the helm, you probably would. So I think that's the risk that a lot of people are starting to weigh in, and the markets are responding to that, which is why NVIDIA stock has kind of tanked a bit. Well, I mean, the market has tanked a bit, right? The market in general, but NVIDIA, even before the market really started to take a hit. Yeah, it's a bad day to own stocks. Hey, if NVIDIA goes under, you guys might see me under a freeway. Just letting you guys know. And then this came out, Jensen is apparently discussing returning, and so maybe trying to figure out what he would have to do or how that might work. But it sounds like he's maybe starting to think through his options and try to make a play to return back. It is funny, because of course you had the whole Sam Altman, which OpenAI wasn't a public company, so a little bit different set of regulations, requirements, all that. But the whole Sam Altman thing where he was ousted by the board, and then employees kind of rose up, brought him back in. Maybe the same thing happens with NVIDIA, it'd be certainly a different scale. Yeah, like the whole world scale. Yeah, but pretty crazy nonetheless. So what do you think? Is he coming back? I really hope he comes back, for my own stock portfolio's sake. But it's just such a weird, it's just such a strange case. I also was watching All In. There's always a conspiracy with a lot of stuff, right? And so politics comes into play again. I just fucked up. Yeah, I mean, there's a lot of speculation that it was like a hit on Jensen, because people didn't like the way he was pro-AI for everyone, pro-open source. I don't know. They said it could be political, could be motivated by others that are less pro-AI for all. I don't know. I don't know if I believe any of that, but you have people coming out, and there's always conspiracy theories. They're entertaining. I just, you know, I don't put a lot of stock into them until we see some proof. But I do think, overall, I don't necessarily agree with his decisions, but I do think that there's probably no one I would rather have still being CEO of NVIDIA than Jensen. So that's the confliction that I would be in, is I would probably want Jensen back too. I own a little bit of NVIDIA stock. Not a lot, but enough to where, you know, I'd like to have him back. It just shows you when you're a CEO, you can't really, or when you're in a position of power like that, you really can't slip up, dude. No. You can't get caught. If you're going to do stuff, just don't get caught, man. It's much easier if you just take the ethical and moral high ground and not slip up. And just as a caveat for those watching, we're talking a lot about stocks. Nothing that we say is financial advice. Yeah, no financial advice for sure. No financial advice here. We're just bullshitting. We're having fun. All right, let's talk about new models. So meet Gemini 4 Hydrogen. This was from November 12th. This is from Google DeepMind. It's built for rapid coding, multimodal synthesis, and everyday workflows. It finishes tasks 2.5x faster than 3.8 flash and surpasses argon on the artificial analysis intelligence index. Interesting. So if you want to look at some of the benchmarks, you can see it's compared to Sonnet, it's compared to Terra. Over the weekend I tried Gemini Argon, and it was pretty good. I was not doing coding work. I was just doing some reasoning type stuff. Yeah. It's like the 16th element. You got carbon, which is like six, right? Or something. And then you got hydrogen, which is one. So hydrogen, then carbon, then argon. Everyone has a stick, you know? They all gotta have their stick. Flash wasn't good. They say Gemini 4. No one's like, no one cares about that shit. But now we have Gemini Argon and hydrogen. It sounds fancy. It sounds good. Yeah, it's like I'm gonna get one of those periodic table shirts so I can look down and know which model's bigger. Yeah, but it also separates the model class from whatever we were talking shit about before. You know what I mean? Rebranded. Yeah, it's a little confusing, but I suppose over time everyone's got their thing. So this came out yesterday. Meet Quen4Max, your new AI teammate for work, creativity, and discovery. Quen4Max is built on a new mixture of experts architecture with native vision and a 1 million token context window. So let's take a look at some of the benchmarks here. You can see for coding, you can see a lot of different benchmarks. There are more work-based benchmarks. You have HealthBench, you have Humanity's Last Exam. It looks like it's comparing it to Quen3.8 Max, which is the previous one. It obviously scores better in every benchmark compared to the old model, which makes sense. And it compares Opus 5.5, GPT-61 Sol, GPT-6 Astra. And it is, looking at it, good at a few of the tasks, right? Deep C, it's apparently number one now. Terminal Bench, it's a little bit further behind. Some things like Program Bench, which I'm not that familiar with, it's a little bit behind. So some benchmarks, it's very competitive though. Biomystery Bench, apparently top of the charts. Dang. There you go. If you're doing bioinformatics research. I haven't tested it yet, it just came out. I'm always interested to see how these things land. But it's good to see more open-weight models continuing to come out and still be competitive. Now that we're paying for tokens, we've got to figure out what the right token mixture is. Yeah, it turns out when you have a factory that just burns tokens, subscriptions don't go very far. No. You've got to test out open-weight models, keep things reasonable. We are deep seek users now because of the price. We are deep seeking our way to victory. Three editions, one year. The agent world moves fast, but we've got you covered. 2026 is all about coding agents, so we've got skills, sandboxes, and file systems. Get the third edition now at Mastra.ai. I wanted to talk a little bit because I saw a few posts throughout this week, but can agents actually learn? So let's look at a couple things. The first is a post from Daniel Lenton. It says, Open Claw, Hermes Agent, and Prime Agent don't meaningfully learn from experience. On Continual Arc, our new benchmark at Let's Unify AI, none of them were able to memorize anything of value, and none were able to generalize to new tasks. Training a small 9 billion models weights did both. So the argument here is that the memory systems in these agents couldn't actually learn. It maybe memorized certain facts or whatever, but it couldn't actually learn and discern. The benchmark basically has a set of puzzles trying to see if it would actually learn it. Did you have a chance to look at this? Yeah, I have a little thing I can share. Professor Abhi now. I'm going to become Professor Abhi. This article is super relevant to us at Mastro, so I wanted to learn more about it. Because we're doing agent learning as a company. And so I first was thinking, how do agents learn? And learning is obviously a very loaded term, but first you learn in the weights, which is like fine-tuning, continual learning. Or you could learn through context where you edit your prompts and you have better skills and better memory, and you have knowledge primitives and all that stuff that make your agent more efficient with context. Or you have to learn in the system where your tools fucking suck, your workflows do not actually help you, you have terrible retrieval metrics. You've got to learn all these things about all these different levels. And you can improve your agent in all the different levels, for sure. Especially if you don't know how to learn the weights, you could totally make really good progress just doing the other stuff. But when you're talking about the weights, it feels to me like this ambiguous thing that's going on that none of us know how to do. And I want to change that, just personally. I feel like I want to change that for myself. But there's this thing called ARC, which means abstraction and reasoning corpus, where can an LLM reason about unfamiliar problems, rather than doing pattern-based matching, which a lot of these, learning in the context, is about patterns. So it does the right thing based on what it did before, and you kind of nudge it to do that. In ARC, the problem is completely new. And so it needs to figure out what to do about it. So they do it through these puzzles, these are ARC puzzles, and you give it an input and it tries to do an output, and that's all cool. Same kind of puzzle. I think they look really cool. They're all visual, you know, they're like visual puzzles, and that was dope. So I looked into this continual ARC thing, and it's kind of the same concept, except when an AI encounters the same kind of problem, did it learn from it last time? And then how can it actually progress through this problem, like in the current iteration of the problem? So it has many steps. First step, you encounter a puzzle. So you give it the same puzzle. And there's this, the interesting thing is like how, what it does next. Next, it can pay for feedback. I really thought that was an interesting term, pay for feedback. What does that mean? You can actually do a bunch of shit. One, you just solve it, and you have no feedback. So how this works for paying is the run has a budget for what it can do to learn in the run. So first, you pay nothing, solve it. You just say, fuck it, whatever. If you get it wrong, you get minus one point. So like solving and getting good points is encouraged. There's like this teacher, so you can ask for past examples. And that costs like, let's just say that costs like two points. So you have to reason, are you willing to pay two points to get the example to fix the problem? Or are you going to submit an incorrect answer to cost yourself a point to move on? Or can you bank this later as a similar problem that you want to reference when you encounter it again? I just think it's quite fascinating to pay for feedback. Once you do that, you either get the information, submit the incorrect answer, find out what is right, whatever. And then you save all that stuff, and then you update the model weights, and then you do it again. And if you encounter the task again, if you've already seen it because you paid for it, can you do it again? Can you build a skill that does it the next time so you always do this the right way? If you haven't seen it, then you go through the flow again. All right, finishing up here. So arc versus continual arc. Arc is the puzzle. They're both puzzles, but in continual arc, you're having a sequence of puzzles that you learn over time. In arc, you measure things on individual tasks, but continual arc is a long horizon. Over many tasks, how's the retention? And then do you get better with experience? This is a continual learning type of setup. So yeah, that's continual arc, and that's the thing for today. Yeah, and I think it was very interesting that if you looked at OpenClaw, Hermes Agent, Prime Agent, they didn't really learn, right? But if you just trained a much smaller model, and I think they used a bigger model to do the training, to help with the training, to pull out the examples. But then they trained it, and then it was able to outperform, to actually learn. I didn't catch that if it outperformed the much larger models. But you could prove that it provably learned and got significantly better. These puzzles are interesting, kind of, it's not real to me. It's kind of like when they solve math problems. It could be very important, potentially, but also not super practical for the average person day to day. It's not a super practical task. Yeah, and I still don't understand, like, once you do a run of the arc, how do you change weights and stuff? I guess we'll learn with everybody else as we go. So this is from Jacob Coxon, who was previously at Anthropic, had a bunch of drama at one point for leaving. Says that this slide, which was from the Anthropic IPO Roadshow, has kind of been making the rounds lately. All this slide is showing is that it's getting easier to train models. It's taking less compute, so they're getting more efficient, significantly more efficient. And this is like a log scale, so it's actually super exponentially more efficient at training models. Which in a lot of ways is a good thing, because it means it's cheaper to train larger models than it was in October, than it was two years ago. By a large scale. And it seems to be accelerating. It's not just accelerating, but it's also super accelerating, because this is like a log scale. So it's very, very exponential. In my opinion, that's a good thing. It means that models should be able to get better faster over time. But obviously, some people think this is a concerning thing as well. Let's get into the last part, the quick hits. Langchain raised $300 million at a $3 billion valuation. Nice job, dudes. Congrats. Makes me happy. Today, we're introducing UISDK. This is from ShadCN. Agents are becoming both the builders and the users of interfaces. So we asked ourselves, where should the abstraction live? Take a select. There's a pop-up, selection log, keyboard navigation, and focus management. Those same behaviors are useful in lots of other components. So we pulled these pieces apart, tested them, and made them reusable. The behavior stays the same whichever framework renderer you build with, or the harness generating the UI. That's the idea behind UISDK. Give agents the pieces to generate the interface you need, right when you need it. Even interfaces we haven't thought of yet. They're getting into the CPK territory. Yeah. I mean, I think just the whole idea around how UI is changing. Of course, generative UI, I think, is going to become more prevalent. I'm really bullish on MCP apps. I'm just like, the UI now lives in your whatever app you're using to chat with your agent, whether it's Codex, ChatGPT, whatever, Claude, GrokBot. You can actually bring your UI to users. I really think that those applications are slowly becoming the new browsers for people, for some folks. They're not opening Google, they're just opening up their desktop app. And so I think bringing UI to those places is very interesting. And so just the whole concept around UI and how it's changing over time is very, very intriguing to me. I feel like MCP apps, I think we should have apps in the protocols and not just MCP. If I'm in a TUI, I should be able to render whatever, but in TUI form. So you think every interface needs its own protocol? I think for every agent, it should have the ability to generate what we're calling apps in MCP, but it's just part of its protocol. That's probably the future in my mind. All right. Interesting. All right, let's continue on. EVE 1.0 apparently is going to be released soon too. So that's big news for those of you that are building agents and want to use EVE and not use Monstra for some reason. Congrats to them. Getting to 1.0 is obviously a big milestone. Dude, that's the show. That was quick. Follow us on X at Monstra. Go to YouTube, Mastra-AI. You can follow me on X, SMThomas3. Follow Abhi at Abhi Aiyer. What do you have for us? Parting words of wisdom to leave. Everyone have a great week. Stay safe with your CVEs and security. Peace. See ya.