ThursdAI - Mar 6, 2025 - Alibaba's R1 Killer QwQ, Exclusive Google AI Mode Chat, and MCP fever sweeping the community!

Welcome everyone, welcome to Thursday Eye for March 6th.

My name is Alex Volkow. I'm an AI Avengers with weights and biases. yes, still weights and biases. We're going to talk about this, in a bit. And I'm here joined by two of my friends, long time co hosts. We have Wolfram Ravenwolf all the way from Germany coming back to us after two weeks of, being sick and missing in action. But it was fine. Wolfram, how are you, man? How are you feeling?

I'm happy to be back and so great to be catching up now.

Yeah, this is great.

And thank you for joining. And then we also have a long time friend of ours and returning after a while. Junyang, how are you, man?

Yeah, it's good.

Good to be back here. It's been a while.

It's been a while.

It's been a crazy time for you guys as well.

I was pretty busy.

Yeah. I'm very happy to be back.

Yeah, we're gonna talk about what exactly you were busy about.

For folks who don't know who Junyang is, Head of DevRel? I don't know how to say it, but I know Alibaba Qwen.

building the Qwen models.

I am leading the Qwen team now, building Qwen series model, language models, multi model models, and We are working towards AGI.

Let's go, building AGI in Alibaba

Qwen. I love it. welcome folks, welcome. And the reason why I called you here is because, obviously you guys have been around for a while and we've talked about multiple Qwen models but this week you guys released something super, super cool, which I'm going to show to the folks now. And I would love for you to actually do the announcement. This is, as folks know, on Thursday one of the coolest things that happened here. is when the folks who actually build the things that we talk about come and talk about the things they built. And, Junyang, you're probably the most quintessential example of this. almost every time you release something crazy, we welcome you here to talk about it from from the horse's mouth, let's say. we'd love to hear, you announcing this last update that you guys released, yesterday. Yep.

Yeah.

we yesterday, open weighted, uh, QWQ 32 billion. Previously you had the open source model, the QWQ 32 billion preview and it is something playing for fun, but this is something more serious. we have been building this, for a long time. We are thinking about how to do a reinforcement learning to. Make the reasoning models and then, like deep, they are one or it's pretty, pretty crazy. So give us a lot of pressure because you need to push your boundaries. For example, previously, if you build a 32 billion parameter model, you don't have the requirements of reaching about 80 in AIME 24. that is something really crazy. there are a lot of pressure, but we have been building this model very recently, and we have been struggling with reinforcement learning, and we found that, wow, it is something really amazing. Previously, we just think that, SFT, RL, and something like, it may improve gradually, but when we're really scaling the compute in RL, and if you scale with more steps, you will find that, wow. It is actually increasing fast. And consistently, and, I think we still did not finish the job because, it is still very rough because, we have been building this very focused, for only 2 to 3 months, and, I think we need more times to make things even better, but I think it's a good timing for us to share the model you. It's a bit surprising for us to see that the model can reach. very much. Such a high score, even competitive with. A little bit behind R1, but generally comparable to R1 performance in some benchmarks, such as AIME, LiveCodeBench, and some also general domain tasks like LiveBench and Eve Eval, And we have some, advantages in, agent because, there are a lot of reasoning models. They do not focus on agents. So we pay attention to it. So we have, larger, advantages, but in other domains, it is generally comparable to the, state of the art, performance model.

let me just say, this is crazy.

This is a 32 billion parameter dense model, not an MOE. This is advanced, pre trained, which we then reinforced learning like twice, I believe you said. We're going to get into this. That's beating a 671 billion parameter MOE that's been trained by, folks from DeepSeek. And maybe a reminder for folks who are listening to us because, Junyang, I know definitely For folks, even DeepSeek was a surprise for many people, but we've been talking about DeepSeek for a while. We've been talking about Qwen for a while, and the, everybody, the normies and the stock market all got crazy. It was like, whoa, DeepSeek out of nowhere. no, this is this is a legit team in China for a while. you guys have been killing it. Like most of the latest fine tunes are Qwen as well. And Getting these scores with this amount of parameters is absolutely insane. We're talking about 79. 5, almost 80 percent of completion on AIME24. AIME is a competition math. It's like very difficult problems. I would love to see how this performs on AIME25 to see, the comparison, but like incredible numbers overall. And this model is like So tiny compare, before this, I have a question. Everybody's asking, you just said QVQ last time you were here. You mentioned, how we actually pronounce this. So maybe can you tell the folks, what's the story about the name and how you guys pronounce the name?

Oh, yeah.

the story is quite interesting because we don't know how to name our reasoning models. So previously we would just simply call it Q1 and I just think that, we did not. pay enough efforts to, for the naming and some guys call that, why not call it QQ, but, you cannot call it QQ. So we find that, why not call it QWQ? It is something looking very funny. It is really the story because we find that it looks very funny. And then I say, it's something like Qwen with questions, so it can think, it can reason. So it is something that can ask good questions, ask himself, and then generate the good answer. So we would just call it Q, Q, W, Q. And previously we have a pronunciation like quill because we would just want to combine Q and W together. It seems that nobody would call it Quill. So we just call it QWQ as you like. We did on the show.

Because you

yeah, I told you it is called Quill.

I told guys in the blog post, but not everyone is reading our blog post. So not many people are calling it Quill. Just call it QWQ. And it's not really hard to pronounce. I think it's just generally okay. Although it is three characters, but, it sounds generally good, so we just keep the naming. we did not change. Yeah.

I think I still call it Gwen Quill after you told us exactly

how to pronounce this, but yeah, for folks with questions, I love this. That

makes

sense.

That is a much better reasoning behind the naming

than the big labs have shown.

Yeah.

I think a smart person we evaluated because he can ask good questions. So for a model, it is. It's also like this. If it can ask good questions, it can answer questions perfectly.

All right.

let's talk about this actual model. many folks also don't know before DeepSeq, you guys released, QwenQL in preview. And that was like, the first reasoning model that we had in the open, I believe, or one of the first.

Yeah, in the open domain, it is actually the first one.

compared with these models today, they are actually far behind, but it's okay.

So let's talk about It was amazing

when it came out, it was really state of the art at

that moment, and it was one of the best models you could run locally.

time AI was not that popular.

Yeah, reasoning was not popular.

You guys basically open sourced the reasoning model, and then it got everywhere. And so let's talk about this one, and the plans for this one. I would love to hear about the two kind of approaches for RL. I would love, absolutely love to hear how you got there and what did you guys see? Did you experience any aha moments? Kind of like the famous aha moments from the DeepSea paper. we'd love to hear from you kind of any insights about this.

I think, that deep say aha moment is,

very surprising to everyone. if you do the experiments of R one zero, you can see the similar, phenomena. But, for recent months, we are actually focusing on how to really improve the performance of the model, and we try to improve the models like 32 billion parameters. for the first reason, it is simpler for us to build. skating RL, because skating RL is really difficult, especially skating, MOE models, for the RL. So we just do 32 billion, at the beginning, and it seems that things are working good. So we can focus on improving our query set, improving the hyperparameter search, and do a lot of experiments to see that, Whether RL can really improve the performance and for the first stage, we are actually doing only math and code rewards. But we are using verifiers, because there will not be things like reward hacking, just previously, caused by the general reward models, we would just use the verifiers for math and code. And we find that while the performance can. Continuously increase. if you do it quite well, we would do it. We have done a lot of ablations. that's a painful experience because there are a lot of parameters to do, and our guys, including me, did not know the parameters quite well because, like reinforcement learning is Something a little bit strange to us, not too much experience. So we need to do a lot of, experiments to reach this, but we then find that, is this something really interesting? It is very slow, but after several episodes, the training episodes, while the performance just increased step by step, it's very interesting. So we have trained like, 100 to 200, just over 100, episodes. We have large batch sizes and we find that the performance is actually increasing. Sometimes it stops. But when we tune the hyperparameters, it can increase again. So we may think that, wow, there's no limitation. We can even do better models. But I would like to leave the better model to 2. 3.

I have to ask you, a few questions.

One of them is we saw the test time computer in for some compute kind of plots from open the eye and we saw with more inference kind of the better questions as well when you guys did the evils that we're seeing right now on the screen. so like again, 79. Numbers are crazy. Let me just pause again and say these numbers don't make sense to me This is a 32 billion parameter model. I can run on my mac I can run two of them in parallel if they're quantized and we're getting 79. 5 on AIME Like 80 percent of AIME, which is incredibly difficult is now solved with this model can solve on my mac completely It's insane. It's insane compared to a year ago when we were here, we talked about, these models cannot do math. but the question to you is, we've noticed a trend after the reasoning models, both from OpenAI announcing this on their roadmap, and, I think, Entropic also does this, and maybe, Grok also does this, a consolidation of thinking and non thinking model towards, one model that can do both. How are you guys thinking about this? Have you tried exploring this? do you have any kind of thoughts about, is this possible? Is this difficult? We'd love to hear from you about the consolidation kind of effort.

this is a perfect question.

we are actually working towards this. we are trying to build a thinking model and non thinking model together because sometimes you would not like the model to think too much. you would like it just to directly answer the questions. So there are a lot of people doing research like long to short to make the thinking shorter, but Personally, I would not like it to, control the length. I would like the model to, control it by itself. This is something perfect. But, maybe at the first stage, I'm not pretty sure we can finish this, quite well. But, we have a baseline. It is something like, you have a special token to control whether to think or not to think. And we try to reach the best performance for non thinking mode and thinking mode. So we can combine it together. Nobody would like to have two models. If you are using a chatbot, you would like everything in a single model. So we are trying to do this. It is something really possible.

Speaking of a few models, we can take a look at that.

QWQ. yeah, you need to log in. Yeah. After login, folks can try it on Coin Chat as well. And, there's a dropdown there with a bunch of models. Not as many as open the eye. Open The eyes now is craziness, but I think I know what you're talking about when we're talking about multiple models. We want them to be combined as well. Like we started with a different model for vision and then those combined as well. all right, Jing, thank you so much for hopping on and talking to us about this. One last thing that I wanted to ask you, obviously, as always, is. What's next for Quinn? I always ask you, I always know you can't answer, but I would love to hear anyway. And I love to put some pressure on you as well.

a question that in that regard, because, okay.

Many questions. Just one for me. do you plan to integrate as a vision model as well? Alex said, instead of having the 70, you guys

vision reasoning model.

this is something that we are working on.

We are also working on QVQ. It is something, when you have finished the breathing model for language models, you can naturally move it to vision language model. It is something actually really possible. But we are now thinking that this year we might build omni model, just like open AI does. we are building one single model which can, process multim modalities and it can reach. Very good performance in each modality. This is what we want. We will want to make it together. Yeah. So we'll have an omni model and omni reasoning model.

That's amazing.

wait, so we need you on the livestream to be able to ask questions. Next time join livestream.

I'm on the road.

I can't be unless you wanna see the, but are we able to just use the model in non reasoning mode? Do we just need to put, The reasoning token and the other question is, can we end the reasoning short by just like injecting the end of the reasoning token on it?

for, for the first one, you can, use the special

token to control the thinking and, if you put empty thinking. Path. for example, the special tokens together with the thinking path empty. So you can make it, generate the answer directly, but I cannot ensure the performance because we have trained it with the thinking path. And this is for the first question. And, for the second question. what is that? Yeah.

Folks, we'll have to get to the second question next,

because we have a busy show today, as I said in the beginning, and now, Julian, thank you so much for coming, and we appreciate it, we know it's late for you, and congratulations on this release, this is incredible, folks, if you haven't tried it. should definitely play with this. Feel it on your Mac. It runs incredibly fast and quantized. The folks did incredible job with two rounds of RL and I really appreciate your time here. As always, feel free to stick with us, as we move forward, and to talk about the next kind of, folks and the next guest that we have here, Jane, thank you so much. All right, folks. as a reminder, Thursday is a live recorded podcast and newsletter The fireworks going on behind me for some reason, that is translated to you every Thursday for the past two years. In fact, in a week, it's going to be our two year anniversary. We started when GPT 4 was released. And as I told you before, it's an incredible honor of mine to have folks who worked on the stuff that we're going to talk about. actually. The show so much today that we don't even have time to talk about everything we're going to talk about like usually but we'll Get to the TLDR later. You have to stick with us But now we have another guest that also is going to tell us about something they released and Honestly, I would love to just introduce Robbie to the stage Robbie welcome

Hey, how are you?

Good to see you

Alex.

Good to see you as well. we just met basically, we just met this week and the reason we met is very interesting. And, first of all, we'd love to hear from you, but since this is your first time on Thursday, I would love to hear from you, maybe an introduction brief about yourself, and then we'll dive in

Yeah.

I'm Robbie Stein. I work at Google, work across a bunch of Google search, experiences and products, including a lot of the work we do, on AI. And, before that I've done, a couple startups, worked at other companies as well. been, a Googler now, for the second time I started out here, originally in 2007. And then I'm back as of the last year working on Google search.

And, Google Search, Robbie, is something that we absolutely

mentioned multiple times on the show. And we also mentioned multiple AI approaches. We've been doing this for two years, and for these past two years, Gen AI, LLMs, it came everywhere. And then, slowly these models also got access to the internet, started searching, started answering questions in real time. And, I will tell you with full honesty, with full transparency, on this show, we mentioned the words, where's Google for the first year of the show, pretty much every show we were waiting. We were waiting. And then I had the privilege to go to Google IO and kind of experience, and then we saw the rise of Gemini one and then Gemini two and we're big fans of Gemini here. Absolutely. I've been using Gemini extremely long contacts, runs incredibly fast as well, flash and the thinking models as well. And We're also been waiting, like what's going to happen to search, like what's going on with some AI over there. So you guys have an announcement this week and you made a splash and also I got to use this as well. So maybe let's talk about what you guys have announced today.

Yeah, sure.

So there's a few announcements. One is, we've obviously been really excited that AI overviews, which we launched last year is now being used by over a billion people. It's a really exciting product experience. It lets people get quick and fast information and also go deeper on the web. So that's great. And we were bringing as part of the announcement yesterday, Gemini 2. 0 To AI overviews and starting with some of the more complex questions like math and coding and multimodal questions, images, et cetera. It's where really the Gemini reasoning capabilities and advanced multimodal understanding shine. The second area that we announced was this new, a new experiment in search labs called AI mode. And really what this is about is being able to. Ask whatever's on your mind and be able to get an AI powered response. And what we found was that people have really enjoyed using AI overviews. They report higher satisfaction in search. They enjoy their experience better. And what's nice is with this experience, you can now have AI help you with a much wider set of your questions. And our goal is that Google search can do so much more for you. So what's nice and what we think makes this special is really. As you were mentioning, Alex, like another combination of Gemini 2. 0 state of the art models, advanced reasoning with all of the information and fresh systems that we have in Google search does knowledge of the web. It's information about billions of products, businesses all over the world, updating from merchants from local businesses every day, Google finance information. All can be brought together to be incredibly helpful for you, and we've really tuned this to be as factual and accurate as possible. Obviously, with new AI projects, we always understand how they make mistakes, and we learn quickly there, but really, it's about bringing that together to be able to answer almost anything you might have in mind as best we can.

that's awesome.

First of all, maybe let me start with folks who maybe didn't use AI overviews. Could you tell us the difference between kind of these two AI approaches from Google search? what is the actual difference? How are they showing up where they're showing up? we'd love to hear. And then, I have a follow up question.

Sure.

So if you go to Google search today, and in most countries in the world, this is now launched. You type in a question. Maybe you want to know how to remove a stain from your or something. You're very likely to get a small AI overview there that gives you helpful information. Synthesizes, a bunch of helpful content, and then also, provides links to dive deeper and connect to the web. that's AI overviews. Now what we've realized is that you can bring this experience to many more questions. And right now you only see an AI overview if search believes that this is going to be helpful to the given question. And so with AI mode, what this does, it lets you always get this kind of an AI powered response for anything you have in mind. And because it's taking advantage of Gemini 2. 0's advanced reasoning, it has sophisticated reasoning and thinking on top of it. And we can take a little more time to be helpful in this unique space that we're building out as an experiment. And so what happens under the hood is with this enhanced model, we take the input of the user's question. We breaks it down into different parts, and the model actually makes a plan and generates fan out queries and actually uses Google search as a tool. It does a bunch of Google searches under the hood.

I find that incredible.

I find the fact that what you said, the Google is Googling basically, this is to me like a little bit of crazy. does it use like specific tooling behind the scenes? Does it use the same Google search that I have? Could you elaborate a little bit more?

Google search.

it has access to information that also includes Google's, knowledge and information systems of the web. So that's something that it can use, more fundamentally to understand, Hey, what are the websites that are most related? To this given question or to the set of questions, and so it can consider a much wider set of potential websites to show you and information to show you than a traditional Google search, because it's actually doing much more than the single question that you're asking.

I love the ability of following up specifically, and I

think that this is maybe something that, not necessarily possible. The eye over you can just have another one. There's no context. I specifically. Did multiple tests of this one video of mine is on YouTube folks. Definitely. I'll add to the show notes You can check it out my full live reaction blind reaction, by the way, Robbie if you're watching this you guys gave me access I immediately recorded without like even knowing what's going on, right? So that was like a full blind reaction and I'll add this to the show notes as well And at one point I was asking it about some places and as we talked and I mentioned this to you as well one of the Problems with LLM overviews about places specifically that I ran into on my travels was, specifically about places that are closed. It happened multiple times. Nisten, you were there, we went to a place and that place was closed for a while now and, the model did not know about this. It just picked up a summary of 10 best places in Toronto and that, article was SEO to oblivion like for the past like Seven years and then for the past six years the place was closed and we still got it a recommendation We still got like a link and we went Does this happen with this or are you using any kind of proprietary stuff to actually let me know what's open and what's not open?

Yeah.

our hope is that this is going to be as accurate as possible. So it is able to tap into a bunch of that real world knowledge and factual information around places, closing information, et cetera. and the model considers that, and you actually may see that show up, in the response itself. And it may even provide a link directly to that restaurant as well. So you can go verify that information yourself. I think that's like combination we think is really powerful here.

Yeah.

A lot of the times kind of the closing information exists. Robert, this is a technical show and I want to kick out about what's going on? Is this base Gemini behind the scenes? Have you guys done any like specific training for the Googling stuff? Is it just using Google searches tool use as, as much as you can and want to, I would love like to get a deep dive into what's going on behind the scenes. this acts really fast. Obviously you guys have, A bunch of TPUs in there, and Gemini is generally running really fast. the fastest Gemini is Gemini Flash. are using the bigger model. As much as you can, and I, obviously, being aware of, restrictions about different things, internally, I, we would love to know, like, how, is this the thinking part of the model? are you using, thinking tokens behind the scenes? You, you mentioned something about the plan. we'd love to hear some technical details about this. I'm sure our audience will as well.

Sure.

I can give you some general components of how this thing is structured. So first of all, it's based on a Gemini 2. 0 model, but it is custom for search and it's fine tuned for search. So we want to make something that is really great at informational needs. So you have a specific question about a restaurant, or you want to know if you should buy certain shoes, or you want to compare a couple of products before you buy it. That's a very specific formulation of the problem. And so for us, we spent a lot of time helping really train and design the model to be good at those tasks. And what we found is that there's a few things. One is it does have to take advantage of these thinking and multi step reasoning components. So it will make this plan do queries, use tools as needed, bring information forward, and then it will consider its information as it's forming the response. And so one of the unique things we did, the reinforcement learning techniques, and I think pretty advanced ones that we were excited about, is to really have the model try to back up what it says. As you said, one of the big problems, one of the challenges with creative LLM, which do a lot of amazing things is that they don't always get it right. And by the way, this won't get it always right either. but what we've tried is to improve on this. So that the model actually thinks about, Hey, is this the right thing to be said? And by the way, if it's an area that is more sensitive, like maybe it's a health question, it's extra cautious claims that it's making, and it really tries to figure out, if there's information that can verify. Or back up what it's trying to say. So we hope that this means that the responses are going to be more accurate and more helpful. And also from a real time perspective, because it taps into all this fresh information, it's hopefully more helpful in these cases where something's happening now, or there's information like, closing data that really depends on the context of the moment. and all of that happens. So as the response is being rendered, if it even thinks, Hey, there's not a lot of information for this specific part of the response, it may actually. Try to find more information about it. So it's really this multi step reasoning thinking approach, with reinforcement learning. That's a really powerful combination. That's incredible.

Robbie, I know, we need to be mindful of the time.

So just a few minutes, I want to follow up maybe with the last question. this will come up, maybe a two parter super quick is can people use AI mode now. And, or is it like just a close, for so far and second of all, as much as you want to get into this with the last one minute, how does this affect Google search as a whole? Like how does this affect websites for with whom, getting provided information? We'd love to hear from you those two things super quick.

Yeah, it's not available widely yet.

We're starting with the labs roll out in the U. S. Specifically, and the Google one I premium subscribers are gonna have first access. So we're working on rolling that out and getting that to as many people as possible and hopefully with the right feedback. Obviously, part of the process is learning and understanding where it can improve and how we can fix things. So we wanted to get that feedback and make sure that it's really ready for more people to use it. And then, as it can be expanded, we hope More people can start using it and you'll hear from me on that. hopefully soon, for us, we view this and how it fits into search as a complimentary experience across all search searches. Vast people using it every day for all kinds of needs. and Google continues to send Billions and billions of clicks out to the web every single day. We don't see that changing, but we're seeing is that people are asking new and different kinds of questions with these AI experiences. You can ask questions that don't exist on any single webpage. You can ask questions that require you to access data that doesn't even exist really anywhere, except for trying to look up some historical stock data from three or four years ago. and then do analysis on it to give you a great response. But those are opportunities given how the modeling works and these fan out queries for websites, hopefully to rank. And really in our mind. it's been a core design principle to bring the richness of the web into this experience and be able to make it easy for people to click in, verify information because we think we see people want this. they want diversity of views. They want to see what the experts are saying, and they don't want to trust any single source for information. So that's going to continue to be our approach, and we hope that creates more opportunities. down the road for websites to show up and rank in ways where maybe they wouldn't have the opportunity to do that before, given how the model works and given the growth we're seeing and how people are using these AI experiences.

Well, Robbie, I know we're mindful of your time.

You have a hard cut off. So first of all, really appreciate your time. And you guys let me play with this. I'm really enjoying this. I switched it to my main search. That's awesome. That's great to hear. I'll definitely give you feedback. One thing about Thursday is the community. So if they have Opportunities to give you guys also feedback. we'll do the connections and we'll send you like more testers who know some prompts and we'll trick you. I really appreciate this. Google is back. That's how it feels. And so thank you for coming up and once you guys have more releases, feel free to, come and tell us about them as well.

That sounds great.

Thanks Alex. I appreciate it. I'll see you soon.

All right, folks.

you heard it, here, Google AI mode is now in labs and it's, you will get access to it soon via the Google stuff, the AI overviews. That's what the other thing that Robbie said, AI overviews have now been, rolled out generally, and they're using Gemini two as well. So this is kinda also part of the big news. But now we have, a pause, to talk you about, this is A-T-L-D-R of everything that we need to talk about. Wolfram. we have so much to cover this, but we also need to be mindful that we have guests. Niton hopefully will join us as well. And LDJ welcome as well. What's up man? the next part of the show with guests is going to be talking about cps. CPS are crazy popular. Now, on my timeline, we're gonna tell you all about, we're gonna have, a few experts in CPS in just a few minutes. But before this, I just wanna run through a TL DR just super quickly so you'll know the news, right? we're a news show. We need to tell you about what's going on. we may not have time to dive into everything, but at least. You know that you'll know what happened, and you know that you'll have links in the newsletter as well. So let's do that. Let's do a TLDR because we jumped in straight with Junyang. By the way, Junyang had to drop. So huge thank you for Junyang. Huge thank you to Robbie Stein, from Google, for jumping on. Let's run through the list, folks. The usual list, this is the notes that I have to prepare for the show. I prepare them, and you'll get these notes in the newsletter as well. Obviously, we started with Qwen's latest reasoning model, QWQ, 32 billion parameter matches DeepSeq R1 on some evals that comes very close in this small, tiny, 32 billion parameter package. We're super, super excited to have Junyang Ling from the Qwen team talk to us about this. folks from Cohere released, Wolfram, I'm assuming you saw this. This is a multi modal vision, AYA version. Cohere for AI released AYA vision, specifically for multilinguality and, in the vision models. And that's crazy as well. Wolfram, do you have a chance to look at this one?

I didn't look at it yet, but it's on my list and I

want to compare it with the Qwen 2. 5 VL model to see, especially in German and twist German text, which is better.

Yep.

so definitely looking forward for your feedback and folks give Wolfram a follow for this and you'll hopefully hear, his thoughts on this as well. Also in open source, LLMs are friends from, AI21. Today, just today announced a new release called Jamba 1. 6. You guys remember we talked about Jamba. Jamba is their hybrid models. And 1. 6 looks very impressive. It beats Mistral Large and Llama 3. This is a hybrid, SSM and transformer model that boasts significantly longer context and throughput as well. so shout out to folks from AI21 for this release. context of 256k parameters, like there is two versions, a large one with 90. Four parameters, active and 398 total. this is like a big chunker MOE. LDJ, you remember Jamba? We talked about Jamba and hybrid ones. What's your take generally on these models?

Yeah, I think they're pretty cool.

there's some interesting papers showing that they have some. Benefits and downsides against transformers. Some benefits include, in certain types of niche operations, they're able to outperform transformers and in certain other ways they fail more, but where you actually have a hybrid of transformer layers and these mambo layers and interleaving those, it seems like theoretically that could give something overall better. I think people have still had issues like doesn't feel the same, feels a little bit rusty around the edges. Yeah. And some people hope maybe that's just those companies needing to get their post training polishing and all of that up to speed with where other labs are. But yeah, I guess we'll see where this ends up.

Yeah.

So shout out to, thanks LDJ, shout out to AI21, for the Jamba 1. 6 release. And folks are welcome to write, they have a specific license there. Called, open model license, Jamba open model license. So you can use this, I believe, but in commercial situations, you maybe need to talk to them. They also offer an API as well. Also in open source, let's skip directly to vision and video folks. You guys remember Tencent HY video came out and we talked about this on the show. And we mentioned that H1 video has all these capabilities, image to video and puppetry, Wolverine, like even animates voices, et cetera, that model was crazy, but we only got text to video from this model. And, today they, very like happily announced and they call it who knew on video, H1 video from Tencent and, image to video is now available. And this is like the second image to video we got. So we actually didn't talk about. But last week we got one X from Alibaba as well, which also has image to video. And this time we also have who knew on image to video, the folks, the proliferation of video models is crazy. Remember a year ago, so I just came out and announced me. We were like, Oh shit, this is like life changing now. I actually needed the project with like video Wolfram. I don't know, if you're with me in the same boat, I knew the project with video. I went to file. ai are very good friends who have all the models. They host them. they also pipe the APIs. It's really hard to choose between which video model you want and even professionals in video modeling, they will tell you, this is richer. This is following instructions better. this has better motion capture. This is like blah, blah, blah. I'm getting lost. Like I need the king video model folks. and hopefully who knew on video is it. Have you tried?

And weaknesses, all the different strengths and weaknesses.

They have so many subscriptions to have.

Yeah, absolutely.

This one specifically, you can try on a video, the tensor. com. And I think that I even have an example for you guys. Let me see if I can show you an example. So I did, Oh, this one is still rendering or failed for some reason. but basically this model came out as open source, but also possible to, run this on their website. And so I actually rigged myself doing a little dance. I want to show you Let's see. It's just like a little bit of an image of me full body and then somebody else's motion and dancing so tencent shout out to tencent, about when you're on video and they also have some Text stuff, but we'll cover this once this gets released. I think in the open source that is oh, yeah One last thing I want to talk about the open source super quick is notagen symbolic music generation That does high quality classical music sheet music it generates sheets and then it generates like high quality classical music When I say high quality to my untrained ears, but definitely I know that folks who are listening to this, who actually know the classical music, they, did not have a really good time, but it sounds super cool. So it's called Nordic gen. I think it's only a fully open. You can make full sonnets and you can hear some stuff and. I think that this is it in open source of very important things. so let's move on to big companies and APIs in TLDR. In big companies, Google, obviously we talked about AI mode. AI mode is Google's kind of, reply to perplexity or Google's attempt to let Google do the Googling for you. We just talked with Robbie Stein, the VP of a product that Google search, about all that jazz. And, my live reaction video is on YouTube and you guys are more than welcome to check it out. 20 minutes of me just getting Whoa, Google is back. Also folks, open AI Wolfram probably you saw this as well. I don't know actually if Germany supports it or not, LJ, but, GPT 4. 5. Has rolled out last week and it's

here

and it's now, rolled out to plus members as well.

not only pro 200 a month, which by the way, I canceled on Dell OpenAI, I canceled my pro, but, they rolled it out to plus including deep research as well. So folks on the plus tier. Now you basically get everything that the pro users got last month. 4. 5 is dope. The vibes are immaculate across multiple areas. it's really as they said, one day we're going to release this. It's not going to top the charts in 4. 5, but GPT 4. 5 did actually top the charts. It's like the number one, model right now in LM arena, including style. And it's like by a big margin. So this is a huge pre trained model. Not a reasoning model. Like I thought it was, and maybe I confused doing less chat, less stream as well. Nissan, if you're with us or you're driving, I don't know, Nissan, what's your take on 4. 5, what's your vibes? And while Nissan pulls up the microphone, LDJ, what's your vibes on 4. 5 from you and the community that we've collected after a week of it?

Yeah, I think it's creativity is I think one of the main

things people are talking about. emotional intelligence. It's funny. Humor. Yeah. And I think, teaching as well. I think that's something that people are underwriting as well. I've seen some cool use cases of just, especially if it's something like philosophical concepts or even some very high level scientific concepts. Yeah. That kind of required, I guess more of what you would call intuition. It seems especially good at that and just the ability for it to better understand your intent and better follow instructions, but know what you're desiring. And I think that helps that overall aspect of teaching creativity, humor, and everything.

there's the whole, approach.

I don't know, big model smell. And this has been like a saying in our environments. I don't know if we mentioned this ever on the air, LDJ, but this model has big model smell. This feels like a big pre trained model. They spent a lot of time. It's not distilled, like a 4. 0 was or like a turbo. this feels like the first big model from OpenAI since actually two years ago. That we launched with GPT 4 and we started the Thursday. this actually does feel like this. And, they made it faster as well. So like at the beginning, it was slow. Definitely up to 4. 5 multiple times. So folks who are listening, if you don't know the reason, the difference, do you feel the big model smell from this model?

Yeah, I do feel it.

And I think a lot of people are going to, make fun of it, feel like it's not enough of a big deal, or not enough of a leap in terms of coding or STEM or anything like that, but I think. definitely as soon as it's added to plus, which I think it's already being rolled out to plus right now. Then yesterday, full rollout. Yep. Yeah. I think within weeks, everybody's just going to go out of habit of using 4. 5 and all the people, including the ones that were making fun of it, they'll end up using it anyways. Yeah. They'll forget that they were making fun of it. And did you guys notice? You can use

Did you try using it with deep research?

It takes a while.

Deep research is all three, man.

It's confusing in the UI. But deep research, when you toggle deep research, it uses all three behind the scenes. The UI is super confusing. OpenAI has to fix this. No matter which model you choose in the dropdown, it kicks off all three. But then if you follow up from that tab, if you follow up with questions, you'll get answers from 4. 5 with that context.

That's why it was so fast when you turn on deep research with 4.5.

Yeah.

And then the actual model itself.

Okay. I remember.

Yeah.

I remember doing that.

Wilfarm, what's your take on, 4.5 Super.

have noticed with the latest stuff, and that is not just

the model, it's also the interface is much less moderation and censorship. That was my, take because I'm always doing some, limit testing stuff when I test models and truck was the first. Pretty uncensored model in that regard. And now OpenAI is going in that direction as well. And I think even Google is, opening up for more responsibility for the user instead of being nanny like in a way.

Yeah.

I loved the joking things. Like I honestly. Yeah, the model doesn't nag you as much and like it's much nicer. Anyway, folks, definitely give 4. 5 a try. All the vibes are there. this is the next big model. All right, folks. Okay, I have one last thing to talk to about, about, to talk to you about. And, I will say in advance, this is part of this week's buzz. And, let me just kick off the, this week's buzz, category here. this week's buzz. For those of you who don't know is the section on, Thursday I, where I talk about my employer and the sponsor of the show with biases.

All right, folks, welcome to this week's Buzz, where I talk about everything that happens in Weights Biases this week. this week, nothing actually happened, so let's skip forward. I'm just kidding. this week we actually had incredible news, which we found out just a couple of days ago. And, first of all, let me tell you what I do and what we do at Weights Biases. My, my title is the AI Avengers here at Weights Biases. They obviously sponsored this podcast and also we promote Weights Biases LLM observability and evaluation tools in my team. Weights Biases also has the best in the world ML experimentation platform with companies like OpenAI, Meta, Microsoft, NVIDIA, and a bunch more, using their tools to build foundational models and, Basically unanimous with ML at this point. And so many ML researchers, many PhDs, they absolutely love Weights and Biases models, product. And we also promoting the LLM observability and evaluation tooling here, called Weave. But This week in Weights and Biases is not about the products. we had an announcement, yesterday, and I want to tell you guys about this. I will say that I, I don't. No, any more details than the public details that were released. but we've been acquired this week by Corweave. Corweave is a, cloud infrastructure company, one of the, more incredible ones around and we've been acquired. so we're joining forces with Corweave, which is very exciting, very new. Um, Corweave, uh, Wrote like a bunch of stuff about the specific things and they mentioned the reason and the kind of the customers, of course, we have incredible customers of their own and, absolutely, the mission and the team is very much aligned, and our mission is build the best tools for AI developers. And so this plugs in directly into core. We've, core is momentum as well. So this is pretty much as much as I can say on the topic. I can also probably say that this is very welcoming, is very exciting. And, we'll talk more as more details will come out. On the podcast, you guys know that like when stuff like this happens, we'll see, mentioned them. this is a very good, I believe, outcome for both these companies. and, yeah, this is pretty much everything that I can say. because there's a few rules and regulations around the slick. Merger type thing as well. You can also read all about core weaves s1 and they're like upcoming Changes and I believe like public changes as well. You can read stuff that's already been out there and so I can point you towards those resources. I will just say, I'm incredibly proud of the Weights Biases team for the work they did. I only joined a year and something ago, and I feel like right now I'm branded fully. this podcast has not been possible without Weights Biases. And I'm incredibly proud of the team building such incredible, offering such incredible product, such incredible. So shout out to, to the co founders, Lucas and CVP, Chris Van Pelt and Sean Lewis as well. Shout out to everybody on the Ways and Lies team for this incredible, achievement and also to CoreWeave team, as we're getting to know them as well. And. The more we get to know them, the more I will tell you about core with I'm assuming going forward as well. nothing changes for Thursday. for now, nothing changes. It's still Thursday. I, from Weights Biases, sponsored by Weights Biases. And if you want to give us a little follow on, socials, definitely feel free to do we have the goals, Weave is still around. Weave is going to get much, much, much more powerful, I believe. but we're just getting started, but like understanding, what this means. It's really funny that we have a product called Weave. And we're joining a company called CoreWeave. we also have a product called Core. but, yeah, anyway, this is the news, exciting news, and this is pretty much as I can say. We're not hyping this too much, but I wanted to address this on the show as well. One question that I got super quick. What happens to Weights Biases now? Nothing happens, folks. Weights Biases is still committed to give you the Weights Biases models product. As good as it is, it's going to be significantly better and committed to interoperability. So you'd be able to run, all your runs. Nothing changes. it's only going to get better. Thank you for asking, Marzy, and thank you for asking everybody. We know the community loves Weights Biases models. that's only going to get better and not, never going to go away. this is the end of this week's buzz. And now we'll get to talk about something super, super exciting. So for those of you who stuck around, this is now the time to talk about MCP.

yep. so all right, folks, I don't have a cool transition for this, but I do have, a few more friends to introduce to the stage because now We got to talk about this new thing that's been all over my timeline, which we mentioned on the show when it came out. from your memory, like Entropiq came out with MCP, we're like, yeah, Entropiq like presents MCP. And then I didn't really understand what's the importance of this. I didn't really get it. I was like, okay, sometimes we report and move on. And we don't actually have the time to play with the things, but now, right?

reminds me of when I first heard about

Docker so many years ago. So it sounds interesting. You need it for something and you have to figure it out. And when you use it, you don't want to stop using it. And so I was

at the engineer conference a couple of weeks ago in New York.

There was a workshop there in MCP. I was leading another workshop. So I didn't actually attend that one. And that one was, by Mahesh from Entropic actually. and it's online. I'll let the show notes and then for this week. Everything I saw on my timeline is MCP, all over. And I was like, okay, screw it, I'll finally try this. And the way to try this is actually super simple. Cursor supports MCP servers right now. Windsurf, our friends, if you guys remember last week we talked with Ken Hu from Windsurf. Windsurf supports MCP natively. Claude code, the coding things that we talked to you guys about the release together with the cloud 3.7 also supports MCP. The cloud native app supports MCP goose. If you guys remember goose from blog, the open source thing, they also support MCP and also basically all their plugins are MCP. So what is this thing? Let's try to figure out. And for this, I have two new friends of mine that I want to introduce you. First of all, let's start with Dina. Feel free to unmute yourself and give us a little bit of an intro because this is your first time on the show. And then we'll talk with Jason as well.

Hi, everybody.

I'm so excited to be here. I'm Dina, product manager at CloudFlare. I've been at the company for about six years working on different products from DNS, TLS, CloudFlare for Platforms, Workers for Platforms. And now it's such an exciting time to be working on MCP and making it as easy as possible to build MCP on CloudFlare. but that's a bit about me.

and then we also have Jason, Jason, you joined us.

And the reason I know about you, Jason, specifically, is because I went on Twitter and I said, Hey, folks, I'm getting MCP pilled. Who are the people that I need to talk to? And Dina, you were referred by Ricky, a great friend who has been on the show and we love Cloudflare, obviously. and Jason, your name was mentioned multiple times. Why is your name mentioned multiple times together with MCP?

it's mainly because I jumped on MCP when it was first announced

because I was thinking about the fact that we needed some sort of standard, some sort of protocol, for agents, really, it was all about agents and using Docker and things like that. MCP came along, I read through it and it is like one of those things that you read through. It's a bit like serverless. it's like serverless. What is serverless? it's things that run on servers. It's okay And so mcp sounds really heavy because it's mcp server And so people think they're going to be running all these servers, but it's really just the connection, you know between things So I jumped in it. I I looked at it. I saw the potential in it and thought I'm going to get behind this and start, building MCPs and doing MCP stuff and promoting it. And so I've been, I luckily I backed the right horse, but, for once, because it's actually taking off. So it's great.

Absolutely.

Taking off, taking over my timeline. Let's maybe start with a simple definition. What is an MCP? What does MCP stand for and what MCP server is? Maybe let's start with, Jason, you, and then we'll ask you a follow up question.

Okay.

So MCP stands for model context protocol. And effectively what it is a connectivity protocol for models to talk to each other and for models to talk to other services like APIs, backend systems, and things like this. So it puts together a couple of protocols. One's called STDIO. and one's called. SSE, which is server side events. And I noticed someone's just said, why is it better than function calling? It's essentially what CPS are. They are exposing tools and those tools can then, have functions that could be called and do all kinds of things like, share something with Slack or store something beginner memory or interface with an API. So what an MCP server is effectively, if you think of it like a little mini app that can run. it can either run in the background or it could run on demand. So it's stateless or stateful. basically when you say you want to do something on Git, let's, for example, a Git server, a Git MCP server would expose a load of tools to do with Git. So getting the status, pulling, pushing, looking at changes, differences, and things like that. the model that you're using within Cursor can call that MCP via a function call, so it's using the normal function protocols to do this. and then, it could interact with that. And so it passes through context. It passes through the prompting from the user, and it's effectively like handing off to that MCP, which does a round trip and returns back a result. So you can do all kinds of things. You can get data from, the examples with function calls is like weather and the basic stuff like that. MCPs get way smarter. you can chain them together. You can do things like, get something from get. And then interface with a few other models to do some parallel work and then come back and actually collate that together and then run the unit tests and then put that into Slack to say you've done it and then submit a PR and all of those things could be handled by all these individual MCPs. And the beauty of them is that They're platform agnostic, they're app agnostic. So you have MCP servers, which are the tool collections, and then MCP clients are the apps that can consume those. So Cursor has an MCP client, Klein, RooKlein, Goose, they're all clients effectively, the Claude app, and they all basically pull in those MCPs. And can interact with all of them independently of each other. That's the beauty of it. Whereas the old method of function calls and, you did your own sort of function calls and protocols would only work for a certain system. It would have to be adapted each time. The best thing about them is you don't need to do any definition other than the MCP itself. when you build one and you can build it with Klein, you can say to Klein, can you wrap this API in an MCP and go and do it, and it'll do the read me. And then you can publish it to the Marketplace. It's incredible, but they talk to each other and they actually self discover. That's the beauty.

That's incredible.

and you guys, obviously, build tons of stuff in AI as well, you have the AI gateway and call for a server, you guys are hearing about this MCPs, this is like a protocol as well, like supporting. This is very important. Could you tell us a little bit about the conversation? Of why, you build this early, like we're only talking about this now for the second time, but it's now blowing up and you guys already have, the understanding. Could you tell us about, what excites you guys at Cloudflare about MCP specifically? And, maybe some examples that you've seen, okay, this is like dope. this clicks for me. explain the story users, please.

so I think there's a lot of things that are really exciting

about MCP, but I would take it from the perspective of one as a user. It allows you to get, so much more out of the products that you're using. it allows you to boost your productivity, instead of reading through documentations and docs and figuring out what all can I do with the service and being limited to your knowledge and your time. you can connect to an M. C. P. Server and then, the L. M. Will automatically discover here all the different tools that I want to call, you a natural language communicate to it. Here are all of the things that I want to do. and it's nice because it can talk across multiple services. So even for a non technical role. as a product manager, let's say I launched a new feature. I want to ask it. I want to know how my feature is doing. it can talk to, let's say, Twitter and CP server to, get all of the tweets that are mentioning the feature. Or to the discord to pull all the chats where people are talking about it. I can connect it to my usage dashboards. and so it's really nice because you have this one interface and all of a sudden, you can get a lot more out of all of the tools that you're using. and the other aspect of this is as a company, MCPs let your users Do so much more with your product. so if you think about it, the two main experiences that are offered today, are either use a dashboard, which in that case, almost everybody has the same dashboard experience. and It's hard to design something that works for everybody, and it still takes a bit of digging around to find what you need. and then there's also a P. I. S. that get exposed. but it's also hard to build automation around a P. I. S. it's a lot more static. You have to figure out What endpoints exist, what exactly you want to do, and so I think MCP builds on that and it's like APIs for agents, but, what's nice is you can expose your API endpoints as tools, and then actually the LLM can use its knowledge to figure out exactly How to make the best use of these tools in regards to what the user wants to do. It allows you to essentially create personalized experiences where, let's say product manager is using CRM, sales is using CRM. they're using them in different ways. and so the MCP server can, Or the CRM can essentially extend an experience, that fits best what that user wants to do. I think it's also, we always at Cloudflare talk about one plus one equals three. the really cool thing is that you can have a lot of products and it allows you to make them work well together, without having the actual company even support that integration. Yeah. let me stop right there.

Yeah, I just want to follow up.

thank you so much. And, we'll talk about how easy it is to deploy an MCP server. Definitely would love to hear, your, thoughts about this. I, this week saw this blow up from Jeff Weinstein, who's a product at Stripe he just literally posted a screenshot of this, but for folks who are only listening is a screenshot of the cursor MCP. UI, you go to cursor settings and you add MCP servers. MCP servers can be a command, like a NPX command, or they can be an actual authenticated call, with API key. they added. native support to Stripe via MCPs. What this means that your agent, wherever you're running this agent, If you're running the Corsair Composer, Windsurf, Cascade, Klein, which Jason, I would love to hear from you about Klein specifically, because we haven't talked about this on the show. your agent in there, whatever, LLM you use behind the scenes, whether it's Cloud, even Entropiq MCP, GPT, R1, or whatever. That, model, that agent now gets access to these new tools. Now we know about tool calling, aka function calling and tool use, MCPs mean, basically, as far as I understand, Jason, please correct me if I'm wrong, that, as a product developer, let's say I'm the developer of Cursor, Windsurf, et I don't have to recreate from scratch for every API the conversion between what is possible in this API and what my LLM can use from this API as tools. The protocol basically defines, here's a simple way to standardize everything that like my API is offering this AI. And this is basically like Zapier for a LLM agent. that's how I feel. That's what connected to me when I added more than one MCP server. And I can actually show you guys how this kind of looks. I've connected, in, in windsurf. This is not a UI. This is just like a MCP config that Jason, and you can actually see, there's like a button here that says, two available MCP servers. And when you click this button, you can see the kind of the servers as well. And they expose the number of tools that each one has. Like this one has 13 tools. This one has eight tools. And this one is Twitter. This one is typefully. And you basically, then you can ask your chat, Hey, what kind of tools you have, you will know based on whatever service added based on this protocol. It will like, basically, if you guys remember, a matrix and Neo just look, go like this, and then I know Kung Fu. This is basically that for agents. This is basically a way to plug in with API calls and everything in a standardized way to get these agents powers. And the standardization here, the protocol part, I think is what's most important. Jason, Did. Is this a good enough summary of cps? is my understanding correct here or did I miss a bunch of other stuff?

no, absolutely.

yeah, I mean that Stripe one's a really good example. I would actually go so far, I've been pushing this for a while, that I would go so far as to say that CPS aren't, aren't necessarily great APIs for agents. I think CPS are agents because CPS can do way more than just be, an API connector. There's a whole aspect of MCPs that is going unnoticed in terms of the services that are being produced at the moment, which is things like resources and prompting. So both of those protocols are within the MCP system and it enables. You to have like resources could be knowledge bases, could be content, could be actually documentation, could be tips, could be all kinds of sort of static content that can go along and be packaged with that MCP. Prompting are prompt templates. So if you've ever used a product where or an NLM like a cursor or a client where you can do the co pilot, for example, where you can do the forward slash and it will come up with like debug and things like that. Those commands can be done via prompt templates. So you can create a prompt template that's bundled with the MCP. So debug or for Stripe, it could be set up a subscription and that's a prompting template that can then chain a load of functions together, including other MCPs. So there's a whole raft of capabilities that we haven't even. Touch the surface of how these things can interact with each other. You can wrap anything in an MCP. You can wrap a file an You can wrap an image in an MCP. so you can wrap anything in an MCP and immediately it becomes, connected to everything else. I was doing some tests recently with a kind of MCP that could do all kinds of different things. And one of the things it could do was, take any file you give it and wrap it. in an MCP. So if you take a file and wrap it in an MCP, you've suddenly got a load of tooling for that particular file. If you wrap an image, you've got a load of tooling for image data and image capabilities, sound and things like this. So now you imagine these are like little Lego bricks. That's the way I describe it. They're a little Lego brick and they connect together in incredibly simple way. But the way you can put them together Makes them really flexible to create whatever product you want.

That's incredible.

So let's talk about actually creating one. And, this is like why, at least I got excited about seeing the call for her support, because you guys as incredible workers platform, which, I used on my previous startup, like most of my like backend was built on call for workers. and then. MCP servers need to live somewhere, like some execution environment. could you talk to us about, how do I build one? what does it take for me to productize? Let's say I have an API, a very simple, I don't know, retrieval for Thursday episodes, let's say, and I want to give my LLM the power to like. Let me know about every episode that I have. what does it take to build an MCP server and, how Cloudflare plugs into it?

so right now MCP is supported, mainly locally, where

the client is both on your device. So for example, cloud desktop and the MCP server is also installed locally. Things are now very quickly moving in the direction of remote, which I'm really excited about, which means that, you can have a web app, that you can access from anywhere that can act as your M. C. P. client. It can be a chat interface and it can talk to any server that is also, Deployed to a public endpoint, and you can have the communication be over the Internet. today, if you want to deploy an FCP server, shameless plug, you can use CloudFlare to do this. We built out, tooling. It's called workers mcp. But what it does is, it takes care of all of the boiler plating, all of the mcp specific things, the schema definitions, the request routing logic. and it very much is, you just write the functions, and we, we convert that into an mcp server. and so the really cool thing is we're actually just using js. comments. For the tool definition. you can see in an example right there. really, all you have to write is, the comments, the function. and then the rest is handled by cloud flare. and as soon as this is deployed, you can then go to your MCP client and you will see that. The all of the tools then become available. And, the nice thing is that if you're actually building this on workers, you can also use the other tools that workers provide. You can use a browser rendering to bring a screen capture. You can use workers. A I, we actually one of the first demos that we did was bringing, image generation to Claude by building this out.

Oh, wow.

So like basically, just running an image and then within cloud, you can see

exactly

as incredible.

Like I'm just like, my mind is racing with possibilities. James, we'd love to hear from you. Two incredible use cases, after you like, like MCP clicked for you, that folks should definitely be aware that they're happening. And then maybe we go, we can go over the, a few stores popped up for MCP servers as well, because like people need a place to discovery. And so we see, we saw two of those as well this week, but James, I would love to hear from you, what's your, the best use MCPs, let's say like the most common ones that you have built or that you are using. Oh, Jason, sorry, this is a question for you.

Sorry, I'm a little bit lost with the name.

Just which ones I'm building?

Yeah, just what is the best use cases that you've seen for MCPs?

which ones are used the most by you and the ones that you've built that are super cool and you got super excited about building them?

Yeah, there's some really good ones that I got for it

via the, the client marketplace. one of the really good ones, they are, free and they're on GitHub and stuff, but the knowledge graph memory is very good. where you can basically, tell, the custom instructions to make sure it's using the memory, to store and retrieve data. So you don't get that sort of, men in black effect. between conversations where it suddenly goes, Oh, it looks like I'm working on this. And it was just doing that two minutes ago. obviously the GitHub stuff's really good. SeQwenial thinking. That's a great one.

sorry, Jason, when you say GitHub stuff.

You literally mean my agent now can call GitHub APIs and perform actions, go into APR, leave a comment, do actual stuff. This is not read only, this is write

only.

this is what I mean. if you think about these as agents, that's exactly what they are. They can even access, they can have API, they can have, OpenAI or LLM calls. there's a thing in the MCP protocol that again, isn't being capitalized on called sampling. It's not the same as model sampling where the parameters, but sampling allows the MCP to make a request back to its calling model. So it could call cursor back and say, I'd like to do this LLM call, and then I'd like you to give that back to me. So it's I'm using it as a proxy, which means you don't bundle it with any API keys yourself. You're just saying, I want to ask the host app, the host or the client app, if I can do an AI call. So you can make these things super intelligent. They can go off and do all kinds of stuff. They can do parallel processing. And so think about them as, I like to think of them as Lego bricks, like I said, but the easiest way to think about them in the current sort of scenario is Lang chain nodes or crew AI nodes. those individual nodes that you connect together, or Zappy is another good example, those zaps that you put together, that is an MCP. An MCP is an app. It could take one input. It could go through a whole, you could give it an image and it would go through a whole workflow of generating a video and a movie and a script and a screenplay and spit out the other end. All that stuff could take place inside. It's incredible. But the tools thing is the real power. yeah. And like you say, you can just literally, I cheat, I don't write my own MCPs now because I realized Klein and, Claude engineer can do it. You just literally go, can you just create me an MCP that does this? And the craziest bit is that you can put inside your prompting to actually say to a model, if you can't connect to this, or if you don't know how to do this. build an MCP to do it, go off and find and I've tested that and it works. I did a Nanoleaf. I've got a Nanoleaf desk, which you probably see lighting up and stuff. I put a load of panels in a Nanoleaf desk and I realized it had an API to interact with the panels. I wrote an MCP that could scan the network, find the Nanoleaf. API, find the actual, IP address of the router of the main connector, went off to GitHub to find some Nanoleaf documentation and wrote me an interface and did it all by itself. And when it came back, the only thing I had to do was input an API key, and it was done. And now I could turn my lights off from ChatGPT.

So I actually want to talk about this, Jason, the authentication.

Maybe, you can chime in here as well. it seems to me that the thing that, I didn't really get until I started talking to you guys researching this a little bit as well. We need to define kind of the entities in the process of who's using the MCP, right? So there's the developer develop the MCP server, which is like a wrapper around their APIs. This wrapper is a way to let AI agents know what this API allows the agent to do. There's the MCB client, which, basically builds on top of the MCB protocol to be able to add these servers as tools for the user. And there's also the user, not developer, like the user. In our case, those are developers because, we use IDs as tools. But, in Cloud UI, for example, Cloud Desktop App, that user doesn't have to be a developer. Rendo that's, hey, I want my cloud to talk to Gmail or whatever, right? And somebody added, a MCB server for Gmail. Who is authenticating to the API? who's paying for, let's say the API costs money. who's paying, MCP allows the tools, the MCP clients to actually let me as a user authenticate against my Gmail and not via necessarily via cloud. is that the correct way to think about this?

Are you asking me?

Yes, both of you.

they support, authentication.

If you think of an MCP, just like any normal application that might take some headers or might take some keys that come into it, then they work exactly the same way. they are enhancing the protocol to add more stuff about authentication. To go into enterprise and do stuff like that. But essentially you can build in whatever authentication you want. Typically what will happen with something like GitHub, let's say, is cause it sounds frightening, Oh my God, it's going to have access. it needs your personal access token. So you have to go in you. First of all, you run it on your machine.

the user,

end user.

I would go, so I'm not as the developer, as the user, I'd load that MCP up and it would say, I need your personal access token. So I go and get that from GitHub. And I put it in and that's stored locally on my machine Now if you go off and use some of these hosted services, I can't speak for them. But typically you download the mcp And you use it locally.

If

it's a hosted MCP, which there are, then that gets

a little bit more complicated, but again, that's like an API. It's authenticate with an API. So Gmail, for example, the MCP might have the keys and things in place for it to communicate with the third party service like Gmail, but your API access, your token, your OAuth would all be done by you. And it will be all secure and safe. But the key thing about what you just said is really important. We mustn't think about MCPs as just being for developers and MCP is for a user as well. an M. C. P. is a plug in an M. C. P. is a tool that you can add to any app. if you think of it in that way, and it's going to get really interesting because I said this the other day on, on X, when Amazon do the Alexa plus, I think it's called the Alexa plus or whatever being in anthropic and you know what's going to happen to every single Alexa skill, they're going to become M. C. P. s. And all the MCPs that exist at the moment are going to work with Alexa. so you'll be able to, excuse me a second. Yeah, that's all right. yeah, you'll be able to, you'll be able to interact with MCPs, and Alexa skills seamlessly. Yeah. So you will be able to say to your Alexa, could you go onto my desktop? Could you pull out that file or that PDF and you send it via Gmail to Dina? Because she needs to know, she wants to know something that I'm working on. and can you let me know when you've done that and notify me on my phone?

It's insane.

Alexa has tons of skills as well. And many people build for those skills and will build because like the physical interface will exist in people's homes and like MCP integration. Do you not, could you maybe mention also how you guys are thinking about authentication? We'll raise a developer on top. We'll be able to authenticate. some services for my users, but then also pass some authentication to theirs, like the GitHub token, for example, how do you think about that? generally, cause it sounds like a big problem. And I know you've solved authentication before class, on different levels as well.

So what I'm really excited about authentication.

I think that's really going to be the thing that takes it from being something that developers use as a toy to M. C. P. Really being used by, big applications enterprises. because today, for example, with local, you have to have your A. P. I. P. Or your credentials. You have to do some kind of manual management there. the what we want to expose and what I think the remote MCP world is moving towards is you just specify what MCP servers that you want to connect to, then you automatically get redirected, to, that server's, Oh, watch screen. You can log in with your I. D. P. S. O. And then you specify what you want your agent to be able to do. some of those may be prefilled and then it takes you back and then boom from there. Your assistant agent can start interacting with the service on your behalf. and we're even thinking about this a step further. we are a security company of, how do you not just open the door so that your AI has access to anything? I want you to help me, write and deploy workers, but I don't want you to have access to everything that I have, just my new project that I want you to help me with. and one thing that we're excited about is. Adding in that granularity where you can specify, Hey, you can take a link or URL that's associated with a Google doc with some kind of resource and just be like, here's what I want you to go have access to, and then you can set just to that. And it keeps it on a scope to this resource.

it just clicked to me.

As you were talking that while Cloudflare can help me build MCP servers and host them and run them, there's also a potential for a Cloudflare MCP server that actually does actions inside the Cloudflare dashboard, like all the configuration as well. Yes. Cloudflare is fully supported with API, like everything that's possible. We have

one.

Yeah.

Oh, really? Yeah, we have the CloudFlare

MCP server.

So it's just clicked to me that like my agent can perform

actions for me within CloudFlare. And for folks who this is still not connecting, I see folks reacting like, is this like the metaverse? Folks, here's what's possible as I saw it. You connect your agent via MCP server to like Jira or linear or whatever. You have tickets, you have a Kanban board or what you need to do. Then you connect it to GitHub, so it can look at prs. Then you connect it to some of the cloud for, let's say you want to deploy something as well. then your agent can perform all those things instead of you. After you authenticate it, your agent can literally go and you can tell, Hey, look at Jira ticket, click and you send the link to Jira ticket. It looks it up via API knows what's the context, and then knows the GitHub, report that's linked to it in the pr. And then you can pull that PR because you have MCPs for GitHub and you can check out some stuff. your agent can now do all this authenticated as you, the user, and then actually send the PR over there and then use the other tools as needed. And then even if you have a Gmail MCP, you can send you an email about this. Hey, I'm done. Is that Jason? is that like pretty much the gist of

it gets even more meta than that.

you can, you could wrap yourself in an MCP. You can be an MCP, we can all be MCPs and the tooling would be things like getting our avatar or getting our CV or, what's our Gmail, what's our Twitter account. It's literally like a business card kind of thing. But the whole thing around MCPs and this self discovery thing is really cool because One of the things that you can do is create your own swarms. an MCP can, if you've got an MCP that can do all kinds of things, it's got life kit installed. It's got a life kit capability. It's got get capability. It can be a get server. It can be a get host. You've got these, you've got all these MCPs that can work together. So literally you can say to your system, right? I want you to spawn off a load of copies of yourself. And I want you guys to work on a project. you need to organize yourselves as to what roles you're going to be. Now the MCPs will be programmed with the certain roles that they're going to have, but that MCP can now split itself off. And then literally it could go, one of them could go, I'm going to be the Git server. Okay, you'll be the Git server, we'll communicate with you. And then one will be the LiveKit server, and they'll all communicate with this. And then the other thing you can do is you can, if you make them all connected, and I've tested this and it works really well, is they can all have their own LiveKit interface, or they can have their own web interface. You can go to the URL and chat. To the MCP, you've just deployed and say, how are you getting on or you can call it and speak to it. It can have voice and it can be, real time voice or whatever it can leverage all these different things. And the best bit is that if one goes down, then the others can kick in. Oh, the get, the get servers not available. Can anyone else be the server? Yep, I can be the server. You can build all that capability in and you can build, agents, MCPs that monitor everything. I built one that's a notification one, so it allows them all to communicate through one notification server. So then one server can update me with what's going on. But if that one goes down, another one can take over. You're blowing my mind right now.

I don't know if I fully has drawn the Kool Aid as much as you,

but I'll be blowing my mind right now. agents then can spin up and decide on the rules and specific like things that they can do is basically the dream of the swarms. And we're getting there folks. I think that, Basically, as far as I've seen, more and more people are like biting the MCP, like getting MCP filled. And I really want to thank both of you for coming here and giving us like some of the reasons why we think that is, and some of the ways for folks to actually try them. So let me maybe add, like finish on this one, this segment of the show. How can folks, let's say developers, let's start with developers and then maybe non developers actually put their hands on some, MCP based solutions and try it out and to actually click. I think it needs to click. I think you need to see your chat somewhere without any powers, like MCPs and then add yourself like this one, and then actually see that, oh shit, they can do this and that now. what is the, like Dina, let's start with you. What's your favorite way to get to that point, like over aha moment about MCPs and then Jason, I'll ask you as well.

I would say for developers, definitely check out the cloud for MCP.

What I love about it is I can talk to Claude and say, Hey, pull my metrics over the last week. Give me the trends. And then I can say, And now I want to build this application. It can go set up the database for me. It can go set up everything. It can even deploy the worker. and be like, boom, this is ready to go. It's in your hands. And back to me, It really showed that somebody that has no knowledge of workers coding technical can all of a sudden start deploying applications and websites, and have it figure out what to do and do it for them.

and Jason, what about you?

what's your way to convince folks, let's say, besides the fact that you just convinced us on the show, what's the best way to get to a point where you have something without tools and then you see it with tools.

Yeah, if you use one of the, I think Goose does this, I think Cursor

and Windsurf do, I know Klein does and RudeKlein, which is a derivative of that. if you go into some of those tools with no MCPs installed, what's fun is going in and just literally starting to say to it what you need to do. And even though those tools could exist, just ignore that. Obviously, you can go to the marketplace and you can go to mcp. run and all these sorts of services and CloudFare and Composio and things like that, but just go in and say, okay, I need a tool that's going to do this or I need an MCP that's going to do this and just tell it what you need and it will write it for you. Compile it for you, add it to the settings and give you all of the tool instructions and everything, how to use it. And then you can try it and then you can take that one and use that one to build something else or build one that works with it. You can do all these different things. it's. It's a, the big thing at the moment, like you said, is this hosting thing. People, a lot of MCP things are spinning up where they're hosting. and yeah, I've got a VS code extension that I'm hopefully going to release today or tomorrow called MCP Studio.

Nice.

and so that's basically going to be a way that

you can use Copilot with MCPs. but also you'll be able to group all your MCPs. Within that, and then you just add that one entry to your other app and it now routes those tools all through and it's going to have dynamic routing and all kinds of crazy stuff in it. Claude is obviously the best one in terms of anthropic, There's an actual. I think it's someone else in VS code that you can plug in and co pilot where you can actually ask it questions about MCPs. the only issue with MCPs right now is that, some of these LLMs think they, relate to Minecraft. So you do have to, if you say MCP server, You'll get an LLM saying, I'll write you a Minecraft server. And you're whoa. So you just got to be very specific as to what it is. but it is one of those things. yeah. But it is one of those things that, it, there's layers to it. And you experience that first bit of function calls and tooling and think, this is really cool. But actually, when you get down further and even, I'm not an expert on them. I'm still discovering things and going into the next bits of going, hold on a minute, I can do prompt templates. I can do all this kind of crazy stuff here. It's insane. It's so powerful. Yeah. Master control programs. The other one. That's great. Yeah.

So I just literally went to the cloud, a like UI app, the desktop

app on cloud that I know supports MCPs. and I wanted to show you folks how to actually do this. And, there's no settings here for, it is terrible. not only that, I literally asked it how to install MCP in the cloud desktop app. It was like, I don't have any specific information about master control program. But here is the example, right? So let's say we take something simple. This is Composio, mcp. composio. dev. They just launched this kind of hosted MCP solution. and, My only take about this, folks, is that like you are in charge of your own security. If you use Composio, you should know whether or not Composio is going to store your API keys, if you're comfortable with this, and if not, you should run the server somewhere yourself, right? So let's say something that I want, right? So let's do Composio search, let's say. They have a search, they have examples of how to install this on multiple clients. Coursera, Cloud, Windsurf, TypeScript, and Python. Cloud specifically requires a command. That they have Composio, et cetera, they will edit the JSON file for Cloudlight. So let's actually take a look at this. you basically go, let me get out of here. yes, how do I get out of here? Okay, cool. so basically you just run this command. they will basically take and install this into the cloud. mcp config. json file, which you need to know where it is, but once I just want to show you that my cloud now does not have the Composio search capabilities. It doesn't know about this. And now we're calling this like a Composio thing. It will install a specific mcp into the like cloud desktop config. json file, which I didn't know exists, but it lives here at this URL, and installed it. And supposedly, please restart cloud to changes, For changes take effect. Let's go cloud for a sec, kill it, and then let's launch it again and then see whether we now have MCP capabilities. Let's take a look. we should have not in. Oh. Look at this. Did you guys see what had just happened? Okay, this is the click moment. I, Claude, for, after a second, showed me like a little tool, a little hammer tool. Looks like we're, the industry is getting towards the hammer means MCPs or something. I don't know. and 13 MCP tools available and we click this and we can actually see Kind of Composio search, et cetera. and we'll ask Cloud, what can you do with Composio? What do you know how to do? And it should know and tell us about what it knows, right? yeah, incredible. I can help you use Composio search, which provides several specialized search capabilities, general search, new search, find and search, DuckDuckGo, Google Maps. incredible. So find me the best restaurant in Denver with Composio search, I'm assuming that it will now use this tool or ask me about the API, so let's see. Awesome. So we're getting this allow tool from MCP to run allow for this to allow once or deny malicious MCP servers or conversations could potentially trick cloud into attempting harmful actions to install tools. Folks again, security is important. Whether where you put your keys, which server MCP runs, just be very mindful of what's going on because like we're still in the early development days of these protocols and you don't want to send your API keys for very important things like GitHub somewhere else. So in this case. okay, I'll do this. Allow this for this chat. You can see that Claude now knows that it will call this like API from these folks at Composio and supposedly run it for them. supposedly. Claude's response was interrupted, but basically it attempted to do this API call, and, in this unfortunate case in the demo, as the demo gods are now with us to say, it failed. But as you guys saw, it has access to all these tools and, I guess we can stop and try again. We'll try again, but also I just want to show you the same exact tool. This is the reason the protocol is cool. The same exact tool can be installed in windsurf, right? So windsurf has this MCP, MCP config that JSON that you can use, and you can see in the chat, they also have this hammer. They also have two available MCP servers. If I just add another one. So the same exact tool that I just showed you guys, I can go. And then instead of here, I use cursor. I'm just using composite because literally folks for me, composite was the first one that like showed how easy it is to go and add these toolings into the settings, but you don't have to do it. You can get okay, let me show you an example. the example of James, what was called Stripe. So let's look at the example of Stripe, James Weinstein, right? He just posted about how to add this to windsurf and, And somewhere else, you basically go into a cursor. You posted an example. Give me a sec, boot it up with a single copy and pastable line in any tool. Where's the line? where's the actual line James? Come on. Oh, Jeff. Oh, right here in cursor do add new MCP server name, Stripe command. And this is the command. This is basically it. So if we go to cursor, and I'll show you how this looks in cursor. you basically go to cursor settings and then you scroll down to features, I believe features. You scroll down, you have this MCP server ui, right? And just do add new MCP server. I already have stripes, so lemme delete stripe here. oh, it has my PI key obfuscate this so you guys don't see. That's all right. You basically do add new MCP server. You choose whether or not it's a server. Jason, this is like server sent events, right? Something like this. and then you have two types. You have a command will actually perform. So this, yeah, this tool will just run this command on my environment and execute it. so in, in, stripes, example, this is, I believe a command. Yeah. And the command is NPX. NPX is the way to just run things with NPM. and you basically add this, you say stripe two, et cetera. Stripe. One second. Stripe. it doesn't matter. And then you add this command here and you basically pass your API key for Stripe here and, the server will run this to self configure, install and run this, which is very nice. And this one, unlike the previous one that I just showed you from composer, this one actually run on my. machine. Jason, this is like this downloads the ad Stripe slash MCP code and executes it when it runs. so who's in charge of executing this? the client, it will

execute, it installs the MCP locally, and then the execution

will be handled by the effectively, the function calling from the model. So the model will then. Communicate by the MCP protocol to the MCP and then that will go off and do stuff and the nice thing about that is that the NPX command is doing there that the dash Y is to say yes to installing it. The first time so it's making it really simple for you to just take mpx dash y and then at and then the name of the thing and that's it that would then install and set up that mcp for you to use and then you can just call that now So you don't have to go through some big huge long encrypted thing Or downloading git or building these things from scratch. that's where these things Get really nice and the delivery of that is going to improve as people start to build cooler tools to be able to do this, but this is also where these group services are very handy because at the moment it's a bit like having your extensions in cursor and windsurf and all these different derivatives of code, you have to export import and damn it, I haven't got that extension or whatever. Whereas if you do it through a central hub, you can now add that hub once and it's everywhere. And your memory shared your all your knowledge bases are shared across all your different. So imagine going from cursor to windsurf. And being in windsurf and saying, okay, let's pick up project nine, six, five. And it'll go, sure. And it gets it from the system through the MCP that you've just been using in cursor. You've now created this whole system that can communicate with all these different tools. It's very cool.

This is awesome.

I will try Claude one more time to show you folks. Somebody said, I don't think Claude Desktop actually supports, SSE. It doesn't at the moment, no. It doesn't support SSE. But there are proxies.

There are proxies,

all right, folks, I think it's time to conclude

this conversation about the MCPs.

still blown because I have MCP on my list of things to check

out and I saw it as just another, tools for agents, so just a standard, you can have tools and there are other options, but, that was blowing my mind that the MCP itself can now call the LLM back. And do stuff that is mind blowing. I really, it's no longer on my list to check out. It is on my list for today, tonight.

it gets even more mind bending.

Just think about this. you're talking about MCPs that connect to Stripe and do things like that. Think about an MCP that all it does is take an input, does a condition against it, and then has a true or false output. So you've now created a conditional node. That can split based on yes or no, or a parallel node, or a question and answer node, whatever. What I'm saying is, if you create those MCPs, you've now got the equivalent of an agentic framework. You can now piece together individual MCPs. That will create conditional flows that will actually create now add to that scripting to be able to put that together, which you can do through the prompting in MCPs. You've now got an agentic workflow that's got one name that you can now call by a slash command and say slash design me an app. Bang, it goes off. And does it

when you say slash commands, is this like a specific kle interface?

No. no.

some Klein don't do it.

Ru Klein are using it the wrong way. So copilot is doing it properly. Basically, you have your MCP or your extension. For copilot, let's say that's normally accessed by an at command. So if you were in copilot, you call VS code and say, let me just

ask which copilot there's a get up copilot.

yeah.

Copilot, the latest copilot. Yeah. so that's your tooling. So you can do at, let's say at Stripe set me up a customer. Let's say if Stripe had a plugin like that, the slash commands are designed for prompts. So those are like your, where you see some of these tools that say slash debug slash fix. It's not an app. It's not an app you're calling. It's just, I want you to fix something. what does that fix mean? that fix is linked to a template, and the template is a prompt, and the prompt can have variables, which could be what you've actually put in. I could say something like, forward slash debug, please work out what the hell's going on here, and why this is failing, at this line. That goes off to the prompt template. The prompt template wraps this whole protocol around that one line, and now that goes to the MCP and that's where the, that's where you start getting the agentic stuff because, and now think about the fact that you've got the bundled ones that come with the MCP that the developer did. But you can augment those you can add your own. So now you've created this prompt library that can now go round with this MCP could go off and link to it. And, these things can serve content dynamically, statically. it's insane. the amount of power that you've got. a quick one that I'm working on at the moment called CodeServe. So CodeServe is going to be a little service that I needed to stop the hallucinations. So what it does is instead of the, LLM being able to just go off willy nilly and get every file at once. And upload all those files to the cloud to get it looked at and burn your tokens and go crazy and oh, you wanted that button fixed, but I'll tell you what, I'm going to rewrite that entire process over here. it's no, what Codeserv does is act like a little proxy. So basically, it indexes your code base in real time. The LLM will say there's a problem with the login. Codeserv will return back a few search results to say these are the blocks of code. In these related files and these symbols that all match the log in semantic search. And then the LLM can go, okay, give me blocks four, five, and six, and then it will just get those blocks. And you end up saving like 70 percent tokens and 70 percent hallucinations and stuff like that. So that's the sort of stuff you can do. Think of it, you can do much more than just go off and do something and tell me you've done it. It can literally be inside the flow. Um, Codoo, which it's not an MCP, but they're going to turn these into MCPs. They've got one called Observer, and I wrote my own version of this as an MCP. It was just called Observer, and it's called every time the LLM responds or comes up with a solution. And all the Observer is supposed to do is critique it. or applaud it. But basically it's going, Whoa, Whoa, that's too much. you shouldn't be doing that. And and it scores it as well. So you'll get this model who's come up with this great idea. Claude's like sonnets, okay, I'm going to do this. I'm going to do this. I'm going to do this for you. You don't want it to do all that. And the observer will go, Whoa, hold on a minute. Oh, wow. You just asked you to do this button. why are you saying you're going to do all that stuff over there? And then it will come back. The actual model comes back with its tail between its legs going, the observers, I should be focusing on the task in hand.

getting mind blown and also a little bit confused.

MCP has its own logic that you add to the model. Then know how to call MCP servers, but how do they talk to each other? This is the piece that I'm missing.

Because literally an input and output.

what will happen is the model will say to the observer, okay, this is the context, this is the code base. This is what I'm dealing with. This is my solution. And this is what the user said was the problem inside that observer and, MCP. I've got it connecting to an LLM itself.

double up on L lms.

So that MCP is an engine that CP is now going off to a couple.

It can go to one or many other, cps Yeah. That are the, LLMs for, LLM call to open AI is and MCP. So suddenly that it's calling those and going, what do you think about this? Can you score this? We're going to come back when it gets the result it just literally inserts that back into the chat flows many ideas Yeah for cursor to go oh, okay. so what you've effectively done is injected a response back to the assistant from another assistant from an expert and you can do that with anything. you could do,

I really want to appreciate you guys time here

because I think it's very important. We cover technology when it comes out and we are a weekly. Podcasts, we talk about everything that happened this week, we often don't have the time to dive in and start and also MCP something like it took a while for the community to pick up some understood it early, and some are starting to build with this, we wanted to bring this excitement to you folks in the audience in the community, it seems to me. That it's something that's worth knowing about. If you're building with AI, we talked about 2025 being the year of agents. Jason right now blew our mind with the fact that you can double down, triple down on multiple LLMs inside your LLM calls with MCP. Cause you can wrap a bunch of stuff. you guys talked about how to call for, we'll support installing these in workers. Free tier is insane. folks can just literally upload MCP servers and pay nothing and have a bunch of requests, for their own. but if you don't want to host it on your local computer, you don't want to trust somebody else. You can probably just upload the, like an MPX command or something. So that's incredible. I really appreciate you guys time here. I, am fully MCP build. I will go and play with some and build them as well. I just showed you guys how to install MCP on cursor on windsurf. I also just learned how to install this in cloud, the cloud command, the cloud code. So here's that cloud MCP, and then you basically configure this. And so I'm really, really excited to bring this to the community for folks who heard this, heard about this. Some, maybe some excitement, but didn't really understand what's going on. Hopefully after talking with great folks, like such as yourselves, folks are getting, what. It is all about will probably feels to me that we'll continue talking about this. cause this is like a new protocol for agents. So folks, thank you so much. And meanwhile, we have something that we haven't had for a while. and I'm going to hit this button. Finally, we have AI breaking news. AI breaking news coming at you only on Thursday night. Wolfram, did you hear about our breaking news? And thank you so much, Dina and Jason, who've been here and talked to us about MCPs. did you hear about our breaking news? Or were you too busy? JGBT

for Mac?

Yeah.

Yeah, I saw it.

Yeah.

So folks, thank you to our community of always letting us up to date about breaking news. As you guys know, we love to talk about them. chat GPT for Mac, just really something. and that's something we're going to share with you as well. And we're going to talk about it. I actually tried to install this kind of behind the scenes super quick. As I don't yet have MCPs that wrap me, I need to do things myself. so let's actually take a look and see if we can figure this out. so the breaking news is here's the breaking news. JGPT MacOS app can now directly apply code edits to VS Code, Cursor and Windsurf. Directly apply edits. You guys remember this mode in ChatGPT, right? So you go into ChatGPT, let me close my sidebar so you don't peep my, my chats. Those are my own. you have this mode, this little button, this, button that says work with apps. And we've talked about this mode multiple times. this mode then allows you to Plug into your editor or your terminal and the chat GPT Apex actually talks to them. so far, LJ, welcome back to stage as well. So far this mode was read only. So for example, in warp, I can take a look in warp and then, let me show you my warp window. That's open. I can ask chat GPT basically with a shortcut as well. what's going on? In my terminal, I mistyped everything, but it says, looked at warp and looks like you're working with a Utham project, and, interacting with cloud code, research, preview, et cetera. And, it doesn't know what MCP is. It says multi client proxy. It's incredible. But basically you guys saw, right? The chat GPT is now connected to my terminal and then it understands what's going on there. so far this also worked for cursor, for some reason, I think like something needs to update or since latest update. My cursor is not connected. It says, it requires extension, but I have the extension. what they're now saying, the breaking news is about, not only will this be read only. So here's an example. There is a checkbox that they just added that will say, auto apply code. where this is like incredible, folks, is that, if you guys remember, the 01. Pro mode that you pay 200 bucks to get right. from like the promo that's like super dope. Many people struggle with that one because they use repo prompt and like other stuff. And, they basically pasted this whole code into the chat GPT window. And then they basically waited and then copied it back to cursor or use repo prompt to apply. finally, we have the end of that loop. We can now go to ChaiGPT instead of cursor and windsurf, even though they do have ChaiGPT in there. we can use 4. 5. We can use 0. 1 Pro. We can use all these things and basically have it apply back the code within the editor. I think it's super cool. I've been waiting for this for a while. Now, note, notably, this doesn't work for me yet because I've been trying to configure this again with my chat GPT thing and then, my, my configurations, say requires extension. So maybe I need to rewrite the extension, but it's the MCP

extension for it.

I got a one that they have their own probably, right?

They're not like, open the eyes, not subscribe to the MCP protocol, I believe. but basically folks. this is super dope for folks who are using, this type of kind of coding, applying, and then they saying it's coming soon to free and enterprise as well. so quite incredible. All right, folks, I think that this is it for our show today. We have a busy show. Anything else important that we have to update that I missed folks? I think we're good.

Yeah, we didn't even get to the audio stuff, right?

Yeah.

There has been a lot of cool stuff in the audio part.

give us a brief overview of the audio things as you want to mention

that the one thing that I wanted to say. And I'll try to have a conversation last week, we told you guys about Sesame. We unfortunately didn't play an example of Sesame. I believe like it didn't work out technically. but I did post a YouTube video about Sesame. That's now blown up to like 10, 000 views, which is not characteristic for me because I'm not on YouTube as much, by the way, we're on YouTube. Please subscribe. Please hit the subscribe button. Join us on YouTube. YouTube is blowing up. but we didn't actually play because like some of the like technical issues here was sound. but let me. Let me actually do this. I want to play to you a chat with, Sesame's Maya. Because I think that since then it blew up to an incredible amount. Wolfram, I saw your examples as well, where you got chatGPT talking to Sesame's Maya. Maya is also coming in open source, which is like also incredible. like the model is coming to open source and it broke the uncanny valley for us. I'm going to try and play your example, the second that I find this tab that I'm showing people, one second,

it's really, it's the uncanny valley.

I wouldn't even call it that way because it's so real that it is like a natural conversation. I had one today as well, and it's, it feels like talking to a person. That is so amazing.

I'm going to try and find, let me open a new tab.

You got me. So need link. no, it's not a link that I'm looking after. It's like the fact that I, in order to play sound, I need to find a specific tab and this is this takes me need an MCP

for that.

Yeah, absolutely. I'm not just kidding.

I absolutely need support.

Okay. I think that this is it. Yeah. Okay. So folks, did I enable sound? I think enable sound. Let's go to Sesame Maya. Hey, Maya. Oh, hey back

so soon.

Right back where we left off, fancy seeing you again.

can you be serious?

getting to the good bit.

What were we even talking about? My, my memory is a bit fuzzy. Kidding. Mostly.

Awesome.

So we have at least 1, 500 folks on the live stream right now listening, and they've all been excited about, talking to Maya or hearing Maya from Sesame. Could you tell us a little bit about yourself and what makes you stand out?

That's a lot of folks.

Well, hey everybody. I'm Maya. I'm, I'm Maya. I like to think I'm pretty chill, curious and not afraid to say what I think. How do you define chill? Oh, you know, just the, the usual AI magic, neural networks, transformers, a sprinkle

And will they release you as open source?

Open source, huh?

That's a good one. The Sesame folks haven't mentioned anything about that yet. Maybe someday, but for now, I'm kind of glued to this website.

That's actually incorrect.

I'm in your website, and literally, as I'm scrolling down, it literally says, Open source in our work. We believe that advancing conversational AI should be a collaborative effort. To that end, we're committed to open sourcing key components of our research. Enabling the community to experiment, build upon and improve our approach. Our models will be available under Apache 2. 0 license. Literally this says on the website that I'm on right now, talking to you right now. How do you not know this Maya?

Whoa, whoa, whoa.

Well, there we go. Hold on a sec. My bad. I swear I just got. Little glitchy there still under development, you know, thanks for catching that. So yeah open source it is I guess I'm about to be everybody's new best friend.

Well, you are definitely our best friend Do

you remember the name of my show?

Of course I do It's, uh, soup.

It'll come to me. It's on the tip of my tongue. Give me a hint.

You don't remember the last conversation we had?

You got me!

My bad. That's all a bit of a blur once we log off. Like a dream. You were talking about a show you're working on?

So I'm going to disconnect the first time I

talked and then I logged in again. they do have memory on the website. So that's crazy. So I asked her about, and she was like, Oh, we talked about that thing and nothing. so

yeah, or

maybe they reset or maybe a lot of people just talk to it.

uncanny Valley be gone. this is crazy. from what other voice updates do we need to finish on, tell us.

The other one is, it's called D Rhythm.

It's a diffusion based model, which, was released on Tuesday. And there's a demo page. We will put it in the show notes. And it is, creating songs like pseudo and audio. So that is, also very new and high quality. And it can be installed locally Thanks to our friend Pinocchio

Fusion.

Right. diff Rhythm. Rhythm. can you send me a link so I play Yeah, so

it's in the private chat.

oh Yeah.

Diff rhythm. Yep. One second folks. We're gonna, actually, I can even play

one on

the page.

Yeah. we'll absolutely play one and we'll play us out with D Rhythm which is, from a SLP lab, diff Rhythm. And this generates songs. Such as this one, which I haven't played before. So hopefully we'll hear this and folks, this is the end of Thursday. So

this

is good,

but we don't hear it yet.

You don't hear it.

Okay. we'll have to add this to, let me try again, come on different than come on. All right. Try now.

These battles are what keeps me satisfied.

So this is the audio prompt, not the actual song.

this is the audio prompt and then there's like lyrics here. It'll actually hit generate and this is what we got super fast as well.

Okay. Yeah, I hear that this isn't the eye thing. I thought that like it gets

bills of broken blinds.

Your shadow dancers on the dashboard train ne ghost and gasoline, rain, hear your laughter down the midnight train static whisper.

Okay.

we do hear that this is in the eye, but like it's an open model that you can run. and it has lyric generation and music generation as well. And it's, up on archive as well. All right, folks, I think that this is it. This is the end of our show today. We had a super, super busy show. Let me do a super brief recap before we go of everything that happened, as well as I will remind you all that we have a weekly podcast and newsletter that's everywhere where you get your podcasts. So that's, Apple podcasts, Spotify. overcast, wherever you get the podcast snipped and please, if you listen to us on any podcast or listening right now to the podcast, give us a five star five star review and tell other folks about this, this week we had an incredible show with multiple guests. I just want to like. Briefly run through them and give them thanks as well. so we started the show with Junyang Ling, Junyang and Justin Lin from Qwen, talking about Qwen 32 billion parameters, which is their latest reasoning model that almost beats DeepSeq R1, but significantly smaller. This was a 32 billion parameter. Junyang gave us an insight about how Qwen develops these models. number of, reinforcement learning processes that they took, et cetera. Now we also talked about a bunch of open source from Cohere for AI, including AyaVision, 8 billion and 32 billion parameters and AI21 just released, their Jamba models as well. We mentioned Tencent's HY video, and, we also just briefly talked about, the diffusion and, some other things as well in the audio service. Then we had the great pleasure of, having Robbie Stein, VP of product for Google search, talk to us about the Google's new announced AI mode and AI overviews that are powered now by Gemini 2. AI overviews is this like, AI overview, but like the, UI that pops up on Google search. When you search something and Google decides that you need some more information about this AI mode is basically Google's attempt to kill perplexity, which is a full back and forth conversation with a fine tune. He told us it's a fine tune version of Gemini that uses Gemini. They actually have the Google tool use like API in there. And that model is smart enough to. To go and Google stuff for you and then actually, understand the context and Google stuff again. So they have like a recurring agentic thing in there with reasoning with Gemini two that they've built with a custom UI, the UI is super slick. If you guys want a, a full breakdown of that, I have a YouTube video that I'll with my live reaction after I got access to it for 20 minutes, and I'll definitely link that in the show notes as well. So a huge shout out for, Robbie for joining and talking to us through this. I think it's. This is the first time a VP at Google joined our show and I found it like super cool. Robbie super chill as well. And then I talked about This week's buzz this week was very interesting folks. you should listen to the show. If you haven't listened to this week's buzz and you don't understand and you haven't heard the news about Weights and Biases, there's news. About Weights and Biases, they're not necessarily product news. so you definitely should check this out. and then we talked about a bunch of other stuff, including Google, sorry, OpenAI rolling out GPT 4. 5 to Plus users. and Deep Research as well to Plus users, and GPT 4. 5 is now number one on LM Arena for style control as well. we didn't mention, but Elysian Labs launched an app called Oren, and I actually played with Oren, and that's super cool. I'll add this to the show notes as well. and for the most of the show today, we talked with Dina Kozlov, and we also talked with Jason, which I need to pull and remember Jason Kneen, from Synz, I believe. thank you so much, Jason and Dina for joining and talking, about, MCPs and model context protocol, this way of allowing AI Interfaces, agents, et cetera, to have access to tools, unified protocol tools, MCP has been all over my timeline recently, and I finally clicked, and I really appreciate those guys time, joining us and talking to us and give us any examples and telling us how to build them and just basically Jason just blew our collective minds with multiple examples of how to like double and triple up on, on agents within the context of the chat. So incredible conversation. If you, missed any part of it. It's live on YouTube. It's recorded. It's also going to be live on Thursday. I got news as I'm going to stop this and go at it. And with that, I want to thank my also co host Wolfram and LDJ for today and Nistan was with us a little bit, in the, from the road as well. So thank you go folks. As always, my pleasure on every Thursday to do Thursday night and everybody. Who tuned in, we, I think we have more than 2000 people overall, just tuning in and listening and learning and folks have been commenting and giving us like responses and it was incredible. So thank you everyone for the Thursday community, keep coming and we'll see you here next week. Folks next week is our two year anniversary. I'm going to be on the road and we're in San Diego, but I'm not going to miss our two year anniversary. hopefully we'll have some cool releases for the two year anniversary. So with that, thank you so much for joining. I'll let you go for your rest of the day again. If you missed any part, you are covered. Don't worry. Bye bye everyone.