📆 ThursdAI - Nov 6, 2025 - Kimi’s 1T Thinking Model Shakes Up Open Source, Apple Bets $1B on Gemini for Siri, and Amazon vs. Perplexity!

Welcome everyone to Thursday I four, November

six, and my name is Alex Volko. I'm an AI evangelist with Weights, & Biases, and we are very happy to start the show with some breaking news right off the bat. But before we do, I wanna introduce the co-host for the show. So we have almost a full cast today. Welcome folks. We have Nisten, we have Yam, and we have Wolfram. And, yam is already happy. I think I know what Yam is happy, but we definitely should talk about why all of us are here.

30 AM Pacific time.

yeah, folks, welcome. how are things? I will just say, before I gonna ask you, what is the most important thing in the ai. That you have this week. I posted yesterday that this is the slowest week in AI in the past two and a half years that we've been doing the show. As I was preparing the notes for today, it was not a lot folks, it was not a lot for the first time in. I honestly don't remember for how long. but then we had some breaking news. So, with that welcome Yam Peg. Welcome Wolfram, welcome Nisten Yam. You seem the happiest so far. So let's start with you. What is the one most important AI thing that, you had this week?

Well, Kimmy Ships.

Kimmy Ships absolutely just dropped breaking news as the show started.

Nisten, what's up?

Yeah, I actually really liked the new cursor

integration for a web agent. you have a little window of a browser and I think Cursor like really pulled it off. So IIII like that one.

Houston and Wolfram, welcome sir from the spaceship

with your, you know, makeup on. Finally.

Greetings.

Mm-hmm. And I'm also excited like Yam Said if my, one of my favorite models get the thinking model, and it, it's so good in the benchmarks, then I definitely am super excited. So Kii has always been one of my favorite models since it came out. Kimi K2. And, yeah, I'm excited about this. I checked Op, it's not there yet, but, as soon as I can actually use it,

it's already there from, from another breaking

news flow from five minutes ago. It's on Open Router. Great. straight up, from the Kimmi, Kimi folks. Excellent. righty folks. So I think with this, we, we can probably get started with the show. I will just say mine. probably because we love open source here, right? So, so Kimmi key two, raising a thinking model that is 1 trillion parameters is definitely mine. with only 30 to be active, we're gonna chat about Kimi as, as much as possible. Folks, it just came out, we didn't have time to react or test it. This week I had a chance because it was like a fairly slow week, I had the chance to finally get to some, long awaited automations, that I wanted to do for Thursday. I did them with N8N, which is the kind of the workflow builder that many people use and many people love. And I gotta say, like, after a fairly steep learning curve, I got to a point where I have something that works and even tested it out on Kii this morning. I cannot do the same motions to bring you the news, and I need a specific format. I think I talked about this. and so I built that up and I would love to show you later on the show, so if you're tuning into the show and you never tried NA 10 and you wanna know what's that whole thing all about, it doesn't require any code, well, almost any code and what little code it requires, you can probably use, AI to help you with as, as, as I did. All right, so before I start the show, folks, today is gonna be a very interesting show. we're gonna start with open source. I kind of like underestimated how much open source is gonna be talked about, but it's gonna be fine. we will also have a guest on the show today, because there's a whole thing with CPS not being the right format for AI because blah, blah, blah. we'll figure that out. to help us understand this, we have Kenton Varda from CloudFlare joining us a little bit later. Kenton is one of the authors on the code mode post that I think we mentioned once before. but we'll mention why this is relevant to this week in, in, in a bit as well. So please stick around. We have a great guest, is the TLDR four, November six. Let's go.

All righty. This is everything that we have, including the breaking news from, the AI news world in, in, in, November 6th for Thursday. I, the host and guest today. Well, you know me. My name is Alex Volkov and the AI ventures with Weights, & Biases. We also have our co-host here, Wolfram here in this, Yam Peleg, and, maybe LDJ Ryan will join a bit later. And our guest for today is Kenton Varda from CloudFlare. Kenton, is head of engineering for the workers platform, but also, did some work on MCP, in open source. There's not a ton of news this week besides the big one that was just announced as breaking news. small training playbook from Hagen Face was released. they basically dropped a book. It's a 200 page end-to-end guide, to pre-train and operator lambs, including like post-training and other things. It's really, really like, really extensive. we're definitely gonna point the finger to say, Hey, if you are training models and you are interested in the latest and greatest, the folks from Hug Face released, the mecca for that knowledge, folks from AI two, our Institute for AI released my Earth. It's like foundational models and end-to-end platforms. Fast, high reasoning earth intelligence. And then also Moonshot AI release is Chemic K two. It's a 1 trillion parameter. MOE agent with 256 contact window, and a huge tool calling capacity. This is insane. Folks. We definitely have to talk about this. like they claim some incredible stuff. state of the art on humanities. Last exam beating GPT five, beating CLO on 4.5, sinking. This is an open source model that beats the main guys. This is, insane. We're gonna get to this in a moment. we're gonna move on to big companies and lambs and APIs. And there we have some news, not a crazy amount of news, but we did have finally an announcement from Apple. As you all know, We're waiting for Apple intelligence to pick up and become something. and then we have an announcement, that Apple will pay roughly $1 billion a year, which doesn't sound that much if you consider apple's size and the amount of money they pay for Google search to license a custom 1.2 trillion parameter Google Gini model, for revamped. Siri, we've been waiting for Apple for a long time and we'll talk about what this means. but it's really funny that, apple after all this time is just gonna get Google's model also in the big companies Amazon has sued perplexity. You guys seen this, legal threat to block the comment AI systems from shopping on Amazon, and I have a lot to say on this topic a lot as I told you before, my browser is my browser. I don't want the company to decide what I can and cannot do and which tools I can use. And if I have a mouse that auto clicks or have an AI that auto clicks, they shouldn't give a damn. So I'm really, I was pissed about Amazon's like attempt here, specifically because Amazon was all over the place in news this week. They also announced they multi strategic, multi-year strategic infrastructure partnership was OpenAI. So basically OpenAI will get GPUs wherever it can get them and Amazon will work with OpenAI but also will super plexity, probably completely different team. I think that's pretty much it on the big companies folks, unless you have anything big that I missed. We'll move forward.

Not from the big companies.

I have some open source stuff,

yeah, you wanna add some open source stuff to the DDR so quick?

Yeah, there's a new model that's Long Cat Flash, Omni.

So it's Long Cat Flash, Omni, it's not open source, 560 billion, 27 active omni model which does text image, audio, video input, and does text and speech output. So that as a new model. And there's Oval by dancer's, new Apache license, 1.4 billion, 2.6 billion, two versions, looped language model. So this is a research model which is, not a large language model, but a loop language model that goes through the same layers multiple times to improve its research. So that is a research preview that also just came out.

Alrighty, I robotics.

We're gonna talk about this. humanoid, robot from Xang that exploded all over my timeline. Xang released a Chinese company, and they presented the iron humanoid robots, supposedly, production 2026. And for some reason they have a, a man robot and a woman robot. and we should absolutely discuss why that separation is needed in humanoid robotics, and why, why are they doing this? So we definitely wanna chat about that. and whether or not, people here will get a humanoid robot. And then this week's buzz, is turning into this week's tutorial where I wanna show you some stuff. this week, specifically I have built some automations for us here at, Thursday. I, and, uh, they have been built with, AI agents inside N eight n. I would love to show you what that's about And then, I separated coding with AI into a separate category this week. we have a new segment called Coding with ai. And, we have some news on Anthropic, released a very interesting blog post that says that, It's very interesting from the company that brought us MCP, they brought a block post that say, Hey, we're all using MCP wrong. And the way to use this is to have LLMs right code to execute against the MCP spec. so in order to talk about this, apparently CloudFlare already explored that idea with code mode and we have Canton Varda here to talk to us about why that's the case and they released their blog post a while before Anthropic. That's very interesting discussion. definitely just for coders, the cognition. Windsurf, released code maps and, AI navigator maps of your code base, powered by Swyx 1.5, which we, talked to you about last week and that is really, really cool. so we'll talk with Kenton as well. And then Nisten, you also mentioned that you had a, coding related cursor thing from this week. I think. Can you gimme like a one line summary of what course added? I haven't followed,

they added in IDE browser, so they're doing what browser based does.

Yeah, I saw this, but in IDE and it was very performant, so that's what I, I noticed it was actually very fast. I saw tweet from folks that makes these things that was impressive to, to look at.

I saw tweets from folks that say, cursor is my new browser

and showing they're browsing X in the, in the window pane of cursor and having cursor run automations. in tools and gadgets, the last thing that I have folks, as I told you, besides gmi, not a huge week, sandbar launches something called Stream. It's a voice first personal assistant, that exists in the ring. So they, the, the, the AI hardware assistant format is a ring. so that's very interesting how, that's gonna work. And I think it's time for us to dive into the show, folks. It's not that big as a week. And we're in TLDR for forever. let's dive into open source. I think,

There's Maya one, which is a new voice generation model which

benchmarks, even bad chatbox for instance. So I put the link in the channel and this seems very interesting as well.

Alrighty, I will definitely take a look at, The

link where you send the link. I also saw, if we were talking about voice already, I saw that, artificial analysis posted whoa, whoa, whoa, what's going on? Apologies.

That's the thing.

Yeah.

You see this, the in world, voice ai. Yeah. Okay. So in world TTS is supposedly number one on TTS and artificial analysis. I haven't heard about Inworld, so might as well put it in here in world. So folks, as you, as you in, you notice that some news get updated as, as we kinda understand them as well. but definitely we're starting with open source and breaking news right there. And then

For speech to text, they, also did an update that WPA is

no longer the number one of source and they, posted that now NVIDIA's Cannery for example, or Paraki and even Rockstar Small for missile. they did very well as well.

I saw this and my question to them was, is it multilingual, or

just English that they're testing?

my most important question

Alright, Let's dive in.

Let's start the show and we'll start with open source.

Open source ai. Let's get it started. And we are getting started with open source ai and I'm considering to talk about everything that's not ki super, super quick. So we can talk about gimme at link, and then I can hit that breaking use button. folks in open source, I wanna highlight like one, one more thing as as I highlighted in the TLDR. we have a book essentially from HuggingFace. There's the whole like book on training and post-training models on hanging face. I'll just briefly go through this. I dunno, folks yam, if you have a chance to take a look at that, if you wanna give any comments on, on the depth of, of how much There book

is an understatement and that thing is interactive.

I mean, I saw it. I thought, okay, that's cool. It has a lot of visuals and so on. But then I realized the whole thing is interactive and maybe I can even show you. Yeah, the thing is insane that it goes, it's not just you think about the data when you think hugging face and, and small lamb and so on, but no, that thing goes through everything through the architecture. it go through everything, from the ground up, shows you all the alternatives. just click the link and scroll. It's beautiful as it is. I mean, it explains itself. basically they just went and experiment, reversed engineer so many details that we didn't know, just brute force them through experiments, detail that usually kept behind the closed doors of the people that share that trains, large frontier language models and hugging faced. Just experimented them out, and just shared with us. you see alternative ways of constructing a model, constructing a data set, constructing, evolves, whatever, whatever aspect that you think you have, all the different alternatives and like what happened if you try them. Like what is the result of the experiments if you happen to go and choose this or that, and it just goes on and on, and even ends up with the, how do you train on. distributed system with many GPUs, which is not trivial at all. they also have a completely different section with like 200 pages about that only.

I will just say, the, the folks who released this, been

on the show, Eli Bacouch, and, there's like a whole thread that discusses the whole book as well. It's very, very, very extensive. and they obviously HuggingFace built and trained small lm, in their own family of models that they released and, they like based on that experience. Alright. yeah, we're seeing now just as, make it large a little bit. Yeah.

Okay, so, okay, so they, I just wanted to just show

everyone how beautiful that thing is. I mean, it starts with some charts, but you go down and, and just look at this thing one moment

and it auto switches to dark mode on your system.

yeah.

Oh yeah. I mean, this is the real power of vibe coding. you wanna make a visualization and show everyone the thing that you're talking about. You could just generate that today, guys. this is HTML.

would've paid for this.

Absolutely.

I wanna show you the rope, which is,

can you remind folks with rope is

sure.

So, positional, encoding, transformers. look at this visualization.

yeah, it looks up.

Okay.

Basically, positional, encoding, language models, transformers, can distinguish between, the position of, the tokens unless you specify, exactly like you add this information. somehow into the model. There are many ways to do this. one of the most popular is rope, rotary position on encoding. And there is a little bit of math, about what exactly is going on here, but I mean, I haven't seen this visualization ever. if you appreciate this immediately and it just goes on the whole, the whole thing is just about tokenization and, I wanna share your GPUs, That thing is interactive. This is an amazing, amazing, amazing resource. Absolutely.

I think, insane for them.

But also we should say, and we should bring Eli, to talk about this a little bit at length. they live the spirit of open source, and this is like part of the release, not only releasing, models and putting infrastructure to release models. they also, this is obviously great for everybody who wants to train stuff. There's tons of information here, inside information and it's really, really extensive. So shout out to the hugging face folks, for this. Folks, let's move on. Let's move on. We're moving on to the next thing in open source. We're gonna talk about, we can mention that AI two, Allen Institute released Alma Earth, which is, I'll just show this right here. they have foundation model for, planetary Insights, which is, earth observation foundational models, end-to-end platform, designed to run terabytes of satellite radar elevation, map layer data into decision ready maps and alerts in hours or days instead of month or years. anything we wanna mention about this folks? Have you take a look to this I think it kind of went under the radar this week. It was very chill. but we should shout out L and Institute. They're not only building LLM, this is not on LLM. This is like a very specific, research tool. All righty. Okay. And I think it's time for our breaking news and the thing that we are all excited to hear, and this has just happened, 45 minutes before the show. Let's get it. Hey, I, breaking news coming at you only on Thursday. I.

All right, the breaking news for today is that, even, as of 45 minutes ago, the team at Moonshot released Kimi K two thinking it's an open source, 1 trillion parameter model. It's an mixture of expert model, with only around 32 billion parameter actives. it has 256,000 tokens in the context window and an insane tool calling capacity. I think they're talking about 200 up to 300 sequential tool calls without human integration, They're not talking about how long the agent can run for. They're specifying specifically how many tool calls it can run in sequence, which I haven't seen other models talk about. So it's hard to compare, but definitely very exciting on the evals. You wanna help me cover the evals thing? I'll just like shout out that the humanities last exam here, on humanities last exam, which we know is a kind of,

Human Last exam.

Human last exam, copy of copy of real copy. Title, final. Yes. Final. Final. Yeah.

And they also mentioned this in the blog.

We should mention this as well, that, giving a model access to the internet to hack I face specifically cheats on Humanis last exam. Somehow, I'm not sure, they mentioned that we should take a look. they get a crazy 44.9%, beating GPT five and cloud on 4.5. Thinking, on this supposedly very, very hard last exam, which is, you know, obviously some bullshit. so what else is very interesting here? Browsers, brows, comp. I haven't seen about, I don't remember Amp as a, as a, eval. they're getting 60% here. A IME is quite crazy. 94 and a ME with no tools. and then with tools, I think they have, with Python they get 99% and with Reasoning Heavy, they get a hundred percent on A IME, such as GPT five. So basically a ME is toasted. A ME is, international Math Olympiad, in 2025. So basically a bunch of questions. the model gets 40, 48 versus, higher than 4.5 sonnet and very close to GP five. It gets 84 folks on GPK Diamond. Hopefully you guys are seeing what I'm showing. Yes. I will say, again, this is a open source model, although it's a one terabyte open source model. This is an open source model. And on every benchmark here that they've posted that they tested against GBT five and CLOs on 4.5, and they also tested against Deeps seek and GR four, they're either beating or very close up there with bigger models. from your reaction to this

a really good model overall.

everyone loves it. Everyone that uses it just extremely happy with it. It's different.

The non-thinking one, you mean like, 'cause not a

lot of folks are using this one from the, an hour ago, but yeah.

I've already started using it now that is an opera and my

first impression is really, really good. So I love it. Even the German, the writing style is there. So Kimmi has always been very special in that regard, not your usual AI slot. It's very different and, multilingual. So yeah, I'm very excited about this and I'll use it with my MCP service.

a hundred percent.

the reasoning part is really, really something I had a chance to test the, we sometimes do, uh, with Beth and, um, uh. Ice cubes. And I think besides the fact that like these models always answer this one correctly, the thing that we can get is like looking at the reasoning and the reasoning for Qi just feels different, just reads different. So here we are, we're in open router, we're testing qi. and you can see the reasoning trace stream, like as they go. and let's read some of this, just how and if you got access

to it through API because this is something that is unique to deep seek.

it's on the Kimmi API and also I think, an open router.

And I think Kim API supports the Anthropic angle. And one of the things that the model gets to, and the reasoning trace is they read differently than, Qwen for example. Again, the question that we're asking here folks, reminder for dogs who are not with us when we ask this question, is something about, a few questions from the simple bench. 20 questions. they're designed to trick l lms specifically, to look into the calculation part of the question, whereas the phrasing there is very confusing, to LLMs. So, the question is how many whole ice cubes are going to remain inside a pan at the end of the third minute? the emphasis is on the whole, and the reason is like we all intently understand the ice cubes melt the moment they the actual a hot pan. this is like a physics understanding the humans have in innate understanding. this, question is meant to trick lms usually answer by calculating how many times, numbers, whatever, reasoning. Models are usually very good at repeating and thinking through the process and saying, okay, well did the user intent. So I did see Kimi go into the options and try to rethink this. but for now I'm still not getting an answer. And this is, it's been a while.

So it's thinking a lot.

Compared to other models, it's not very fast. I think a lot of people might be hammering it and the only, service that is using it is their own. So it's their own hardware at capacity, not other providers yet. And it reasons a lot and very in detail that is also much more in detail than most other models I've recently seen.

it's working pretty okay actually with

the, it just plugs in, right? Oh, oh yeah. it just works.

What did you, what exactly works This Can you, can you,

I did a cloud code proxy and I'm using the

Qmi K two thinking model. I just plugged it right in and it just went to work right away. Nice. And I, I, I can show it.

It'll be great actually.

Show us.

I wanna know.

I wanna see how you do it

because for now, Kimmy is still thinking about my,

and it's been over two minutes. I think at this point Kim's still thinking about token. My still, still getting Token. Token. It's streaming. It's streaming the reasoning.

The pricing is also interesting because it has the same

pricing for the thinking and non-thinking models, at least on open router and I guess on key as well, on moonshot ai. So, you don't pay extra for the reasoning, you just pay the more tokens It is.

okay.

So the way that, this works is like you just start off a proxy and then you do the proxy URL. I'm just local host A 69 and I can just say, review the repo, start to subagents to review. And, yeah, it's sending the calls now to Kimi K2, and, by the way, this is my own Ls. it works. it was a little bit slow, but it did do a to-do list. yeah, it's working right away. And I can see the calls here, I can see the proxy calls going to Kimmy.

Mm-hmm.

And, it's not really starting this subagent

yet, but it just plugs right in. Latest CLO code 2.0 0.34. What's it the date? November 6th. it just works right away with thinking on, You can just plug this in. It started doing tasks. maybe if somebody from the audience wants me to plug something in, let it go for like 30 minutes, see if it builds something, we can, but, this, this works, like you can just use this as a clot proxy and, it, it just works.

I think Kimi exposes a Anthropic compatible API even without

a proxy because what you're doing is plugging in the OpenAI, API, right. And then turn it into a cloud. API, if I remember correctly, Kimmi is exposing on Anthropic like kimmi themselves. The, the API tier in Kimmi. I would just say, for now, that, well, Newton looks it up that I'm still waiting folks. this is a little bit kind of like sad. I think maybe my, my, my stream got stuck, of, of the answer. But like you can see all of this reasoning went into the question. Again, these questions like are meant to trick LMS to put them in the loop. but I was expecting an answer by now and I didn't get that.

this is good.

This is very good. I told it just like review the code base start subagents. It just worked, it did todo lists. yeah, this just works with cloud code. You can just swap this in with cloud code for, 60 cents in $2 out. Yeah, this is great.

Alright, let's talk about the capabilities of this model super quick.

and then, tell the folks what we're here to tell them.

just one thing, Alex.

Yeah. Because you didn't get a response. I think that is that server, because there's only the official moonshot AI that is providing the model right now. And I've had stream errors as well.

say even before it got stuck, look at how much reasoning

it did for, for this question. And it, it went back and forth and we saw this before it went back and forth and like tried to understand, things. But here's some, parts of the reasoning. Perhaps the puzzle expects that each minute, half of the ice cubes melt away. and then it goes like, Hmm, not integer, or, it's very interesting to see the, the reasoning traces and, and Kimi has always been. even the, the, the regular model, they've all been a little bit different in, in how they train them. So like, very good models, many people like them from that, perspective. so again, very long context here. 256 tokens for the, for the reasoning models. Ohmic, keep the thinking. I think the highlight you were asking about this, they releasing the reasoning content here. so like they add this to the fields in open I is the K, and also I think in Anthropic is the case. Well in streaming output. this model as far as I saw is this multi-step tool call, they say the communicative thinking is designed to perform deep reasoning across multiple tool calls, inability to tackle higher complex tasks, to get reliable results. Include the entire reasoning content from the context. The model will decide which parts are necessary and forward them to further reasoning. And I wanna highlight that like when we had, do you guys remember we had, minimax two, we had call Meow on the show and he basically said the same thing for m minimax M two to work reliably and much significantly better. something called interleaf thinking is needed where you pass backwards to the model, the reasoning traces as well. And for. Models, like OpenAIs models, that's not necessarily possible. They don't expose the reasoning traces to you. but for open source they notice that like it's significantly better to send not only the inputs and outputs, which most chat bots do, but also to, to pass the reasoning tokens back.

Actually.

They just added this for, I think it was Minimax M two. Mm-hmm. And Kim uses it as well. And by the way, for OpenAI models, they use responses, API, which does the thinking. Well, it's encrypted. The reasoning traces are still there.

All right, so, what else can we talk about the model here?

we can take a look at some of the evals here. basically, I think the highlight there is the sequential tool calls, but I'm thinking about the context window 256, if they talk about 200 to 300 sequential tool calls that this model can go for and for agentic use cases, keep going and keep calling tools and reasoning about them I think this is the highlight of the agentic, model era. And it looks like, at least based on their, evals, this is something notable. every other lab, as we said, measures this by just time to run,

I must say that on this topic, I don't have an eval that I trust

specifically because, simply because vibes and like real life, you can give Codex a single prompt, which is not that elaborate with the PRD document and so on, and it'll go off for an hour and come back done and it'll do whatever, like called a million tools in the middle. You don't really know what's happening.

Yeah,

it's not the same with Claude.

Claude is more like iterative, which is extremely useful as well. It's just a different method of working. So when people say, you can call this and that tool or send it for this and that time, it really depends on the use case. And I haven't seen any, unless you guys know, any evil that you really trust in this topic. I haven't seen any evil that really represent this.

As far as agentic runtime, I think METR is the place where I go to

for those evils, mi miter, they're showing agent execution in hours basically. And I think that this is like a metric that folks are looking into. not necessarily whether or not it'll give you the task at the end correctly, but just like for how long it'll work to give you a task correctly. I saw a funny thing about this where somebody said, G PT five goes off on 45 minutes, comes back, wrong answer. And then, composer 1.5 13 seconds comes back, wrong answer. And so like, you know, at least you fail faster with the faster models as well. I will say that, let's just stick to the open source like, nature where we are right now. communicate to thinking, is. I think they also have built-in tools in the API as well, like web search and precise tool calling abilities. it remains to be seen still every time a model like this releases, especially like a new reasoning model. like as we just said, Interleaf thinking is part of how their API works, for example, built in. And, we'll see other providers put it up. Hopefully we'll put it up also on WB inference and tell you guys about this and give you some, tokens. it takes time to shake out how to exactly use the model in what sense? so Tu one of the leading reasoning, coding models coming out from China. fully open source. I think Apache two, I haven't looked fully. technical report is also available. very interesting report.

very detailed and it's very detailed.

you get the real thing. I'm reading it right now. It's brilliant.

So we're on the technical, blog post of Kimi K2 Thinking.

And now we have graphs that compare 'em to g PT five and to son 4.5 thinking the absolutely latest and greatest models. They don't have Groen here, very interestingly, but that's all right. And on Agentic Reasoning Humanities last exam, they get, 44.9% beating both other models on humanities last exam. That's fine. but we don't see the same benchmark. We do see Swe bench verified. So, Kimi thinking is 71. Percent are GBT 5 74 and sonnet is the absolute beast at, at 77%. So, not quite beating the top labs, but again, this is an open source model versus the latest and greatest. Like we see open source catch up,

you got to see the demos.

Let's go.

exactly this

Yeah.

just look at what kimmi made, like the examples.

Oh, okay.

we're looking at some examples, and there's videos of things that Kimi built and we're seeing multiple examples. This one is a math explainer visitation of gradient descent. Aha. This is dope, bro. This is great. That's insane. That's insane. here is a simulation of virus attacking cells in the bloodstream. Wow. This looks incredible. and virus type is replicating. Oh, this is cool. Yeah.

click five.

Yeah.

What's five?

Five is a video that shows us. Beat simulator music, a beat simulator. The music actually plays folks, you can't hear it on the screen probably because I'm sharing the screen. Browser comp is a benchmark for Gen G reasoning. And we can see an example here with how many, searches the reasoning called. And this is insane. Look at this. I'm scrolling and scrolling and scrolling. And the model just like went into deep research mode and just like kept like think reasoning and then search reasoning and search. This is wow. Okay.

Okay.

I take everything I said back about being skeptic. I mean, I need to try, but we try. This thing is crazy. Yes.

It's a bit of a disappointment that the benchmarks they

did compared with just JGBT on Senate. I would like to see it compared to GLM and Quin as well. That would be very interesting.

Yeah.

There's also something about in four and low bit ization inference, Nisten, you're probably gonna be interested about this as well. they tested all of their, I think all, all of the benchmark were posted are in, in Four Precision. So, they, they got like, a speed up of two X generation speed while achieving state of their performance. That's very interesting. I think to summarize, we, we went a little bit deeper here because again, the release is new. We didn't have as much time to dive in before the show to prep and like organize notes for you guys, but, this seems like an absolutely state-of-the-art model in the open source probably gonna rise and, and beat M two max. We'll see, the scores look absolutely insane. I agree with the, well from, like, it's a bit sad that they didn't compare to the previous open source in, soda, for example the M two max, but it's all right. shout out to the Kimmy team, the moonshot team for this release, and I think it's time for us to move on folks for

Considering that Qwen Max, for instance, which is also great,

Qwen model is not open source, but they made their thinking model open source.

Yeah.

worth chatting about footnotes. They show something that, we selectively employ a subset of tools and reduce the number of tool calls under the chat mode in Kimi. So when you chat on kimmi.com, you may not get to the exact same scores. And it's very interesting that my test failed because of probably everybody's rushing to their service to try it. but still like it should be a little bit more robust than this as far as I'm concerned. folks, let's move on to big companies and APIs. We, we have just a few more things that, Wolf we mentioned in open source. So we'll just mention them and move on. there's a new open source omni model called Long Cat Flash. it's a end-to-end spoken interaction, 128 contact window. And, I don't know what sc MOE is, but like, it's a 560 billion MOE, Omni Omni model. 27 billion parameter active.

it's on their webpage and they have Android and iOS apps.

Oh, they have iOS apps for this or?

Well, I think it's time to move to big companies and talk a little bit about that. What do you guys think? Yeah, let's do that. So in big companies, I think we should start with, alright, cheers. thank you for joining us today, folks. We should start with OpenAIs. Just massive, massive, claim to taking over every possible compute, provider in the world. So, OpenAI announced this this week, or I guess AWS announced this week. new multi-year strategic partnership between AWS and OpenAI, where AWS will provide cloud infrastructure for OpenAI to do inference, to scale GPT, large scale training as well, and inference, in agent workloads. This is a, I think $38 billion deal and

I once knew a tech company that didn't partner with

OpenAI, there is none at this point. Like everyone is partnering with Open.

But, but now we have OpenAI doing business with AWS out of

all places AWS Amazon's web services. And there's, there's this graphic here that I wanted to show super quick and discuss this on the show, from the Coba letter, which they, they talk about like economics, et cetera. OpenAI now has, here's a list of things that OpenAI, deals with 500 billion target deal with, SoftBank, I believe, and Oracle, right folks, correct me if I'm wrong. a hundred billion dollars Nvidia deal, which I, you know, a hundred billion dollars a MD deal. And now $38 billion Amazon deal. $25 billion Intel deal with Intel. $20 billion TSMC, $13 billion. Microsoft, $10 billion, Oracle. And, with, there's also a multi-billion dollar Broadcom deal. This is a lot of money. I'm like, roughly counting. And I think, if you guys remember from last week, we, we showed Sam talk somewhere about having, oh, during the OpenAI restructuring live stream, they showed Sam talked about 1.4 trillion overall commitment to GPUs. Everybody's asking where's the money gonna come from, It, it's a very good question, honestly, to see where's the money gonna come from, because OpenAI is essentially like losing money and the company's only valued at 300 billion, oh, sorry, 500 billion generally. And I think all of these deals kind of add to the valuation somehow, but it's unclear how all of this is gonna get paid for. And Alman went on a podcast where he was asked directly, and his answer to this was basically, we're gonna make a lot of money this year, way more than you thought. to the investor, he got a little bit pissed off and he said, if you want to sell your shares, I'll gladly find somebody who would buy them off you as an answer, which is very, very interesting answer. But then they posted about like, you know, we're not actually fighting. We're good friends. This was just a, remark. So, going back OpenAI looks absolutely terrifying in how quickly they, take over the world. the other thing that's worth maybe mentioning is how much of the US economy and the stock market is now propped by. AI and these deals, and some of these deals are kind of recursive where OpenAI pays Nvidia and then NVIDIA pays OpenAI and OpenAI pays, pays somebody else, and they pay Nvidia. and I think the highlight of all of this is that at some point this week opening ICFO floated the idea that the government, the US government should, gonna like backstop or ensure these deals because we're getting to a point where open is like too big to fail. Because if this fails, the economy will absolutely, absolutely crash. these, like we're talking about trillions of dollars right now. I don't remember offhand the GDP of us, but like I'm pretty sure like if if something happens to OpenAI, stock market will follow.

Oh yeah.

They are pretty much, entangled with Nvidia at the moment and Nvidia is pretty much, single handedly pulling the stock market. So yeah, absolutely. it's gonna influence the stock market. I just want to, I just, you know, just quick note, many people saying it's a bubble, but this money is coming from cash. The companies have, It's not a thin air, it's like it's musical chair kind of way, like transferring hands. But at the end of the day, it's coming from a real, stockpile of cash. So, I think people kind of miss that when they say it's a bubble because the money is real.

we mentioned about the bubble part on the show multiple times,

and the thing with the bubble is, we may not think that the AI itself is a bubble. I definitely don't see a stop to inference, like with the way things are going. and we can talk about why in a second. but it could be an economic kind of over excitement that everybody's like trying to invest stuff, when there's not a lot of, things to be done with it. So we'll see. I think it's very important to show that, like, literally OpenAIs is partnering with every major cloud, Amazon, Google with everyone. I think that even this basic letter thing, there was a deal with Google, I'm pretty sure, and it's not here. at some point with OpenAI it's really hard to keep track as well, at this point. actually wanna move on to, to other deals that happen in the big companies and APIs, apple folks. Apple will pay roughly $1 billion a year. to license a custom 1.2 trillion parameter Google Gemini model. So Apple will make Siri smarter with Gemini. I think I saw somewhere they tested different models in different pricing points and OpenAI. Obviously already has an integration with Apple. and I think it looks like they landed on Gemini. they tested GPT, Claude and Gemini before selecting Google. so it's interesting that this is the first time that we see a size of Gemini. This is a 1.2 trillion parameter version of Google Gemini, which probably means that the actual Gemini is larger than this. because Kimi is 1 trillion parameters, right? Gemini is probably larger. apple intelligence model is 150 billion. So this is like eight times larger it's gonna run on apple's, private cloud which is supposedly, better, for, compute and data stays walled from Google. I think this is the most interesting policy. It's not an API deal. It's an actual, Hey Apple, here's the model that we trained that you can run. Almost like open source, but not really open source or we're not gonna get it. Only Apple's gonna get this one. And Apple will pay $1 billion. A cool, 1 billion annually.

Do you remember the amount Apple spent a day?

You remember these reports from two years ago? 'cause we, we can all get back to remember if you really want to, apple is spending a lot of money a day training a hundred model. this piece, this is the piece from I think 2023 sometime that, at, at the peak of the craze of tragedy, BT we got, reports that Apple is spending so much money every single day training their own. I don't know what they actually did with it, but, like, it's weird. apple has pockets, has talent, has compute, has what is going on.

on the other side of the world, even smaller

companies are releasing open models and apple, nothing happening there.

Yeah, there, there was talking about Apple, like

buying perplexity at some point. Nothing materialized, but yeah, this, it's weird. Wouldn't see anything coming from Apple themselves, but hey, as a user, I don't care. Like as long, as, long as sir is gonna be not dumb. And I'm waiting, I keep waiting for Siri to not be dumb and it's, it's not happening. Alrighty. I think that this is mostly on the big companies as far as I saw. the last thing that I wanna talk about is this like very annoying thing, perplexity, ai, the search company takes on Google, posted on their blog post that They got a threatening email from Amazon that says that, they require perplexity to prevent the comment browser gentech, agent, Gentech agent, to stop using amazon.com, a legal threat to a, agentic company to have them prevent that agent from running on a specific website. I find that absolutely abhorrent. I just wanna mention that like this, this makes no sense and, perplexity replied and said, bullying is not innovation with a whole blog post that says, Hey, we will definitely not use this. And, I find it very interesting previously chatted about like, CloudFlare for example, working. To automatically block AI agents from your websites, unless you like, permit this with robot dxc. And I always thought that there was a difference between automatic AI agents that I can like run up and scale versus an in-row thing that behaves like me and does the shopping for me or whatever. there may not be an actual technical reason to differentiate them, but I always thought that the reason why I open the I is the browser, for example, and Comet, is that this browser has access to everything as me, locked in as me does stuff as me versus the browser. I can run as agents, manness and the previous, open air operator and stuff like this, they kinda like spin up an environment somewhere. I have 17 of them do some stuff to me. There is a difference between them. one operates as me, one is kind of like just does tasks for me, but not as me necessarily. And because the one in the browser access me, I don't see any option for companies to block access to it because how would those companies know If it's me or, or my agent

I can run cloud with computer use, just controls my computer.

Move the mouse. That's literally me. That's me on the computer. It's like I take my friend and put him in front of the screen. Tell him, go and buy something, on Amazon for me on my computer. right now you can distinguish them technically because you can find the user agent. But that's just the technicality. it's really on the wrong side of history. No doubt. because the next step is you need to block Google Chrome and, and good luck with that. So I don't see where it's going.

and we are talking about AI assistant, so

I want my AI assistant to be treated like a human assistant. It shouldn't matter if I set my Amy or my wife to buy something on Amazon, I send a message and say, okay, I want this, buy it. Why should it differentiate if it is an a i agent or if it's my wife or an employee or something? So just because of advertisement or anything that is like saying specific people are not allowed to go to my site. It's not like it's abusing the Amazon system itself. If it's buying and it's a legitimate purchase, what is the issue?

Yeah, I think it should

be allowed

perplexity, it takes the same exact positions we take

and say the future of user agent. User agents are exactly that. Agents of the user, they're distinct from crawlers, scrapers, or bots. User agent is your AI assistant. It has exactly the same permissions you have works only your specific requests and acts solely on your behalf. And I, I like, we're getting to a point. At least I'm getting to a point folks. I know that like sometimes on the show, we talk about the latest and greatest and we bring you some tools, and many people don't use this tools, many people don't know they exist, for example. But we are here to give you kind of like where things are going and where things are going is everybody will have, things like this working on their behalf. pretty much everywhere. And after our conversation with, with Canton in a moment, I'll tell you about my journey with NA 10, for example, and how I used Atlas for that. we're moving toward the point where like all of this is, becoming second nature to many folks. And for Amazon to try and bully the company that's smaller than them, to just force them legally to do so. It's very interesting because Amazon is a huge company with a lot of lawyers, and, perplexity is big-ish as a startup, but not, you know, obviously not work close to, to, you know, Amazon scale. so this bullying will not stand and f you Amazon, Bezos should come back and, and tell people to stop doing this, this bullshit. but I want my AI system to be able to do the things that I want as well. So shout out to Perplexity for taking a stand on this topic. By the way, perplexity has gained a lot of my respect with the stance and response and the fact that they're fully, fully, transparent about this as well. just one thing. Yeah. Because we

can't directly boycott Amazon.

It's so big and we use it all the time, but we could tell our agent to avoid them and go to the source and buy from them.

They can lose some.

the bigger this is, I just wanna

say, it's all of this is moving towards everyone is sticking,

sticks in everyone's wheels. we also saw meta blocking OpenAI on WhatsApp, which is, I don't know. Cool. It's not a cool move at all.

think meta blocked everyone.

Like I think Meta changed the, yeah, meta just said we're gonna change the terms of use on WhatsApp. That only we can offer AI tools as automated bots. Also semi not cool, but I think it feels different because it is like, this is our platform, this is our playground. Only we can offer ai.

Yeah.

It's different because it's their platform

Well, all folks, we have to move on.

'cause we have a, a, a, a new interesting discussion. we're moving to our new segment of coding with ai. I'm gonna add our new guest here, Canton Barda from CloudFlare. Hey Canton. because this week it is very interesting. So we're moving into coding with AI folks. if you're coding with AI or you're doing any type of, agent applications, et cetera, stick with us. I wanna introduce Canton, for a minute here. specifically because this week Anthropic has dropped a blog post. but this blog post is very interesting, because I think. around a couple of weeks ago, we celebrated the one year of MCP existence, MCP being model context protocol, the way to give models like tools and explanations. And we talked about MCP multiple times. And MCP has kind of had a very interesting, sigmoid curve of adoption where in the beginning of the year, everybody's like, this is the only thing you could talk about all your Twitter feed was MCP. and then at some point it's kinda like, all of the companies started adopting this Google and OpenAI OpenAI. Recent apps are built on top of MCP, so the spec one absolutely. However, the excitement kind of lowered a little bit and the importance of people knowing and building everything is MCP lowered a little bit and now Anthropic releases a blog post where they basically say something like, Hey, MCP is great, but it's, overloading the context for the models, these models are much better at writing code than using MCP. So why not write code to use MCP? And I was like, wait a minute. I remember a blog post like this from before. I think CloudFlare released something like this. And apparently Anthropic, tagged. and attributed some of the earlier work to CloudFare in their blog post as well. That blog post was authored by Kenton, as part of the Kenton and some other person. Feel free to please mention their name as well. So I wanna Yes, I welcomed Kenton to the show. I decided, hey, who better to discuss with us than the people who actually came up with this first or before Anthropic? Kenton. welcome to the show, this long preamble just to say, we saw your work on this before Anthropic released it. I would love for you to discuss with us what they released. but let's start with who you are. And for folks who don't know you and your first time on the pod, please give us a brief intro.

Yeah.

Hi. Thanks. thanks for having me. I am, Kenton Varda. I, started the CloudFlare Workers project at CloudFlare, which is our developer platform. started it when I joined about eight and a half years ago. Continued to be the lead engineer on that. a lot of what we do, what we've been focusing on lately is, enabling ai, you know, AI written code or AI agents to be, built on our platform effectively. the blog post that we released about a month ago, we introduced something that we called Code Mode. I can't completely claim to have invented it. there are other people who, have a similar idea in the past. In fact, a lot of people have, explored this in the past. But the idea is basically, So MCP is a protocol for exposing tools to, to AI agents. And they're exposed as these function calls, which is a, just a feature LMS have had for a while. What we did instead, we still take MCP servers, but we take the tools and instead of exposing them as tool calls, we expose a, we convert them into a TypeScript interface and then tell the LLM, Hey, here's a programming interface you can write code against. Please write code to do whatever it is that you wanna do. and we think that's for a couple of reasons. One is that LMS have a whole lot of training data on, how to write code. You know, they have all of GitHub and such. Whereas for tool calls, tool calls are sort of this artificial thing that LMS have been built to be able to do. But they are trained sort of artificially with a training set created by their developers, and not on, a whole world full of examples. And so it's like, it's not their native language, whereas code kind of is. And then of course, the other thing is when you're using code, you can call multiple tools. you can take the output of one and stick it into the next without having to send all of that through the neural network. So it's faster, it's cheaper, you can get more done. and Anthropic had their blog post this week basically confirming that,

you guys were in the right direction and

Yeah.

I love this.

And I think, I saw Simon Wilson, a friend of the path also comment on this, and they said, the exact implementation in the Anthropic blog post of how to actually achieve this is remain this, draw the rest of the out type thing where like, fill in the blank here, you figure out how to actually do this. Whereas MCP has a spec that kind of said, Hey, here's MCP, here's the library from MCP. Here's how to use this. Here's library from servers, here's library from clients. And here they're like, basically we researched and realized that MCP, the way it's implemented right now, and we should, we should show the people how it's implemented when you add an MCP server to like windsurf or cursor, et cetera. that MCP implementation is kind of like a documentation of all the tools available and a long description for DLM of how to use this tool. This is how DLM knows to use the search tool versus the file tool, et cetera, where it has access to a bunch of tools. very quickly I saw that, the context window of your model is getting overblown by the descriptions of these cps, and the more you add the less performant it becomes, can confused, et cetera. and drop basically says, tool definitions overload the context window and, intermediate results consume additional tokens. And they're like, okay, so what do we actually do? Here's how the tool definitions look like. what do we actually do? we basically, we need to figure out a way to have the model right code Canton. Here's my question for you. And yeah, they did, they attributed some of the work to what you guys did, because I think you also released a way to do this, right? Like code mode is an actual implementation of, of such,

yeah.

so good news. we actually have, working code for this that you can run today. in our blog post, on code mode, we extended the CloudFlare agents, SDK, but we extended ours, to implement this model. And, you can run it locally today it relies on a new feature in CloudFlare workers called Dynamic Isolate Loading, or Dynamic Worker Loading, which, basically lets you, spawn workers on the fly that contain, the code written by the LM and run them immediately. And this is really interesting, we think, because, The workers platform has always been based on a different, execution technology. Instead of using containers, we use something called isolates. They're, basically JavaScript on time environments of, V eight isolates. and they start up much faster and they're much lighter weight. So if you're just trying to run a little snippet of code, you can start this up, run the one snippet, throw it away, and that's fine. That's efficient. You don't have to prewarm anything. it takes milliseconds to do. So we think this is probably the way forward if agents of the future are all writing code behind the scenes all the time and running these little one-off snippets. But, yes, we released, all the code necessary to build agents this way and we're continuing to improve it and, you can see it all on the blog post.

Yep.

So that's great. We'll add the blog post to the show notes. Kenton, the question that I have, and I think I had it back when you risk called mode, and just before the question, I wanna show what, Anthropic claims when they can attack you as well. all this work, lets agent load the definitions that it needs for the current task reduces the token usage from 150,000 tokens to 2000 tokens at, and the time cost savings of 98%. So this is insane to me. This is insane. And Anthropic release is a way to define how to give model tools. Superpowers a year ago. Everybody adopts this. And it takes a year for everybody to understand, or maybe less than a year, in your case, maybe like 10 months, that this is counterproductive if you use many tools. And there's a better way, the better way is to have the model right code. And then I am remembering when MCP, the height of MCP excitement was there, many people were asking the question of like, why do we need another protocol where rest exists? Where open API exists where the model can basically read open API specs, try to understand all of your API endpoints versus the select few tools that you decided. and now we're doing this loop and we're back to, okay, the model can write code. captain, I'm gonna put you on the spot here. If you don't have a good answer, we'll figure something out. But why would I have my AI write code for against an MCP server versus an open API server, which is a very standard, like restful definition what many APIs have.

Yeah, so MCP does still provide, a few

things that are useful here. it provides sort of a uniform way of connecting an agent to a particular service and authorizing it that didn't quite exist before. in the past, like we have OAuth for authorizing against APIs, but it's always been the case in the past that you're writing a client for an API you're usually writing a client for a specific API. And so every OAuth service can have slightly different requirements for how you authorize and set that up. And that's okay. 'cause every client, right, so the particular service, but with agents, you want to connect, every agent to every possible service. And you don't want the developers of the agent to have to know about the service you're connecting it to or vice versa. So now it's this instead of one to end, it's end to end. for that a little bit more standardization is needed and MCP is, is providing that. so it could be that, we start to see some of those extra bits being pushed back into open API and all these other standards so that we can then more easily connect agents to all those existing things as well. I think that would be good. but we'll have to see where that goes.

so I think that, it's a very interesting, shift in mindset

from the peak of excitement of MCP which is like a new technology. Many people who adopted it before everybody else, got very excited as MCP experts. And then we saw a bunch of MCP. Storefronts exists where you can look at the tools and collect tools for your LLM, et cetera. And we got to the curve of like, okay, how many CPS do people actually use in their tools? that turns out not to be that much. what's your experience as well? And the beginning of MC was like trying tools and everybody get left with like maybe like two or three tools in their cursor.

I definitely have more than that.

And regarding MCP, I was very excited about it. Not because MCP is something special, but because it was a first standardization effort that got a lot of traction. Before that you had to have, you have had AI write your own tools for every instance, for every interaction. Now you have this standard, if it writes code or calls it directly, it doesn't really matter as much as long as it gets the job done.

Yeah.

And I think it's worth mentioning as well that it's confusing a little bit. The tools that we use, cursor windsurf, OpenAIs Chat, GPT now supports MCP Cloud desktop was supports MCP. They are now exposed to a bunch of tools that they weren't before. Even though before those tools also had their own APIs. It's not, it wasn't like, I think that speaks to what you're saying. it's still worth. figuring out, okay, this new standardization in the world of LLL makes sense even though it gets the context window, blurry.

the integration is a key thing.

Now you can connect your tools or whatever you want to use to JJBT to open web UI to cloud code. Yeah, that is a cool thing about this because you could not write your own, code to do this with these interfaces before and now they all hook up to the same standard that is the SB part of the, MCP where it is a universal connector. And that is a great thing that is very useful. That's why I have so many tools. I don't use them all the time. I specifically enable them, like if I want to ask the LLM in open web or j GPT about my emails so I can hook my email up to it and enable it and it's only in the context if I need it

So I think it's worth, talking about the implementation

that topic talked about as well. And I think, Wolf you're basically doing manually what they are offering as a system to do and what Kenton and other folks offer the system to do instead of like. Logging in all of the tools, sending to the context window. Here's a tool, here's how to use this. Here's the tool, here's how to use this. instead it would basically scan the available tools and then provide just the spec of like, here, like the possible tools. and topic is doing this via their, the skills feature somehow. I haven't really delved into this, but there is a difference in implementation, right?

that's what I wanted to ask.

I mean, if I get you, correctly, the argument is that cps, they sit in your context and they take away, parts of the context and that's gold. That's performance. You're living on the table.

Yeah.

Absolutely.

but what I'm saying is that, okay. What about, progressive discovery through tool, through cps? I mean, what's the difference between, these and actually generating code that queues the functions? just so to make sure that I understand correctly, like what are we talking about?

There are meta MCP tools already and proxy that, decide based

on the request, which MCP tools to give. Mm-hmm. but if the model initially doesn't have the context, available tools and what they can do, it can hallucinate, for instance. So that is also not a perfect solution that I think there is no perfect solution yet, which is why this new approach about the code is, yeah, a new approach.

Go ahead.

there were really, two parts of Anthropics blog posts and only one of

them was really referring to, our work. the two parts were, you know, one using code so that you can string together multiple requests, not send everything back through the LLM, intermediate results. but then the other one was this progressive, discovery where instead of putting everything in the context window upfront to say, here are all the tools you have, it's, a tool that lets you find out about other tools by a file system tool. And then you have to dig through the file system to find the definitions of particular APIs. and that's something we're gonna be adding. 'cause that makes a lot of sense too.

Yeah, I think this was one of the issues against MCP in

the beginning where folks basically said the discovery is still on the user and it should be on the LOM. And I think when we talked about a two A from Google agent to agent protocol, they have the concept of a registry where the agent can basically go to registry, see what are the tools that are available for me. And now we're seeing Cana Anthropic like getting there, with the second part of what you said. And basically. If you, well, they're doing everything via the skills thing and the file system, which we haven't dove, dove in deeply into the skills thing. basically Anthropic is able to discover whatever skills they have. So now they're kind of connecting the MCP in the skills and saying, if their AI is able to go and use the file tool to go and see which skills it can have generally, then it will pull on the skills and their definitions, the full definition of how to use this, which parameters to send, what responses they will receive. Like all of these full definitions, that take a lot of place in the context, it will only go and tag at them when it needs to. This is the discoverability part. so it looks like another philanthropic is connecting the two things. well, it's very interesting. Ken, thanks so much for joining us. Our discussion for this. I will just shout out super quick. I did not know that you, started the Workers project. I love Workers as a platform. Incredibly, like I built a startup, a full startup on top of this. I was part of the first launchpad of CloudFlare workers as well, with tar. I still have a bunch of stuff that just runs there and I don't even know where, where it runs. It just works entirely like automatically. So, it's great to see that you guys are like, you know, building tools for AI as well. And, agents is the case, part of the workers, platform as well. so shout out to you for that incredible work, which I'm, a very big fan of and have been for, for a long time. Great. Having you on the show. Anything else that you want to add? Where to find you, how can folks like tune into what you, whatever you do, and shout out other folks. Feel free

Canton, Varda on, Twitter or X whatever.

Awesome.

thank you for joining us. All right, folks. are moving on to, the, the, I think the, maybe the last piece here, of coding with ai. There's like two launches that we had this week, where, both, windsurf and Cursor, ton's not here with us. so we're not gonna cover the cursor part that he got excited about. But basically the cursor part is they launched an integrated web browser within, the cursor interface. And that web browser can basically, show everything for you.

I also saw that they have a feature where you ask

all the l LMS at the same time. They all give you their answers, and then you have an LM that condenses the answers from everyone combined. Oh, wow. They have a new mode for that. Oh, interesting.

Yeah.

but the other thing in the world of coating is that Windsurf, which is kind of the competitor, they're launching their own like different stuff. And they launched, code maps. And code maps are super, super cool. I wanna show you, let's see if I can get to a code map. basically they're saying that, understanding the code is as much part of, being very productive as running it. And so they're showing something like this. I'm gonna show you an example without understanding what my code is. It's gonna be super clear, but still it's gonna be super cool. let me show you the window here for windsurf, Alright, so, you can see here on the right, this is the agent and then, this kind of terminal. Okay? here you have a new feature called Code Maps. And when you open this, you will see this kind of interface, or you can switch to diagram view, and you can see the diagram of what happens

what exactly are these?

Okay, so, as, I have a bot.

That, let's add Nisten. 'cause it, it's very important. So my bot for tar video, the translation server that I still own and still runs, is basically streaming from Twitter and then receives or used to stream from Twitter when the a BI was available and then gets the pings and understands the video there, downloads the video, then starts translation, then posts something else, right? This is generally kind of the bot. So I asked Windsurf to go look at the code without me saying all this, that I told you without my understanding of the code to generate a code map that does this. so here's the code maps for this whole workflow that I've asked. All I did by the way is ask for like a i new code maps and I wrote com, like Twitter bot video translation. And then it went and looked for this thing. So it separated it to, several flows. One of them is Twitter streaming, and it shows you which task happens when. for example, for Twitter streaming, you have to set up rules. Then we have the filtering, then the trigger validation. and then after the streaming from Twitter happens, forward the backend basically a way for you to understand my code, right? And this is very important when you in a new code base, it's really hard to know what's going on. So you do code maps, and then we have the Twitter processing.

That's brilliant actually.

it happens. It's obstruct, man, obstruct and really cool. It happens

on every, every like, really like every, every,

report that, that we can have. and then you have my pipeline that it understand that I have queue tasks. this, by the way, happens with Al Workers here. I have a transcription generator, and then the, the, the captions thing, it, like, it literally reached out and figured out every little process. That's not easy to do. So I have another code map, for example, for my docker install. and then it shows me exactly what installs and how it installs Docker. And I think it's brilliant because like it explains to you, I have like a very specific, like docker set up for local and then for, GPU enabled machine, for example. I think it's brilliant. So shout out to the wind of commission folks as well.

I can show you the multi-agent.

Yeah. okay. So, there's a video from the cursor team. Just look at this. you go to your agent and you do what you want. You specify how many parallel agents you want, and now you have volleyball tubs for G pt, five, sonet 4.5, also a cursor composer. And they all generate their answers in parallel, and I have no idea how they combine 'em at the end, but, I guess we're going to find out. But now you have multi-processing pool of LLMs at your fingertips.

That's crazy.

That's pretty cool because then if all of them got to the same answer, you kind of like mm-hmm. But I don't, I don't see this yet. Or maybe this is an example or maybe I need like a s Pro that's

Cur cursor now, natively supports git work trees apparently.

So, that's from cursor

thank you, for this.

And, all right folks, I think the coding with ai, is over. I think, we have audio and video if Wolf. Are you ready? Do you wanna cover some of the Maya one and the vol tts super quick. and then we'll do n again.

you can open the link for the Maya.

Maybe we can play the audio back. So the thing about this is that you give it not just the text prompt, but you also describe the voice and it creates a voice according to your specification. you can even say the gender, the age, the kind of speaking, all of the things. And, so you don't have to select from a bunch of voices. You create a voice on the fly. the demo has been pretty impressed. There was a villain that was talking like a movie. It has over 20 motion tags, realtime streaming support, and the backbone is a three B Lama model for the understanding. So it runs on a single GP or even No.

Alright, let's take a listen to an example.

This is a description of this voice, female in her thirties with an American accent, and, is an event host, energetic and clear diction.

This place looks even better than I imagined.

The lights, the music, everything feels magical. I can't stop smiling right now after such a long week. This is exactly what I needed.

That was not the best example Blazer.

That is much, much better.

Welcome back to another episode of our podcast.

Today we are diving into an absolutely fascinating topic.

You dare.

That's pretty cool. And this is a tiny model and, it's open sourced, speech model built for excessive regeneration. yeah, this is great. Apache two license three build parameters, runs on a single GPU.

What's it called?

Maya

one, yes.

Maya one.

Alrighty.

well, it's great that it's open source. So Maya one, I'll definitely play around with this some more and I wonder if it does, do you know if it, offers cloning or No cloning for this one?

you don't clone, you specify it.

Describe voice.

yeah, voice designed to natural language descriptions.

Maya research. this is dope. absolutely dope. the playground does not seem to allow the same thing, like from what you said. And the other thing in voice and audio super quick is that, artificial analysis, the place where it tests, multiple models showed they have a new number one on there called inworld, TTS. Also, we have an example of this. Okay. So this is in world TTS. They claim the state of the art in voice ai. It's really easy to claim state of the art, but let's play. and they also have like emotional tags. Looks like angry, for example. do they have laughter, laugh. Let's play an example here and see and compare.

It's quite a long time to foolish mortal.

Your courage will be your downfall. Soon you will learn that no one escapes from here.

my left act didn't work, but this was really good.

Amazing.

Yeah.

this sounded like solid, I can see why this is like, it takes number one, it's multilingual. They have, Korean, Chinese, et cetera. 5 million characters, $5 per million characters. It's supposedly cheap as well. they compete with Cartesian 11 labs on, world error rate as well. And they tend to be blazing fast. Again, Cartesian is less than 50 milliseconds. This is less than 250. It's not open source, is it? No, it's not. It's not. it's a competitor to like 11 labs and, and, Cartia. It's a whole company. Alright, this is, this is the updates invoice as well. If you want the top one. It looks like, one of the new ones is, you know, in world TTSA new one to remember. Left Labs, cartia inward. TTS. Alright folks, I think it's time to, we talked about coding with agents. we only have like 10 minutes left on the show. So despite me feeling that there's not gonna be a lot to talk about, we somehow film the show. this week, as I said, my highlight of the week was, building things with AI for myself. I finally, finally got to a point where I built an N eight N workflow. what is N eight 10? N eight N It's spelled like just the world. N eight N I'll put it in comments. it's a workflow builder that really quickly adapted to the world of ai. It's a node based interface, which I'll show in a second, which allows you to add sequences one after another without knowing any code. And they have pre-baked things, for example. So we can show, I can show an example of something that I built and I definitely would love, folks here to also comment and say things. I have built a Thursday eye flow that will do something like very interesting.

a good comparison is to zaia, which is a big one.

Most people know of this, one version you can run locally. So it's partly open source is not fully open source as far as I know. But it's, the source code is available and it's a German company that is behind it.

Yeah.

And definitely Zapier is a competitor, but in the world of ai, I think OpenAI is kind of, agent Builder is also a competitor because OpenAI agent builder is the workflow builder. And for example, for the task yesterday, I didn't even consider OpenAI is agent builder. Like this to me was clear that NA 10 needs needs to this. So here is, a workflow that I've built and I hosted this on $5 a month on head center. And I can talk about hosting this as well. I first started doing this locally and then I moved in, put it into cloud, and took me like two hours for $5 a month. It's literally nothing. Here's a workflow that I built. It took me, around four hours, including learning, building the things, and like learning and coding a little bit as well. What it does is a very interesting thing for Thursday. I, we need to understand and tell you super quick, what's going on. And a lot of our sources are on Twitter and for example, Wolfram today sent me this link for, I think this is the tts, right? The Maya one. or another example is this morning Kimmi came out and I really needed all the right details. Super quick. This requires multiple steps. If you guys follow the show for a while, you may remember that the way I used to do this is with Agen ai, for example, come or agents, I give it the tweet, it logs into the tweet, it looks up all the links in the tweets or the thread, right? So it's not like sometimes people because of, you know, they put the links in another, reply, et cetera. So my agent needs to go and get that all information, then go independently into all of the links and try to figure out together from all of this context, build the context for the ai. And then an AI summarizes this and gives me, a few facts to read out to you on, on the show. So in my case, I have the notes that I used that you guys see in the TLDR section. So it needs to also place it there. many manual steps in there where I used to do it with, with, and I now automated this completely, with a trigger from Telegram. So here's what an end, end, end workflow kind of like looks like. At first you have triggers and you have to define which triggers you have. and you can define a very, a few triggers. they have a bunch of automations, Here I have a telegram trigger. Basically I built a telegram bot and it's building a telegram bot requires exactly two steps. You open telegram, you go to bot father in Telegram, and you say, create a new bot and it'll give you an API key. You're done. There's nothing to build there. It's like, it sounds more, complex than it is. It's literally just that, like building a telegram bot is going to the telegram bot father and saying, Hey, build me a new bot that gives you an API key, which you plug in and you're done. You basically have a way to trigger, things also, and then has this form submission, trigger. Basically they'll build a form for you that you define the fields for, and then every field will get passed into your workflow. So this is another entry into the workflow. In my case, this is a Twitter or conversation ad. And the last one is a chat message. this is very similar to the way OpenAI API works, You have this, chat interface here. And this chat interface, whatever you input here, will go to the next node as input. So you can build the chat and then you can reply to the chat and get answers, and you can build your own chat with ai super quick as well. So I have these three triggers here. And what I did, I unified all of them with some code just to make sure that I can start my workflows from any of these three. Basically, every other node will give you a different name after this. There's a bunch of stuff to do processing for. here is basically the workflow, the way I built it, as an example, because it requires Twitter. And so I have two notes here and, it's just a HTP request node. It's a basic request note when you go, you don't have to code for this. because you can just call any API, with the credentials. The credential management here is pretty cool. you have credential management. So I'm not risking showing my credentials to you because they have credential management built into NA 10. You can decide which API keys you have, and then you can use them in workflows, but you're not putting the actual credential. You just like pointing to the credential. Why this matters, this matters is because I can share this workflow with Wolfram, for example. Wolfram can run this and then use his own credentials. So I have the credentials from Twitter and here's what I do. Basically, I get the first tweet, with the regular, Twitter API, how to wire those nodes up is that on the left you can see all of the kind of node executions that happened before this node. Okay, so let me show you, a trigger from Telegram, for example. So let me go here. Executions. So here is a, is a workload that succeeded. For example, some of them fail. and then I will, copy it to editor. I'll show you. So here is a, a workflow of mine that started with the. telegram node. Telegram trigger. So when you click this, you can double click this. Even you can see, that you got some events from Telegram and this was the event and a bunch of things at this event. Well, what you need to know is that there's a URL here, inside the telegram bot, the IT points to Twitter. What we need to get is the Twitter id. you go to the second node and you kind like unify this. Okay? then my code that runs and gets the Twitter ID runs here and you can see, from the URL, which extract only the id. The ID is used for an API. So we take this id, and on the left side of the node you can see all of the steps proceeding to it, what the input was and what the output was. This is very important because you can take the output from a previous note from before or the output from the beginning of the flow. It's very, very useful. And then you can basically use this tweet ID inside the API here, like so, to pass it forward. I think this is the connection point. This is the most interesting part about NA 10 and how you connect nodes. You take an output from the previous node, and then you use this somehow in the output of the next node. In this case, I go to the Twitter API, and I need this Twitter ID as a kind of, collectible. And you can see the result of this expression here. So Twitter, API will go to tweets and then the Twitter ID that I passed via the Twitter API, And then all of the other definitions for API are standard. Basically, I have some other stuff like sorting and cleaning stuff. I kind of had to get to learning how to do this because, of my specific workflow. But I wanna get to a point where I have an AI agent. So after I collected all my things, I now have something that kind of looks like this, I have, I need to execute the previous things. Okay. I have the first tweet and now I wanna send it to my AI agent. it's very simple to create an AI agent. In NA 10, they will receive a prompt for you and respond in some response. you create an AI agent node like this, and then you connect to this node. Any AI agent AI that you want. And in this case, this is an OpenAI. I can have cloud on Anthropic, we can have Kimi, It builds an agent that does something with GI and see how that performs. actually maybe worth testing it out because they also have an open router, extension. Do you guys wanna do this super quick as an actual thing to try and connect the Kimi K2?

test of this new model?

Yeah.

Okay. So, Kimi K2 is probably gonna be slow, but we can test it out, So, it was pretty good.

you an IP if you want.

Yeah, if you can send me something that I'll not show on, on

screen, directly to Kim or open route. I think I have, I can get open routed

Steve.

I think it's, worth, just mentioning, how do you set all of this up? Because you started telling us, but like, we don't know how to set it up. okay. So we know how to start.

starting is easy enough.

If you know what Docker is, if you don't know what Docker is, it's, it's easy ish. You go to NNA 10, NA 10, documentation. But basically where to start. you can host NA 10 on your computer or you can go to NA 10 cloud register and start there. That's 20 bucks a month if you want for free, you can host it on your computer. It takes a little bit more. The thing about this, and I really wanted to highlight before we get to implement some of the stuff is that, I use NA 10 with the Atlas browser with OpenAI and the Atlas browser was open all this time here and I. Was so delighted about how quickly I was able to learn what to do with the browser looking at my workflow because I could ask it like, hey, especially if you have voice on how can I replace the summarizer g PT five node with an open router node in NA 10 N Atlas browser has the context.

sick, man.

Yes.

the browser itself realizes what I'm doing and it knows an eight 10 because of like examples and I use the fast GPT and you want to swap out the summarizer with your open router step by step. this is literally the future. I will not be able to build and do this whole intent workflow without the browser just telling me what to do.

actually the future will be when it does it,

I'm pretty sure that the agent can do it.

this time.

Yeah.

I think there's an agent here and I'm pretty sure that the agent can do it.

cloud code can do this through the config for sure.

Cut thing, everything.

the whole point of cloud code is like, it is writing code.

Okay, folks, Nisten gave me an API key. And, so we have, we wanna replace this guy, our AI agent that summarizes in the end with the Kimi Moonshot to test it out, right? so I can, I can delete this connection. So now I have this like free-floating AI agent and I need to define the new agent. Let's see how this is done. we can do open route chat model, OpenAI chat model for advance user AI chain.

yeah, either open router.

Yeah, I'll just

Nisten thank you.

But I'll go through open router just for the, the ease use of this. Okay. we go open router, select credential. Nathan doesn't have any like, open router credentials in my case. So I can create a new credential. I will hide this window super quick from you guys, so you'll not see my open router credential. And I will get a, a key here that I supposedly works. I say this, it's a quick

If you have a home assistant, it's also a plugin

for home assistant and add-on, you can put in home assistant.

And then it can run completely.

I actually need to set up on my home assistant.

What is Home Assistant?

Is that on Apple

or

Yeah. It's a home automation alternative to Google Home or, Alexa, for instance. So you can, set it up, it's open source software, then you can do the home automation with that. It has connection protocols to all of the devices.

yeah.

Open source home automation system. okay. So it's so much small by the way. It's, it's incredible.

media as well.

Yes.

you guys see I connected my API to open router, and now I have all the models including communic key thinking, like literally took me a second paste an API key, and then they do all of the, all of the work behind the scenes. So I now have an opener chat model. I just, I need to name this, like here, I need to name this guy. You don't have to rename, but like Kimmi K two thinking. Okay. Rename. And so now instead of this GPT five mini, I have communicated thinking and I can execute this note from this place. let's take a look at how this looks execute. no, but I won't. Okay. Just play. So you can see now this agent runs and then it goes to communicate thinking. I believe if you double click, you can actually see, output here from mto. Now, MTO thinking, it depends on whether or not they enabled the thinking, Traces back in the UI or not. This may take some time for GPT five, for example. It doesn't show me anything 'cause GPT five thinking it takes a while. this step does take a while. Kim two is done. And now it calls the MCP client. remember how I told you that every tweet has a bunch of links? I have connected an MCP tool here, that uses, Gina ai. If you guys remember Gina, we talked about Gina. Gina allows you to do something like Gina Reader. you basically give it to URL and it presents the URL back to you in markdown. so here's an example of Gina. You go to stack overflow this question and then you do fetch content. Gina will give your AI a markdown format of that webpage so that you don't have to use HTML, for example. As you can see, it's ugly. And Gina gives you this like title URL source markdown content. It's really nice for LLM contacts. It reduces the tokens that you need. So I connected Gina as an MCP and I basically told my ai, Hey, if you have any amount of URLs in those tweets, use the generator, go get their text, and then give me an output. So it looks like my, communicative thinking, does know how to use these tools as, they promised it, would. I wish I could see the full output and show you kind of the full output, but it, only shows you at the end of the node what actually happens. But here we did see that, we got the output of this webpage from this URL, the Ha and face collections by dense aura, which is the example that we're showing. And now we're gonna have to wait for Kimi. I just wanted to show what happens at the end of the workflow as well. What happens at the end of the workflow is that you have something, and that's something needs, the state. You need to put it somewhere. So I have kind of two outcomes. First of all, I have a channel, which all of you are welcome to also join. there's a channel announcement channel. this is the discussion channel? This, there's announcement channel for Thursday and news updates. Everything that I will use will get posted here as well. and this channel is basically the output where I send all of the summaries. sometimes there's no picture. So you use Thursday? here's an example of this thing we're working on by dance. This is Aura Looped language models. This is the link that Wolfram sent me in the beginning. the beginning of my workflow starts with Telegram. I send the link and the end of my workflow lands in the channel also via this API, and the links to HuggingFace and everything is here as well. This is the problem of showing AI agents, especially the ones like Kimi, they take a while to load, but you saw how easy it was to connect a new one. very similar.

You may not even have to use the thinking version for many

workloads because, I used the normal chemi with MCP and it was faster than the thinking when it gave the same results. Yeah. So depends.

I probably don't need to use the kimmi.

And I think I have an issue here with some Jason that, that it responded with. but now we can see kind of the process. So, this was Kimi MCP client gave me some, more context and then, it took Kimi quite a while. All of this to think how much is this, in It took 1 68 seconds, almost three minutes to think, through this. And then at the end, it didn't give me the, the right output, as I asked for it. So JG PT is really like listening to this, but you can see the error. the output is, I'll help you prepare the notes for Thursday. I podcast about Biden's aura. Let me gather more detailed information from you having his small pages. I think it's because of the, the thinking. Maybe open order. Gimme the thinking first. Ah, I think it

maybe the overloading because I had so many

interrupted tool code as well.

but if we delete the QI one and then we get, GPT one

back, as you can see, I can try out, test out new models super quick. this will play out much faster. And we'll see how the rest of the workflow goes.

Remember what, Skylar meow, so last week that the prompt format

actually matters a lot for the, the API. So it might also just be that, but it might also just not be good with this task too. It's, yeah, it was pretty good for me.

we have some comments from, Joseph that says this, should

be turned into a vertical flow. it's probably easier. but it depends on, I'm assuming Joseph, you're talking about just the ui, folks are commenting and saying, anytime we also implement a text to workflow feature that creates an workflow using natural language. That's great. I could absolutely see that happening. but for now, again, this essentially is not that far away. If you just use, Chat GPT atlas and you ask, what to do and this is great because, yeah. Okay, so GPT five execut are node, successfully and now we can see that on the output of this, you can see the output fully Exactly based on how I require it, the Jason I require, et cetera. you can see the 10 factoids model families or available sizes, 1.4 and 2.6 parameter. I asked the AI model to extract based on all the links and basically, the process to very simply summarize input from somewhere, then preparing this input to get more information, building the context for Theis. So I have one AI agent that builds the context, and then I have another agent that actually does the thinking, the summarization, and then the output is sending it somewhere. In my case, there's the telegram channel that I told you, you can join and also reflect append so I can show you how this works. Reflect is my notes, taker and I have an append system and it's done. And then if we go to reflect. and I go to my daily notes here. We can see, it added it here, by released or looped language models. And you can see the executive summary intent factoids. This has automated so much work that I did manually, It's a little bit slower than using a browser, but it's definitely worth, checking out. hosting N 10 is possible, n at 10, the cloud, which is $20 a month, or on Hener, which I chose. It took a little while to set up, but again, with Atlas it's really easy to just have it tell you exactly how to do the domain, the DNS, the HTPS certificate, et cetera. and once you set it up, you can automate a bunch of stuff, which I plan to do as well in the future. So from coding with agents to non-coding and getting agents on Thursday this week, hopefully this will help. This was helpful for some folks. Definitely. I feel like I've unlocked a new superpower in automation, and as Wolfram said, if you're running home assistant, you can run your workflows on there. think with that folks, we've managed to go through a whole show, even a little bit longer than we expected to. for November 8th, who can believe it's already in November. we are expecting a lot more interesting news for the end of November. a reminder, if you are coming to the ny, an AI engineer in New York, I'll be there. we're gonna have a show from the floor of an AI engineer, as well probably gonna work on getting some great guests in there. If you're coming to Ai.engineer code, please let me know. I'll be glad to meet you there. If you missed any part of the show this week, the show is getting turned into a newsletter and a podcast on a weekly basis called Thursday News. Please subscribe to Thursday News. If you haven't done already, if you hit this QR code here, you should be very easily be able to get redirected to subscribe. I wanna thank my co-host here, Nisten Ya Wolf, Wolfram, and, Kenton Varda from Lauter who joined us before to talk about MCP as well. I look forward to seeing you all here next week, which should be a very, very exciting week. shout out to everybody who did the releases, hugging face folks, montra folks. OpenAI folks. Cheers everyone. we'll see you here next week. Bye-bye.