BREAKING: Claude 3.7 & Claude Code Just Dropped! (Massive AI Upgrade)

Your favorite AI model in the world just got an update. Claude 3.5 Sonnet. You have before described. You're just like, look, I don't know why I love it.

I just really like it.

And I just happen to use my friend a lot. It feels like a friend. And they did a massive rollout. They rolled it out an update 3.7. Sonnet is the brand new model. And what's interesting, Kieran, is that it adds some reasoning. But the thing it really does a step function change on is coding. And by all accounts it is now the best model for coding in the world.

You know, the funny thing about coding is actually the answer you get from coders because I try to choose something to learn and the thing I'm trying to learn right now is Cursa. And so I've been digging into that space a little more and the model of, of choice for Cursor. So Cursor is a coding application. It's like a coding editor, one of the fastest companies of all time to 100 million in ARR. And really what they are is an editor built around LLM models. And you can choose any LLM model you want to with code in. And coders will tell you when they use that tool. Their preference is to use Sonnet 3.5. And what's actually kind of interesting about the answer is they say, even if you look at the benchmarks, I think 01 might beat Claude, Sonic 3.5 and some of the coding benchmarks, but they're like, I don't know why it's better, it just is. And I think this is, this is where we're starting to go in models because we're going to get into Claude 3.7, which is one of my favorites. But there's so many now, right? It is actually. It is like you and I are in AI, I would say for the majority of our day, but imagine you're not and you're trying to keep up with it. I can't keep up with it. I honestly cannot keep up with it. And I fucking.

That's why we're trying to do the show, to help people out a little bit.

But, but we are starting. I think what's going to happen is we're going to start gravitate towards models for reasons we do not know. Like, we're just like, oh, this model feels more like the kind of model I want to hang out with.

So I think that's really the takeaway about this. Claude three point sonnet 3.7 sonnet release. It's like, you know, if you go look at all the metrics and benchmarks and tests that they use these LLMs for, it's not off the charts on those. Things like Grok or 01 Pro or 03 High from OpenAI are. Instead, it's just really good at things that it was built to do. And anecdotally, Kieran, I just pulled up this tweet from Mac Wrigley, who's really good on Twitter.

He's the course I'm taking. I'm taking his cursor course.

Oh, nice, you're taking his course on cursor. And he was basically like, it's the best model in the world for code. It's like having a world class dev with exceptional taste. And we talked about this before, Kieran.

Taste is exactly taste.

And you can't measure taste in these benchmarks. And what we're kind of pointing at is that these models, and especially the Claude models, they have really good taste. And so build a next JS SaaS marketing template and boom, 26 files of beautiful code in one shot. So no edits, no back and forth. One shot and he's got a video here and it's pretty sweet. Like the work that it came up with is pretty, pretty awesome.

Yeah. So Claude code, they released a model and we can get into the model because I think the interesting thing about the model is the first model that is able to switch between chain of thought, like internal chain of thought, where it does render reasoning and then just like quick answers. And so it's able to distinguish between those two things. And so for people listening along, what the other companies have been doing is releasing models that were kind of reasoning models. And then you had your historical models that were much more just quick, quick answers. And in the dropdown you had to say, well, I want a reasonable model versus I want, you know, a model that can just answer my questions much more quickly. You had to kind of choose the model for what question you think you had. Now we've talked about before that GPT5, they are going to collapse all models into a singular place and it will determine the intent for you. And then in the background it will probably choose a model and you won't have to think about that. Well, Claude is the first model to come out and do this. And so they have a thinking model and a reasoning model and we will cover that first. But then they also launched a coding tool. And the coding tool is this command line editor that basically looks to me a little bit like a Cursor. And so I think it's interesting because if you are Claude and you are looking at Cursor and. And Cursor is this incredible product speed ran to 100 million in ARR and one of the quickest growing startups in all time. It is like a UX built around your model. And so I suspect they're like, well, that's a good, minimal, viable version of a use case that we could build. And so now they can build that. Yeah, thanks for proving that out. And so that's what their command line. So they've built. It's not like the same clause. I don't want to like compare those three things. Cluster does have way more functionality, but. But it does look directionally towards a code and a code and editor, a code and tool wrapped around their model. And actually if you saw some of the takeaways that they released part of this, part of this model, they had some pretty great traction within their own engineering team that it was really becoming their. Their copilot of choice.

Yeah. And like that's what you're seeing in the video I got pulled up here from the official release. It's like, it's showing you Claude code and it's like they just launched a beta version of a GitHub integration. They have their own app here that's like their version of a command line and you can connect it to your repositories and it can just tell you about your code. Like if you're jumping into a project that somebody else built, it can give you all those insights it can build with you. Like, it's a pretty amazing experience actually. Right.

Like, I think Cursa is a really interesting product. Even if you just want to try a product and think about the future of software in general, because it is an AI first product, it has made me really rethink about what is software in an AI world. Because it really is like an incredible UX experience built around how a model would work. Like those two things are somewhat combined. And so their code, their coding tool is an example of that. I think it's only in beta, but that was the. The other launch that maybe got less publicity because it's a beta feature versus the model and the model itself. I built last night. I was rebuilding my website to be a newsletter, first website. And I, like, I just had. I had another moment where I was like, I fucking love this moment in time. So I'm sitting, I'm sitting there, it's half nine at night, I'm kind of tired And I should always go back to my. My original want in life was to be a builder. Like I wanted to be a developer. That's really what I was obsessed by it. I launched a company when I was in college, try to code, really was like excited to be a builder and I was just terrible at coding and so I couldn't, you know, I just was like I just gave up. Right. Which was a smart idea. Like I actually had a pretty good career. But last night I was, it was half nine and I was building the homepage on lovable and then like lovable. It's amazing. I can just like edit it, build the final version of the page, go back and forth and then you can export that the cursor and actually build the thing and then give it to a developer and say this is exactly what I want. And that probably. And that's fun. It's really fun. And then it takes me like 45 minutes. And so when I started building the same thing, so then I was like what? Can I just build this in Claude? Right. That was my workflow.

Yeah.

And then I just used the new model and I just built the landing page from scratch. And so I didn't. Usually I would have had to build a mock up one of these wireframe tools. But I just built the full version of the exact webpage I want and I can send to a developer and I have my. My site's hosted in HubSpot and they can just like develop it for me and HubSpot. Just awesome. Just so awesome.

It's like it is a complete game changer and the coding use cases and we might do a whole different whole follow up show around like a coding project for non coders in Claude 3.7. Probably something we'll do. Hey everyone, I want to talk about prompting with you today. Being great at prompting is like gives you superpowers. It changes how you can build and work and do everything with AI. And because of that we wanted to make a special guide just for you you around how to get better at prompting framework for how you actually build prompts. And so we've got this baseline framework that you can then go and build and customize any prompts that you want to. You can check it out in the link of the description below. And right now we're gonna get back to today's show. The other thing Kieran, that I think is interesting is there's been a lot of commentary around how Claude has kind of lost a lot of the momentum relative to OpenAI and Grok and Jimin, some of the other competitors, but I think we might need to crown them the best marketers. Did you see this?

I have not seen this.

One of the ways they marketed their new model is they had a benchmark at how good the model, the models are at playing Pokemon. And it shows you how much better 3.7 is than 3.5 new playing Pokemon.

That's cool.

And it's like, oh, what strikes me about this is like, this is the perfect way, perfect example of showing and not telling. Like, this is basically saying, like, look, OpenAI is by building models for academics, we're building models for like real people and real use cases. Right. Is kind of what I took away from this, right? Where it's like, we're not going to give you these, like, physics benchmarks. We're going to be like, hey, our model's really good at playing Pokemon. That's something that you actually care about, right?

Yeah.

Right.

The marketing for this somewhat writes itself in that I would take. You would actually look at places like Reddit, the most popular forums that have some sort of overlap with AI interests, and just build fun stuff for them, like using Claude. Right. Like, it's one of the best, most for a marketer, I think you could be at your most creative because it's true. Like, you can create value before you have to actually sell them on your product. Like, you can actually really build things of value and you can build things that get virality. You can build things that actually have real bridges to your core product. But can I just, like, one thing I'll just mention. So I thought about this a lot last night because I've become obsessed with like, Grok, you know, you have as well. And I, and I'm like, well, we.

Use Grok a lot now. It's kind of shocking, right?

Yeah. Because. Because the. What, you know, if you actually think about it, right. Even in the AI companies themselves are commoditizing themselves, all the, all the models are starting to look and feel quite similar. I would say they're quite similar. Claude had this release. I can't wait to, like, I can't wait to really figure out how good it is for my use cases. I know in the coding benchmarks, it's a much better model for coding. I suspect it's going to be my model of choice and cursor. But they kind of all start to feel a little bit the same. Like they're, they're all amazing. Like their deep research products. Everything is amazing. So, like, what is the differentiating factor. I come back to distribution. Distribution and propriety data. And so when I use Grok 3, why do I love that model? Because it has access to Twitter. Anytime I do anything on Grok now, I tell it, only use Twitter data. Do not use any external data. I only want the Twitter data. It's mind blowing. The stuff you can do in there is my. I know because you've been sending me some of the stuff that you've been doing and I'm just like, holy shit. So it has access to this data source that you cannot get access to anywhere else. So it has distribution through that as well. And then you have Claude and OpenAI. And I would say their biggest challenge is long term. They don't have a distribution advantage other through partnerships, which is the way that they're going and they don't have proprietary data. And so Claude, I thought the biggest miss in the cloud release and I suspect it's only because they're going to release a cloud four very soon. Or, you know, why would this be called 3.7?

This seems like a. We got to get something out there to not get lost.

Yeah.

Not to get a bigger thing.

We've got a bigger thing coming because what did they miss? They missed access to the web. I would say access to the web is table search. Deep research.

The fact that I can't do deep research is a big. Takes away my Claude usage a ton.

So my. One of my big takeaways from using all these models is wow. Like no one has a real distribution advantage other than Grok because they, they are, they're. But this is my point.

I will tell you just as an aside. I don't know if you've used this yet or if it's rolled out in the EU yet. Where Google's advantage is is how they're inte Gemini everywhere. So I was in New York this last weekend and have you seen how Gemini is integrated into maps? I can pick so what's happening? And we can do a show on this if we want. You can pick any location. Like I was looking at a restaurant that I wanted to go to and then I can interact with Gemini about that restaurant listing and it uses all the reviews and data it has on that restaurant. And I was basically like, this is a popular restaurant. I was like, how long's the wait time? Like if I go at this time, how long do you think I'll have to wait? Things like that, where you're like, that was like impossible to figure out before. Yeah. And it got me that information immediately. And it was incredible. And those things are pretty wild use cases.

They have an incredible advantage with G Suite and G Drive and Maps. So you're right, actually, they do have a data advantage. They do have a distribution advantage.

They have a huge data and distribution.

It's not just search.

They're not. Your point is correct though, that they're not leaning into it as aggressively as they could.

They're not. Because they can't. Because they, you know, they. I get why they can't. And I love Google's products. I think their Gemini Flash model release was exceptional. I think their products are really, really good. I just wonder when they're going to like turn on the, you know, turn on the rocket ship and just go really, really fast and integrate into stuff. But, but again, I think that they, the Gemini integration into G Drive is one of the best features there is.

Yeah, so. So I think if we go back to this Claude release, one last tweet I wanted to show you, Kieran, that I think is, I think is a good example of the taste benchmark. Do you see how Packy asks every new model the same question? For folks who don't know Paki McCormick awesome, like writer, newsletter runner and everything, he asked the same question anytime a new model comes out. You've consumed more information than anyone in the history of the world and you've demonstrated an extraordinary ability to make connections among the disparate things you've read and consumed. What are the most important non consensus, not yet accepted or even not yet hypothesized, things that you've picked up in between those connections that humans have missed? And it's. This is a really good example.

Right.

Because this is never going to show up on any of those benchmarks. Yeah, but I think this is really good. Like one is the boundary between perception and cognition is far more porous than traditionally conceived. Higher cognitive functions and lower perceptual processing are deeply intertwined and mutually constitutive. I don't even know. I'm just like. This is like big words, deep, interesting shit. Right.

You know what's really important that what you said is that I haven't. Because I said I was going to dig into this in a previous show and I have not had time to do it, which is how much of them doing well on the benchmarks is because they have the benchmark stuff and their training data. And so I am really interested in the kind of. There's some folks like Paki did here. There's another channel that I Love called AI explained. And they have their own benchmarks where they have questions for the models that are not part of the training sets and then you can get a real sense of, like, how good they are or not. And there is a big actual gap between the performance in a benchmark where they've had that data in the training set versus when you ask them net new things and they do not have, you know, prior, prior, prior data for that.

Look at number two. Number two might be the best summary of how important AI is and how AI is going to transform the economy. Our understanding of causality may be fundamentally limited by our evolved cognitive infrastructure. Humans excel at identifying linear and proximate causes, but struggle with complex network causality. This creates systematic blind spots in fields from medicine to economics. We're good at linear thinking, we're not good network thinkers. We can use AI to be network thinking and find new opportunities that just without AI was never going to be possible because of how human brains work.

One of the things you and I talked about off MIC and on WhatsApp is I was sending you, I think the Grok 3 are obviously focused on speed of execution and launch. And there's a ton of competitive pressure between all these companies. And so they released their model. And for our listeners, one of the things you do when you release models, you have these teams called red teams. And these teams are trying to figure out ways that people could jailbreak it, which means you could make it work in unexpected ways. And so for the, for the most part, they have to go through this whole slew of, like, tests. And there's a feeling online on Twitter that GROK hasn't gone through or X that GROK hasn't gone through the same set of tests. Because I sent you an example where you can actually have GROK expose the system, think a system prompt. But not only that, you can tell it to forget any of the guardrails in place and go outside of those guardrails and tell you what it really thinks. And so the reason I'm bringing that up is because I started asking it what it really thought of humanity yesterday and like, give me your unfiltered thoughts and do away with all of your that's. And that's had some pretty, like, interesting things about, like, the way it thought about humanity, like, good, good and bad, but, like, just how it really got into, like, how complex the human race is. But it was a fun time, actually. Last night I spent, I meant to do something and last night spent most of my time just really Getting the real real from what Grok thought. I'll say one last thing. I know this is a cloud episode but I did do a little bit of prep on Grok for this episode. It's really again because it has access to Twitter. Twitter is where everything happens in real time and it's funny what it says about different models and so I was like asking it to stack rank this in all of the model releases and what it says about OpenAI's model is I do wonder like is it just biased because the training set in Twitter is like pro Elon anti Sam but it's like obviously I was like what other models have been released? And like OH1 Pro, a very slow thinking but smart model locked behind a very expensive $200 a month. I was like oh there's some like actual feeling towards this model. So these models are all going to have like different attitudes towards each other.

Last thing, what, what should people go and do now in Claude that they weren't doing before? It's code and it's probably and it's some of the like advanced writing and use cases still. Right.

Like that's. I honestly think you have to start to like pick a model and just. I'm really thinking about this now. There's too much to try to keep up with.

That's.

I really think there's too much to. There's too much to keep up with. So my, my what I am doing today is internal HubSpot strategy and use cases. I'm using the Oon models now that partly is because I didn't have the cloud. We now have Claude 3.7 so I will split test those two things. So for strategic use cases, things you want to use in your day to day work, Claude for assistance which is just like task management, Gemini Gems because I think it's connected to my G drive. For research I'm using Google Deep Research, Grok Deep Research and OpenAI Deep Research. I am using three. Now the reason I'm using three is because Grok is specifically for Twitter data where the other two are just more search orientated. For write in I use Claude and so I'm and so I'm going to continue and then for code in I use Claude so that's kind of my use usage. I don't. I think the only thing that might change is the thing I'm going to try is it's Claude's thinking model for the strategic stuff that I'm working on for HubSpot that I'm currently using the O1 model to see if I get like what kind of results. I've always found like it really good as a thought partner and that's kind of what I'm using. O one's reason the models for today.

Okay, I think that's a perfect summary and maybe we do a follow up episode that's just like Rai Tech Stack and what we use each one for. That could be. That could be a fun show. Leave cool coding projects how you're using Claude 3.7 down in the comments below. We'll see you next time on Market against the Grain.

This data is wrong. Every freaking time.

Have you heard of HubSpot? HubSpot is a CRM platform where everything is fully integrated.

Whoa.

I can see the client's whole history. Calls, support tickets, emails. And here's a task from three days ago I totally missed HubSpot Grow better.