applications. I tested this by asking it to analyze salary data and create visualizations. Instead of just giving me charts, it wrote Python code, ran statistical tests like ANOVA and T tests, created interactive violin plots and bar charts, calculated correlation coefficients, and even generated a full HTML dashboard with filters and hover effects that I could share with clients...completes tasks. It thinks through problems step by step and actually gets stuff done. And here's what makes this even more crazy. You can integrate this with Claude code right now. That means you get Claude's nice interface with Kim K2's insane coding power. It's like putting a Ferrari engine in a Tesla
Поиск по транскрипциям
clean them up so you get a sense of what to do in your own code base. That will be available for free in a section on my Master Claude Code class. And if you want to apply these ideas to really large code bases of hundreds of thousands or millions of lines, then I cover how to do that...Context Engineering class right over here. A bunch of teams have applied this to code bases consisting of hundreds of thousands of lines and have seen better performance with Claude Code. So this is perhaps the most advanced class on context engineering that you will find on the internet
start off by saying that Claude Opus 4. 6 is my daily driver model. I use a lot of tokens in this every day, both in Claude code, but for also other uses and stuff like that. And so, for a long time, I've been very partial to the actual Claude models. I like the tone of them...goal here is to build a model that's reasonably cheap for co-work. So, Anthropic introduced Claude co-work which they sort of describe as Claude code for the rest of your work. And this kind of plays out here when we look at the sort of knowledge work benchmarks that we're sort of seeing of office tasks
Claude, then it can be a bit annoying, but let's try this out. Then we're just going to log in. We'll shut down Retrieve AI and Claude code is now set up in the terminal. Hit enter to continue. And then we'll just keep going from here. And you can see that can basically run from...actually works. So, obviously the market is closed. That price is way out to be fair. I just want to make sure sec. Yeah, it doesn't look like Claude code can use computer use either. As you can see right here, we'll terminate that. Now, another
catching up to do. It's just been so long. And if you look at the context windows and everything else. So, we've got Claude Opus 4. 11 coding over here. And we have Deepseek V3. 1, the new version coding over here. And then we'll compare it to the old version as well. By the way, whilst...easy. So if we come back to Claude now, you can see here that we have the output from Claude way faster than DeepS. So Claude is a lot faster for coding and the output looks pretty good. So let's test this out. Maybe not actually. Look at that. It's breaking straight away
that Claude-4. 0-Sonnet performs pretty badly when it comes to web searching. And I found that to be the case myself when I get Claude Code to do any web searching through the web tool that it has available or to use a Tavily MCP server to find information online. Then it performs pretty badly compared to when...high setting and also like web search enabled to find information online or documentation or find an implementation for something. And then I pass that information back to Claude Code because GPT-5 is just better at web searching when you have the right settings. And you can use something like RefTools, which provides context. It's like context
right, so a couple hours ago, Claude Code released their biggest update yet, which is 2. 0. And this includes all these updates over here, and we're going to be going over them and talking about what's changed, and also talking a little bit about Claude Sonnet 4. 5, which is now the best in the world coding...model. And you can see that right over here. But let's focus on Claude Code
just basically messing everything up. This was over the course of like six seven hours of Vibe coding and I found its planning abilities and planning mode in Claude Code to be on par with Opus 4. 1 because usually they'd recommend using Opus 4. 1 with planning but I think Sonnet 4. 5 like just matches Opus...that level which may make Claude Code cheaper as well because it probably means that you don't have to use Opus 4. 1 nearly as much but it did still have some blind spots like one of the blind spots that I had over here is that every time you press sync, it likes to sync the entire local
tell Claude, "Make me a simple to-do app where I can add tasks, mark them done, and list all my tasks. Watch what happens. " Claude doesn't just give me code. It builds the entire thing. Front end, back end, working buttons, real functionality. You can add a task, type it in, hit enter, it shows up on your...tried, what worked and what didn't. You don't have to reexlain everything every 5 minutes. Fourth, better reasoning across domains. This isn't just a coding tool. Claude Sonet 4. 5 is better at math, better at logic, better at financial analysis, legal work, medical research. Anthropic says this is the most aligned frontier model they've ever
what that'll do is actually take the project files from the whole project that anti-gravity's created, but using Claude, which is like a superior AI, particularly Claude code. Then it will look through the Claude codebase, analyze what to improve, and say, "Okay, do you want to improve this? Yes or no? " So you can see here...submit a responsive mobile app to the app store. And basically, what you're going to do is use something called capacitor, right? And this takes the code that you have, so the HTML, the CSS, the JS, and it packages it into an iOS project that can be opened in Xcode. And when it's packaged into that, then
what that'll do is actually take the project files from the whole project that anti-gravity's created, but using Claude, which is like a superior AI, particularly Claude code. Then it will look through the Claude codebase, analyze what to improve, and say, "Okay, do you want to improve this? Yes or no. " So you can see here
deal. I don't know if I'm bullish on any AI vibe coding tool that's not built by the model makers. So, basically, Codeex and Claude Code. I don't know if I'm bullish on any vibe coding tools outside those two. The issue is the pricing just doesn't make sense. The pricing will never make...people that work there. Great, great. Right. The issue is I don't know if the economics make sense. These companies are already losing so much money like uh Claude Code and Codeex. They're already losing money on these apps because it's so expensive. all this inference, right? And then you got these companies like Lovable and Replet
lead on real software engineering tasks. They're 65% less likely to use shortcuts or loopholes, and Opus 4 has dramatically improved memory capabilities. But the real gamecher is Claude Code. It's a command line tool that lets developers delegate coding tasks directly from their terminal and it's getting integrated into VS Code and Jet Brains. That means...Claude isn't just helping with scheduling and emails anymore. It's actually writing code, real production level code. And if you think that's impressive, wait until you hear about the browser integration that's coming. Claude for Chrome is already in testing
zero solo victories and one co-ind. One out of seven strong performances. But here's what really matters. The insights. Claude for Sonnet isn't just good at content creation. It's exceptional at complex coding tasks. Game development, interactive applications, advanced programming challenges. When you need something that actually works, Claude delivers. Gro 4 and Gemini...doing content marketing, this is your AI. The quality is consistently excellent. The engagement is superior. The readability is perfect. But that's not all. Claude also dominates complex coding tasks, game development, interactive applications, advanced programming. When you need code that actually works, Claude is your AI. Grock 4 is the analysis beast. Competitive research, strategic planning, comprehensive reporting
zero solo victories and one co- win. One out of seven strong performances. But here's what really matters. The insights. Claude for Sonnet isn't just good at content creation. It's exceptional at complex coding tasks. Game development, interactive applications, advanced programming challenges. When you need something that actually works, Claude delivers. Grock 4 and Gemini...doing content marketing, this is your AI. The quality is consistently excellent. The engagement is superior. The readability is perfect. But that's not all. Claude also dominates complex coding tasks, game development, interactive applications, advanced programming. When you need code that actually works, Claude is your AI. Grock 4 is the analysis beast. Competitive research, strategic planning, comprehensive reporting
code. Um, I feel like there's tons of things like that could exist. And maybe like the simplest example is like claude artifacts where uh Claude can help you write, you know, code that gets hosted right there um in Claude AI to display. And of course that is safe because it is very limited. There...tests to make sure that you know if I agree with the tests and the tests pass then I feel pretty good about the code. Um that works best if you can encourage Claude to write sort of very minimalist endto-end tests. Thank you for the very fascinating talk. Um, I also appreciate that
literally take a screenshot of any NAN workflow you find online, any automation you see, any template you want to copy, paste that screenshot into Claude and boom, you get perfect JSON code that you can copy and paste directly into N8N. The entire automation gets created instantly. Every node, every connection, even the sticky notes. I'm going...replicate, just a simple automation with a few notes. I pasted this into my custom Claude project, asked it to recreate this workflow, and look what happened. Claude gave me perfect JSON code. I copied this code, went to N8N, pasted it in, and the entire workflow appeared exactly as it was in the screenshot. Same nodes, same connections, even
literally take a screenshot of any NAN workflow you find online, any automation you see, any template you want to copy, paste that screenshot into Claude and boom, you get perfect JSON code that you can copy and paste directly into N8N. The entire automation gets created instantly. Every node, every connection, even the sticky notes. I'm going...replicate, just a simple automation with a few notes. I pasted this into my custom Claude project, asked it to recreate this workflow, and look what happened. Claude gave me perfect JSON code. I copied this code, went to N8N, pasted it in, and the entire workflow appeared exactly as it was in the screenshot. Same nodes, same connections, even
comparison. Claude code versus open code plus anti-gravity editor freedom. Clawed code says no. Open code says yes. Subscription reuse. Vendor lockin clawed code is high. Open code is none. IDE experience. Clawed code is limited. Open code is full. Risk of access loss. Clawed code is high. Open code is low. The pattern is clear. One option gives
over the last few weeks, OpenAI's Codex CLI has gotten quite good, and I'll be explaining how I've been using it alongside Claude Code over the last few days to achieve even better results than using Claude Code alone. Firstly, why has it suddenly gotten good? That's because GPT-5 was launched about 3 weeks