The Claude Code Situation

Key Takeaways

So, I think it's safe to say that the past month or so has not been the best time for Anthropic. We just had the Clawed Code leak that happened. Then, even a few weeks ago, which is what I want to cover in this video, Claude Code actually masscommitted 51 times to try to build a C compiler from scratch. And you may have heard, but none of it actually compiles. Now, I'm going to put a bunch of stuff on screen so you guys can see what I'm talking about here, but Enthropic put 16 clawed agents on the task of building a full C compiler in Rust. They did 501 commits, created thousands of files, they were targeting four different CPU architectures, and Claude announced this and on paper it sounded pretty incredible. But then the community actually tried to use it and it couldn't even compile hello world. So it was almost 500 people that reacted to that exact issue on GitHub, which I'm going to show on screen here. But here's the thing. Somebody actually went out and fixed every single problem with this C compiler and then built an even better compiler completely from scratch using AI just with a different methodology which I'm going to get into here in just one second. All right, so let me set the stage here and kind of explain what happened. Now about a month ago, Anthropic published a blog post about a project that they called CCC or the Claude C compiler. Now the idea here was pretty ambitious. They were going to take 16 cloud agents, have them work together under the guidance of a senior researcher who's kind of prompting the models, going back and forth with them, and then build a full C compiler in Rust that's targeting different architectures like x86, risk v64, i686, etc. Okay, now this is pretty serious scope. This is a large project and something I couldn't from scratch. I'm sure many of you couldn't do the idea was the agent should be able to communicate with each other and build this. And to be honest here, I got to give a little bit of credit to Enthropic. This was a genuinely impressive experiment. And the fact that the AI agents could even attempt something like this is a pretty meaningful step forward for agentic coding. If you told me a year ago that this was even remotely possible with AI agents, I probably wouldn't have even believed you. And the thing is, a C compiler is one of the hardest pieces of software that you can build. So the ambition here was definitely real. But obviously when the community started digging into this, they found a bunch of problems. Now, the whole project actually relied heavily on one human expert. Now, this was a senior researcher who logged over 2,000 interactive turns, writing test cases, resolving merge conflicts, debugging, communicating with the agents, and the agents couldn't really self-correct or validate their own output at scale. So, they were producing a bunch of code, but it didn't really lead to the result that we were looking for. And here are some of the issues that a bunch of the community found in this C compiler. The ARM assembler had instruction encoding bugs. The x86 assembler had broken conditional processing. There were cross architecture alignment bugs. There was entire chunks of the C11 that were missing. And there was no real optimization pipeline. And there was no meaningful test infrastructure as well. Now, all of those are definitely problems, but the one major thing that caught a lot of people's attention is that not a single commit of this entire GitHub repository actually produced a working build. Meaning, if you went, you pulled it down, you tried to run it, you could never get something functional. Now, I'm not just saying all this to dunk on Enthropic. What they showed is that AI agents can take on massive complex engineering challenges. And this is pretty cool that they could even get something remotely close to working. However, it did reveal something important. Without a human expert who's acting as the architect, the context engineer, the QA, right? Basically project managing the entire build, these agents can't get to the finish line on their own. And they produce the scaffolding, not production or enterprise software. without that person sitting there coordinating the agents, they can't really do anything meaningful. Now, this is where the story starts to get really interesting and I'm going to share something with you that you probably haven't heard about the CCC or Claude Code compiler. Now, there's a company called Blitzy. I partnered with them a lot on this channel and they saw that Claude C compiler was broken and said, "Hey, we can fix every single one of these issues." And they went even a step further and they said actually let's try to build our own C compiler completely from scratch to compare our AI model or scaffolding you know orchestration whatever they have which we'll talk about in a second to clawed codes. Now Blitzy isn't just an LLM or another vibe coding tool and in fact it's not actually even trying to compete with Enthropic. What it is is a fully autonomous AI software development platform. Now the way that it works is fundamentally different from what Anthropic did and that's why it kind of makes it interesting here. Now, Blitzy starts by building a deep understanding of your codebase. So, it actually maps the entire codebase and constructs a dynamic knowledge graph of the entire system. So, dependencies, architecture, various patterns, conventions, whatever. Right? Now, they call this the infinite code context because it can handle code bases with over 100 million lines of code. Now, it doesn't do this in 5 minutes. When I've used Blitzy in the past, it can take hours. It can take even days depending on how complex and how large the codebase is. and it will generate a massive document that's usually hundreds of pages long. And I've shown this in many videos before that maps everything out so that when the AI does start making changes or producing code, it actually knows what it should be doing and it can follow this code base and all of the conventions and patterns that already exist. Now, once Blitzy understands the codebase, it starts orchestrating thousands of specialized AI agents that plan, build, and validate your code completely in parallel. Now, we're talking about agents that again run for hours or days. In my cases, I've had tasks as short as 5 hours, as long as 3 or 4 days. And what these agents do is collaborate across the entire codebase. Now, the platform here can actually batch build up to 50,000 lines of code in a single run without you having to interfere, ask the agent anything else. And there's actually multiple QA agents that are checking each other's work consistently before any code gets delivered. So, there's a built-in review loop and it's not just raw generation. And that's kind of what makes this unique is the orchestration of literally thousands of different agents. Now, this covers the full development life cycle. So, everywhere from requirements, design, code, test, PR, etc. And they recently scored a 66.5% on Sweet Bench Pro, which is independently audited and at least to my knowledge, one of the highest scores, if not the highest scores for any AI platform, tool, LLM, whatever you want to call it. So all of that is to say that effectively Blitzy works a lot differently than something like Claude Code or Codeex or these other tools even like cursor models whatever that you're using where those really rely on you sitting there in the terminal orchestrating the agents sending a lot of different messages opening up different sessions pressing approve reject whatever right you guys know the process and what AI coding typically looks like with Blitzy you literally submit the entire codebase it understands it you then give really detailed long prompts I've shown this in some other videos it then goes orchestrates the agents as I explained spends multiple days in some cases building the code using those different agents and then delivers you the result. So with all that said, what Blitzy decided to do is say, "Hey, okay, Claude Code wasn't able to do this effectively. Let's see if we can first of all fix any of the issues that they had using our approach and then can we build it from scratch and that's what I want to get into right now." So what Blitzy did here is ingest the CCC repo that had all of those different issues, right? It built out that knowledge graph of the existing architecture. So whatever cloud code had built itself and then they fixed all 13 critical PO regressions. They fixed the ARM assembly encoding issue, the x86 conditionals, they fixed all of the major problems that you saw in the original build and that people were pointing out on GitHub. And then on top of that, they built a real tiered optimization pipeline. They had loop ruling, tail call optimization, etc. And for the test suite, they had 753 of 753 unit tests passing. All of the integration tests passing with zero issues. and everything actually compiling and working properly. I'll put the repo on screen. You guys can check it out if you want. I'll leave it in the description as well, but that's just the start. So, like I mentioned before, Blitzy didn't just stop at actually fixing the issues in the current repository. What it did is attempted to build its own compiler completely from scratch because it's one thing to fix existing software. All of the scaffolding architecture, you know, more majority of the code is there and it's just fixing the bugs, but it's another thing to actually build it from the ground up without relying on any pre-existing code. So, what they were actually able to do here using the agent platform I talked about is rebuild everything completely from scratch with two prompts. I'm going to get into what those are in a minute, but that is just insane compared to the 2000 that was used by the expert that was working with the claw code. So, for this project, they named it the Blitz CC compiler or BCC. They generated 229,983 lines of Rust. They had 129 source files, 2,271 tests that were all passing with zero failures, and they had zero warnings, zero formatting diffs, zero external Rust dependencies. And same thing as before, they targeted four architectures. So they did the exact same as Claude's x86, i686, AR64, and Riskv64, risk 564. And here again is the most important thing, right? Claude's compiler needed the senior researcher who logged 2,000 interactive turns over two weeks. Whereas with Blitzy, they coordinated 3,600 agents autonomously just through the platform. They had a total engineering time of 624 hours of parallel agent work. And that was compressed into about 4 and 1/2 days of execution. Then with the human it was just two turns, so literally two. They had one to submit the prompt and then one to refine the final PR so it would work for Linux validation. So effectively on the first run almost everything was fully functioning. They added one more prompt and then everything was good to go and that's what you're looking at right now. Now look, I'm not going to get into all of the technicals here. You can read it from the blog post which I'll link below. But effectively the proof that this worked is that the BlitzC compiler first of all compiles and it runs SQLite readies Lua quick.js zlib and the big one it actually compiles the Linux kernel and boots successfully. And if we go through all of the details here, the BlisseyC compiler leads in 15 of the 22 measurable categories when comparing the two different um you know AI generated software, right? And this is backed by all of the committed test results which you can go and see for yourself. So all of this is cool, but I want to zoom out here and talk about why this actually matters. Now look, this isn't about compilers, right? No one really cares that much about C compilers and if AI can generate them, but what's happening here is we're seeing really the frontier of AI generated software. We've seen that certain approaches just don't work that effectively. When Enthropic's team, or in this case, the expert researcher was just using Claude, right? The Enthropic models or Claude code, it required 2,000 interactions. It got something very cool and very close to working, but still with all of the issues that you saw. And then with something like Blitzy, this is kind of showing where we're moving to in terms of a Centic software development where rather than having these short sessions where you're going and chatting with the model and going back and forth 2,000 times, by having the right orchestration framework and in this case 3600 individual agents, it can actually build fullcale enterprise software with the context engineering and again the multiple days of work. Normally when we're writing code or we're doing the AI gen it's back and forth very quickly you know two three four five minutes of generation in this case you spend more time on the context up front more time on the prompt specifying exactly what it is that you want and then the AI agents can go out and in this case work for 624 hours which is like multiple days right in parallel and generate fully functioning enterprise software which is kind of crazy to think about that we're already at that stage and we're just at the beginning here of 2026 now Of course, everything that I mentioned in this video, I'll leave in the description so you can check it out, read it for yourself. Obviously, you don't need to take my word for it, but I want to let you know quickly about really cool opportunity because on my channel recently, I've been talking a lot about OpenClaw, AI agents, AI orchestration platforms, and if you're someone who's interested in that, you've been building OpenClaw bots, then I want to let you know about a really cool opportunity that Blitzy actually brought to my attention. And this is called claw comp and it's a program where teams of two to three people can apply to each get a free Mac mini shipped to them to build a workflow or a product using openc clock. Now the competition culminates in a week-long hackathon where you'd get flown out to link ventures. This is a venture fund in Boston Massachusetts. All of the travel housing all of that stuff will be covered by link ventures for you. And honestly it's a great way to get hands-on experience with the kind of agentic technology that we've been talking about in this video and a lot on this channel. Now, if you're interested in that, again, it's called Claw Comp. I'm going to put the link in the description. Really cool competition. Definitely give it a shot. And I mean, if you get a free Mac me, that's pretty cool alone. If you get flown out there, I mean, that's going to be a great opportunity, something to build the resume, etc., and probably just a lot of fun. So, look, to recap all of it and close it out here, Claude built a C compiler that couldn't compile Hello World. Blitzy fixed all of the issues and then built an even better one from scratch with 230,000 lines of Rust code. Pretty much zero human code, very little human intervention, just a few prompts that I talked about that was able to boot the Linux kernel. Now, this is the future of software development that we are seeing right now. It is crazy what AI agents are capable of, but it is very clear that the way in which we orchestrate them is really what the future is going to be about here. We have really good base LLMs, but they need to be used in the correct environment. And if they have the right harness around them, what they're capable of building is really mind-boggling. And I don't know what I'm going to be talking about in one, you know, two months from now as things get crazier and crazier. Anyways, let me know what you guys think in the comments down below. And I look forward to seeing you in another