{
  "video_id": "GtTJ6LcqdCE",
  "channel_slug": "techwithtim",
  "channel_handle": "Tech With Tim",
  "title": "The Claude Code Situation",
  "duration_seconds": 803,
  "url": "https://www.youtube.com/watch?v=GtTJ6LcqdCE",
  "upload_date": "20260410",
  "transcript": "So, I think it's safe to say that the\npast month or so has not been the best\ntime for Anthropic. We just had the\nClawed Code leak that happened. Then,\neven a few weeks ago, which is what I\nwant to cover in this video, Claude Code\nactually masscommitted 51 times to try\nto build a C compiler from scratch. And\nyou may have heard, but none of it\nactually compiles. Now, I'm going to put\na bunch of stuff on screen so you guys\ncan see what I'm talking about here, but\nEnthropic put 16 clawed agents on the\ntask of building a full C compiler in\nRust. They did 501 commits, created\nthousands of files, they were targeting\nfour different CPU architectures, and\nClaude announced this and on paper it\nsounded pretty incredible. But then the\ncommunity actually tried to use it and\nit couldn't even compile hello world. So\nit was almost 500 people that reacted to\nthat exact issue on GitHub, which I'm\ngoing to show on screen here. But here's\nthe thing. Somebody actually went out\nand fixed every single problem with this\nC compiler and then built an even better\ncompiler completely from scratch using\nAI just with a different methodology\nwhich I'm going to get into here in just\none second. All right, so let me set the\nstage here and kind of explain what\nhappened. Now about a month ago,\nAnthropic published a blog post about a\nproject that they called CCC or the\nClaude C compiler. Now the idea here was\npretty ambitious. They were going to\ntake 16 cloud agents, have them work\ntogether under the guidance of a senior\nresearcher who's kind of prompting the\nmodels, going back and forth with them,\nand then build a full C compiler in Rust\nthat's targeting different architectures\nlike x86, risk v64, i686, etc. Okay, now\nthis is pretty serious scope. This is a\nlarge project and something I couldn't\nfrom scratch. I'm sure many of you\ncouldn't do the idea was the agent\nshould be able to communicate with each\nother and build this. And to be honest\nhere, I got to give a little bit of\ncredit to Enthropic. This was a\ngenuinely impressive experiment. And the\nfact that the AI agents could even\nattempt something like this is a pretty\nmeaningful step forward for agentic\ncoding. If you told me a year ago that\nthis was even remotely possible with AI\nagents, I probably wouldn't have even\nbelieved you. And the thing is, a C\ncompiler is one of the hardest pieces of\nsoftware that you can build. So the\nambition here was definitely real. But\nobviously when the community started\ndigging into this, they found a bunch of\nproblems. Now, the whole project\nactually relied heavily on one human\nexpert. Now, this was a senior\nresearcher who logged over 2,000\ninteractive turns, writing test cases,\nresolving merge conflicts, debugging,\ncommunicating with the agents, and the\nagents couldn't really self-correct or\nvalidate their own output at scale. So,\nthey were producing a bunch of code, but\nit didn't really lead to the result that\nwe were looking for. And here are some\nof the issues that a bunch of the\ncommunity found in this C compiler. The\nARM assembler had instruction encoding\nbugs. The x86 assembler had broken\nconditional processing. There were cross\narchitecture alignment bugs. There was\nentire chunks of the C11 that were\nmissing. And there was no real\noptimization pipeline. And there was no\nmeaningful test infrastructure as well.\nNow, all of those are definitely\nproblems, but the one major thing that\ncaught a lot of people's attention is\nthat not a single commit of this entire\nGitHub repository actually produced a\nworking build. Meaning, if you went, you\npulled it down, you tried to run it, you\ncould never get something functional.\nNow, I'm not just saying all this to\ndunk on Enthropic. What they showed is\nthat AI agents can take on massive\ncomplex engineering challenges. And this\nis pretty cool that they could even get\nsomething remotely close to working.\nHowever, it did reveal something\nimportant. Without a human expert who's\nacting as the architect, the context\nengineer, the QA, right? Basically\nproject managing the entire build, these\nagents can't get to the finish line on\ntheir own. And they produce the\nscaffolding, not production or\nenterprise software. without that person\nsitting there coordinating the agents,\nthey can't really do anything\nmeaningful. Now, this is where the story\nstarts to get really interesting and I'm\ngoing to share something with you that\nyou probably haven't heard about the CCC\nor Claude Code compiler. Now, there's a\ncompany called Blitzy. I partnered with\nthem a lot on this channel and they saw\nthat Claude C compiler was broken and\nsaid, \"Hey, we can fix every single one\nof these issues.\" And they went even a\nstep further and they said actually\nlet's try to build our own C compiler\ncompletely from scratch to compare our\nAI model or scaffolding you know\norchestration whatever they have which\nwe'll talk about in a second to clawed\ncodes. Now Blitzy isn't just an LLM or\nanother vibe coding tool and in fact\nit's not actually even trying to compete\nwith Enthropic. What it is is a fully\nautonomous AI software development\nplatform. Now the way that it works is\nfundamentally different from what\nAnthropic did and that's why it kind of\nmakes it interesting here. Now, Blitzy\nstarts by building a deep understanding\nof your codebase. So, it actually maps\nthe entire codebase and constructs a\ndynamic knowledge graph of the entire\nsystem. So, dependencies, architecture,\nvarious patterns, conventions, whatever.\nRight? Now, they call this the infinite\ncode context because it can handle code\nbases with over 100 million lines of\ncode. Now, it doesn't do this in 5\nminutes. When I've used Blitzy in the\npast, it can take hours. It can take\neven days depending on how complex and\nhow large the codebase is. and it will\ngenerate a massive document that's\nusually hundreds of pages long. And I've\nshown this in many videos before that\nmaps everything out so that when the AI\ndoes start making changes or producing\ncode, it actually knows what it should\nbe doing and it can follow this code\nbase and all of the conventions and\npatterns that already exist. Now, once\nBlitzy understands the codebase, it\nstarts orchestrating thousands of\nspecialized AI agents that plan, build,\nand validate your code completely in\nparallel. Now, we're talking about\nagents that again run for hours or days.\nIn my cases, I've had tasks as short as\n5 hours, as long as 3 or 4 days. And\nwhat these agents do is collaborate\nacross the entire codebase. Now, the\nplatform here can actually batch build\nup to 50,000 lines of code in a single\nrun without you having to interfere, ask\nthe agent anything else. And there's\nactually multiple QA agents that are\nchecking each other's work consistently\nbefore any code gets delivered. So,\nthere's a built-in review loop and it's\nnot just raw generation. And that's kind\nof what makes this unique is the\norchestration of literally thousands of\ndifferent agents. Now, this covers the\nfull development life cycle. So,\neverywhere from requirements, design,\ncode, test, PR, etc. And they recently\nscored a 66.5% on Sweet Bench Pro, which\nis independently audited and at least to\nmy knowledge, one of the highest scores,\nif not the highest scores for any AI\nplatform, tool, LLM, whatever you want\nto call it. So all of that is to say\nthat effectively Blitzy works a lot\ndifferently than something like Claude\nCode or Codeex or these other tools even\nlike cursor models whatever that you're\nusing where those really rely on you\nsitting there in the terminal\norchestrating the agents sending a lot\nof different messages opening up\ndifferent sessions pressing approve\nreject whatever right you guys know the\nprocess and what AI coding typically\nlooks like with Blitzy you literally\nsubmit the entire codebase it\nunderstands it you then give really\ndetailed long prompts I've shown this in\nsome other videos it then goes\norchestrates the agents as I explained\nspends multiple days in some cases\nbuilding the code using those different\nagents and then delivers you the result.\nSo with all that said, what Blitzy\ndecided to do is say, \"Hey, okay, Claude\nCode wasn't able to do this effectively.\nLet's see if we can first of all fix any\nof the issues that they had using our\napproach and then can we build it from\nscratch and that's what I want to get\ninto right now.\" So what Blitzy did here\nis ingest the CCC repo that had all of\nthose different issues, right? It built\nout that knowledge graph of the existing\narchitecture. So whatever cloud code had\nbuilt itself and then they fixed all 13\ncritical PO regressions. They fixed the\nARM assembly encoding issue, the x86\nconditionals, they fixed all of the\nmajor problems that you saw in the\noriginal build and that people were\npointing out on GitHub. And then on top\nof that, they built a real tiered\noptimization pipeline. They had loop\nruling, tail call optimization, etc. And\nfor the test suite, they had 753 of 753\nunit tests passing. All of the\nintegration tests passing with zero\nissues. and everything actually\ncompiling and working properly. I'll put\nthe repo on screen. You guys can check\nit out if you want. I'll leave it in the\ndescription as well, but that's just the\nstart. So, like I mentioned before,\nBlitzy didn't just stop at actually\nfixing the issues in the current\nrepository. What it did is attempted to\nbuild its own compiler completely from\nscratch because it's one thing to fix\nexisting software. All of the\nscaffolding architecture, you know, more\nmajority of the code is there and it's\njust fixing the bugs, but it's another\nthing to actually build it from the\nground up without relying on any\npre-existing code. So, what they were\nactually able to do here using the agent\nplatform I talked about is rebuild\neverything completely from scratch with\ntwo prompts. I'm going to get into what\nthose are in a minute, but that is just\ninsane compared to the 2000 that was\nused by the expert that was working with\nthe claw code. So, for this project,\nthey named it the Blitz CC compiler or\nBCC. They generated 229,983\nlines of Rust. They had 129 source\nfiles, 2,271\ntests that were all passing with zero\nfailures, and they had zero warnings,\nzero formatting diffs, zero external\nRust dependencies. And same thing as\nbefore, they targeted four\narchitectures. So they did the exact\nsame as Claude's x86, i686, AR64, and\nRiskv64, risk 564. And here again is the\nmost important thing, right? Claude's\ncompiler needed the senior researcher\nwho logged 2,000 interactive turns over\ntwo weeks. Whereas with Blitzy, they\ncoordinated 3,600 agents autonomously\njust through the platform. They had a\ntotal engineering time of 624 hours of\nparallel agent work. And that was\ncompressed into about 4 and 1/2 days of\nexecution. Then with the human it was\njust two turns, so literally two. They\nhad one to submit the prompt and then\none to refine the final PR so it would\nwork for Linux validation. So\neffectively on the first run almost\neverything was fully functioning. They\nadded one more prompt and then\neverything was good to go and that's\nwhat you're looking at right now. Now\nlook, I'm not going to get into all of\nthe technicals here. You can read it\nfrom the blog post which I'll link\nbelow. But effectively the proof that\nthis worked is that the BlitzC compiler\nfirst of all compiles and it runs SQLite\nreadies Lua quick.js zlib and the big\none it actually compiles the Linux\nkernel and boots successfully. And if we\ngo through all of the details here, the\nBlisseyC compiler leads in 15 of the 22\nmeasurable categories when comparing the\ntwo different um you know AI generated\nsoftware, right? And this is backed by\nall of the committed test results which\nyou can go and see for yourself. So all\nof this is cool, but I want to zoom out\nhere and talk about why this actually\nmatters. Now look, this isn't about\ncompilers, right? No one really cares\nthat much about C compilers and if AI\ncan generate them, but what's happening\nhere is we're seeing really the frontier\nof AI generated software. We've seen\nthat certain approaches just don't work\nthat effectively. When Enthropic's team,\nor in this case, the expert researcher\nwas just using Claude, right? The\nEnthropic models or Claude code, it\nrequired 2,000 interactions. It got\nsomething very cool and very close to\nworking, but still with all of the\nissues that you saw. And then with\nsomething like Blitzy, this is kind of\nshowing where we're moving to in terms\nof a Centic software development where\nrather than having these short sessions\nwhere you're going and chatting with the\nmodel and going back and forth 2,000\ntimes, by having the right orchestration\nframework and in this case 3600\nindividual agents, it can actually build\nfullcale enterprise software with the\ncontext engineering and again the\nmultiple days of work. Normally when\nwe're writing code or we're doing the AI\ngen it's back and forth very quickly you\nknow two three four five minutes of\ngeneration in this case you spend more\ntime on the context up front more time\non the prompt specifying exactly what it\nis that you want and then the AI agents\ncan go out and in this case work for 624\nhours which is like multiple days right\nin parallel and generate fully\nfunctioning enterprise software which is\nkind of crazy to think about that we're\nalready at that stage and we're just at\nthe beginning here of 2026 now Of\ncourse, everything that I mentioned in\nthis video, I'll leave in the\ndescription so you can check it out,\nread it for yourself. Obviously, you\ndon't need to take my word for it, but I\nwant to let you know quickly about\nreally cool opportunity because on my\nchannel recently, I've been talking a\nlot about OpenClaw, AI agents, AI\norchestration platforms, and if you're\nsomeone who's interested in that, you've\nbeen building OpenClaw bots, then I want\nto let you know about a really cool\nopportunity that Blitzy actually brought\nto my attention. And this is called claw\ncomp and it's a program where teams of\ntwo to three people can apply to each\nget a free Mac mini shipped to them to\nbuild a workflow or a product using\nopenc clock. Now the competition\nculminates in a week-long hackathon\nwhere you'd get flown out to link\nventures. This is a venture fund in\nBoston Massachusetts. All of the travel\nhousing all of that stuff will be\ncovered by link ventures for you. And\nhonestly it's a great way to get\nhands-on experience with the kind of\nagentic technology that we've been\ntalking about in this video and a lot on\nthis channel. Now, if you're interested\nin that, again, it's called Claw Comp.\nI'm going to put the link in the\ndescription. Really cool competition.\nDefinitely give it a shot. And I mean,\nif you get a free Mac me, that's pretty\ncool alone. If you get flown out there,\nI mean, that's going to be a great\nopportunity, something to build the\nresume, etc., and probably just a lot of\nfun. So, look, to recap all of it and\nclose it out here, Claude built a C\ncompiler that couldn't compile Hello\nWorld. Blitzy fixed all of the issues\nand then built an even better one from\nscratch with 230,000 lines of Rust code.\nPretty much zero human code, very little\nhuman intervention, just a few prompts\nthat I talked about that was able to\nboot the Linux kernel. Now, this is the\nfuture of software development that we\nare seeing right now. It is crazy what\nAI agents are capable of, but it is very\nclear that the way in which we\norchestrate them is really what the\nfuture is going to be about here. We\nhave really good base LLMs, but they\nneed to be used in the correct\nenvironment. And if they have the right\nharness around them, what they're\ncapable of building is really\nmind-boggling. And I don't know what I'm\ngoing to be talking about in one, you\nknow, two months from now as things get\ncrazier and crazier. Anyways, let me\nknow what you guys think in the comments\ndown below. And I look forward to seeing\nyou in another",
  "transcript_chars": 15889,
  "ingested_at": "2026-05-21T19:01:25.693815+00:00",
  "source": "retry-no-transcript",
  "yt_meta": {
    "view_count": 10582,
    "like_count": 258,
    "channel_id": "UC4JX40jDee_tINbkjycV4Sg"
  }
}