This is an iHeart podcast. Guaranteed human. Hello and welcome to this week's Better Off Line. I'm your host at Zitron. So yes, you're getting two monologues this week because we had a guest pull out last minute leaving me with just my keyboard, my microphone, and a Diet Coke, him one of those weird vacuum sealed tubes that really only real Diet Coke freaks on. And today I'm going to start with a few questions. What's going on in Anthropic? What is Dario Amade up to? Why do things keep breaking? Used to be the coders could just launch Claude and run a bunch of sub agents to look productive. Think you can rely on Claude code? Guess again, Dario Amade doesn't care about you. And he certainly doesn't care about your family. Anyway, to set the scene, I need to give you a few details about Claude code in general, which you don't pay for Claude code itself. There's not like a Claude code subscription. When you subscribe to Anthropic service, you pay for a Claude subscription. And you either pay 20 bucks, 100 bucks or 200 bucks a month with each tier having more access than the last and access meaning great limits so you can use it more. Though if you're wondering what that actually means, that's really by design. As all Anthropic tells you is that you've got a five hour limit and a weekly limit on them models with a separate one for Claude Opus, which is their more expensive, more complex one. An important detail is that these subscriptions are also heavily subsidized. When you pay Anthropic $200 a month, you're not paying on a per token rate, which is what AI startups have to do. They pay per million input and output tokens. No, no, no. You just do stuff and stuff comes out. And you're able to do the same things you would if you use in the API, you're burning tokens and you can burn over $2,500 in model tokens on a 200 bucks a month subscription, or at least you could. And some users have been able to spend upwards of $5,000 on the same subscription. Rate limits are generally used to restrict people from burning as much, but the problem has become that for Anthropic to approach anything close to profitability would have to rate limit people into the, well, into the depths of hell at this point. Moving on, in the middle of March, Anthropic started a two week long promotional campaign where they doubled rate limits for off peak hours using Claude, set to end on March 27th, 2026. A day before it was set to end on March 26th, 2026, Anthropic would announce it was starting peak hours with Claude code users maxing out their sessions faster between the hours of 5 a.m. and 11 p.m. Pacific, Monday to Friday, with a spokesperson limply adding that efficiency wins, unnamed of course, would offset this and only 7% of users would hit the limits. All of this was sold as a result of managing the growing demand for Claude. Yeah, spoiler alert, more than 7% of users appeared to have hit the limits and nobody seems to be feeling particularly efficient. Don't know whether those gains are. One user on $100 a month max plan complained about hitting 61% of his session limit, the five hour one, after four prompts, which he found out based using a tool called CCUsage, cost $10.26 in tokens. He still spent 10% of his subscription and four prompts. Another said that they hit 63% of the limit on their 200 a month plan in the space of a day and another hit 95% after 20 minutes of using their max plan. Gonna guess $100 a month on that one. Another person hit their max limit after two or three things, don't know what they were. And another vowed to cancel their $200 a month subscription after hitting their weekly limit in the space of a day saying that they, and I'm going off of a translation from fucking Grog. So forgive me, expected a premium experience between $200 and what they got was constant limit stress. I've linked a lot of these in yesterday's free newsletter, the subprime AI crisis is here. Really advise you read it, but also you can see how many people are mad or just go on Twitter and search Claude limits, it's not great. Now while Anthropic technical staff member Lydia Halley posted that Anthropic was aware of people hitting usage limits in Claude code way faster than expected and that some investigation of some sort was taking place, it's hard to imagine that Anthropic had no idea that these limits were so severe, any of this was a surprise. Now, as I wrote this sentence that I'm reading on Tuesday, March 31st, it doesn't appear that any changes have been made. People are still complaining about hitting their limits in a few prompts and Anthropic has yet to update anyone in my opinion because these are the rate limits they decided were necessary to keep the business going and roll their nasty ass into IPO. A few days previously though, Anthropic could also accidentally, and I put that in quotation marks, leak the existence of their Kappie Barra and Mythos models to Fortune magazine, by which I mean they had a data cache to quote Fortune with over 3000 assets that was left open on the internet and Fortune somehow found it. You know, it kind of reminds me of like a skeezy bloke dropping a magnum condom out of his wallet in front of women being like, hey, hey, you see that? Oh, whoops, whoops. In any case, this massive leak or some included absolutely fucking nothing about the models themselves other than that they're a step change better and that their cybersecurity features were so very scary that they would have to roll them out slowly. I'm, I don't know, man. I know insult to the people at Fortune, I just don't think they're on fucking showdown looking up AWS buckets. So wonder how this got there. Now, I will say at first I was completely sure that this was a deliberate leak because 3000 assets left open on the internet and none of them have any info on the model but just scary things like, oh, it's so much better. And I'm not so sure because yesterday Anthropic accidentally leaked the entire source code of its Claude code coding interface. The leak was a result of a reference in Claude code's MPM package, the thing that you queried to download Claude code onto your computer via the terminal that led right back to a zip archive on Anthropic servers per the register that contain the source code. I will add that there's an ongoing discussion about what actually caused this problem with someone hack a news, hack a news, hack and use, I'm just gonna keep it, saying it might be a problem with bun, the packaging tool used to allow people to download Claude code, that Anthropic acquired in December of last year. To be clear, this isn't a leak of Anthropic's models but it's still an unbelievably large leak, one that exposed Claude code's innards to the entire internet and all of their competitors. And while I imagine using the source code is illegal on some level, I can't imagine there's any reason their competitors can't take a look you think that they're all sitting around being like, oh, I absolutely can't, I mustn't, it's not okay. Especially when this is a company that fucked over just about anyone building anything that you build on top of a Claude code subscription. Thinking about how they treated open code, by the way. Anyway, now is a great time to remind you that Claude code creator Boris Cherny said at the end of December that 100% of his contributions to Claude code were written in Claude code and told Lenny Rachitsky in February, 2026, the coding is now solved for most use cases. I assume that the use cases that haven't been solved include making sure that there isn't a direct link from the Claude code installer to a Cloudflare storage bucket with its source code ready to download. That one just doesn't seem like they got that one pinned down. I'll also add that the same interview added that Cherny hadn't written a single line of code since November, 2025, which is, I mean, those of you playing at home, you're all, I know many of you really enjoy my game show. Is that good? And I just have to ask you, is that good? Anyway, I feel like I very recently warned everybody about the very obvious dangers of allowing LLMs to write all of your code. These models do not have thoughts or knowledge or really anything other than probabilistic generations of outputs based on training data. And a lot of it, and they're quite complex. Nevertheless, this means that any time you choose to just accept the code they generate without reading it thoroughly, you're choosing to trust something inherently untrustworthy that doesn't think or have knowledge. This is just the beginning of us finding out the ugly cost of software engineers trusting large language models to write their code at scale. LLMs are good at writing lots of code, which in turn means that the code requires far more time to review, which assumes you do review it, which becomes far more difficult when you're constantly pressured to ship more and more and more software every fucking week. Or if you're Boris Cherny and you believe that shipping software fast is the same thing as shipping good software. I'm now thinking about the launch of Anthropics' Claude Cowork, a product that can allegedly manage files on your computer or draft documents or some such bullshit and nobody seems to be able to explain it. At the time, Boris Cherny proudly boasted that Claude Cowork was built in around a week and a half, which makes perfect sense, says almost immediately a story came out of a guy who tried to reorganize his wife's desktop files using Cowork, which is one of the literal things on the Cowork website. This is the future that these so-called large language model companies want for you. Bad software, shipped quickly, hyped by a captured media that doesn't give a damn about whether the services are the things you pay for are functional or useful and doesn't even bother using the tools or understanding them. And nobody else appears to be discussing how inherently deceptive these companies have become. A subscriber to Claude Cowork said, I'm not a subscriber to the company, I'm a subscriber to the company, I'm a subscriber to the company, A subscriber to Claude who played for an annual subscription in December 2025 now was a product that has significantly less value thanks to egregious rate limits. And that's also like if you can even do your job anymore. I imagine people in the 20 buck a month subscription are really fucking suffering. And in general, Anthropics models seem to go and look at the Reddit if you want to understand more. They seem to oscillate in utility and efficacy based on the time of day you use them and their proximity to a minute model launch. This is a really weird story that I've never been able to get to the bottom of, but there have been reports for over a year of different models from Anthropik, from OpenAI, from whomever, getting dumber at random times. No one's able to get to the bottom of it, I've heard people saying they quantized the models make them smaller during the day. I dunno. If you have any information about this, shoot me a piss on plurk or email me at easy at betteroffline.com that's echoesator at betteroffline.com. What we do know is that Claude Code, one of the few popular AI products, is built with the same disregard for safety and customer happiness as the rest of Anthropik's astonishingly shitty business that burns billions of dollars with no end in sight. Chief Claude Code's sloppagandist Boris Cherney doesn't give enough of a fuck about his customers to actually read his code and I imagine the same goes for a lot of other engineers within the company and big tech at large. The consequences are already obvious. In the last few months, AI coding tools brought down AWS twice and lost Amazon hundreds of thousands of orders and led to a security breach inside Meta mere weeks ago. And now these tools have leaked Claude Code's source code. How much is enough to make people wake up to the inherent dangers of using these models to write software at scale? I guess we're gonna find out. I'll see you on Friday for another monologue. Hans, the GC here. I'm whispering because... As the queen. Queen of social media? It's about time for my ASGMR series. So I'm recording this on my phone and then I'm going to use Canva to edit and upload it. Oh, sorry babes. I'll make that whisper when I edit it. Anyways, Canva makes social media edits so easy. I'll upload this in a minute. Canva, make everything. Iconic. How do I stop recording, Zaryn? This is an iHeart podcast. Guaranteed human.
