Home Creators Posts Import Register Favorites Logout

Downloads

Content

Ben and Emil of the Ben and Emil Show take us on a journey through the financial landscape of the tech sector and a glimpse of what to look out for in the coming months. Follow them at https://www.youtube.com/@benandemilshow

Files

Comments

Tom

I hate the Tesla Roadster for numerous reasons but one of the main ones being it's not even a roadster anymore. What a terrible excuse for a car

Theodora

As someone who’s been trying to get to the bottom of all this cybersecurity noise since the Mythos announcement, I have a few things I think are worth bearing in mind vis-a-vis the PHP stuff. This is an insanely long post but I'm completely obsessed with this so: (1) Anthropic’s pivot a few months ago to cybersecurity was, if not entirely for PR purposes, calculated. Their entire approach to hyping their products throughout 2025 was “We will replace all jobs!” But when it became clear that that was simply not going to happen, that’s when they started talking cybersecurity. (2) Much of the narrative emerges from straightforwardly insane people. Everyone in Anthropic is insane at this point. The New Yorker profile of them by Gideon Lewis-Kraus is great evidecne of this, even though he thinks it’s a well-reported investigation into the possible consciousness of Claude. They are all genuine TESCREAL weirdos who believe themselves to be building God. Much of the cybersecurity narrative comes from a group named METR, who are a bunch of insane 25-year-olds directly downstream of Eliezer Yudkowsky and co. According to his LinkedIn, their “lead scientist” did a master’s degree in training computers to play Atari games, interned twice at EA orgs (Future of Humanity and one other), then dropped out of his Ph.D. program to work at METR. Their investigation of the Hugging Face (god I’m fucking sick of saying and writing “Hugging Face”) was literally going to OpenAI HQ and asking ChatGPT to summarize the “reasoning output” (which has no direct relationship to the actual “reasoning”) from ChatGPT’s epic hack, then writing up a summary of what ChatGPT said ChatGPT did. They got written up in the Times for their work! It’s nothing. It’s straightforwardly nothing. The reason I said Anthropic’s pivot wasn’t entirely cynical is because nothing they do is entirely cynical, because they are insane. Building an Ultimate Hacker is a self-fulfilling TESCREAL prophecy: they think it’s one step on the way to building their epic machine god, so they do shit like run ChatGPT while it’s plugged into the internet in an obviously-insecure way (this is a bit technical but suffice it to say the means by which it was accessing the internet was not even designed for public internet-facing services; it was designed for use within secure enterprise networks) for four days without ever checking on it, having given it a prompt it is not permitted to stop running. Meanwhile, Hugging Face’s cybersecurity setup was such fucking trash that their systems had been penetrated for like two days before they even noticed what was happening, and their first response was to track down an open-source frontier LLM and ask it how to make it stop. (3) Cybersecurity always sounds fucking insane; it’s just that it wasn’t the object of such widespread attention before now. The book “Fancy Bear Goes Phishing” is a great example of this. Before AI, random fucking teens in New Jersey were building botnets out of smart toasters and almost taking down Google itself with DDoS attacks. It’s madness out there, and you can always make it sound fucking crazy. Some of this is just that. “Agentic swarms” are just malware. (4) Everyone is doing a horrible job differentiating between different kinds of, to borrow some horrible NYT-speak, AI-involved attacks. The major cybersec firm Crowdstrike’s most recent report -- cited in news articles as evidence of the unbelievable catastrophe of epic AI hacking -- says that AI has led to an increase of eighty-something percent more exploits. But later in the report (which is one of the ugliest and most bewildering documents I’ve ever seen), they let slip that they’re lumping cool hacks in *with exploits that occur thanks to insecure AI software*! That is, companies are installing stupid AI bullshit that’s probably vibe-coded and then getting their shit rocked and that’s considered the same kind of thing as AI doing cool hacks. (Sorry for the self-promo but if you’re curious about this kind of thing I wrote about this blurring here https://ncgenuineintelligence.mataroa.blog/blog/going-rogue-two-tales-of-ai-and-cybersecurity/ (or here for a dastardly substack link https://ncgenuineintelligence.substack.com/p/going-rogue-two-tales-of-ai-and-cybersecurity ) for a newsletter I write for the higher-ed union I work for -- the same week as the HuggingFace hack, I think, it was announced that Microsoft had refused to fix for 180 days a vuln with Copilot that made it so simply opening a Word document could result in your computer being “prompt injected,” which is to say that your AI services could be made to do shit that you don’t want it to do. That’s the real AI cybersec crisis, not this shit!) (5) We don’t have a ton of evidence one way or another, but Anthropic’s big cybersecurity push -- “Project Glasswing,” where they provided their big boy Mythos only to trustworthy companies like, uh, Microsoft -- seems, even on their terms, to have been total dogshit. All the flagship bug findings they reported early on were reproduced by LLMs from 2022 that you can run on your laptop in like two days. The only reason they weren’t fixed is because they didn’t matter. (More on that in a sec.) But to notably little fanfare they finally released a report and only 0.8% of found vulns have been fixed. This is like the worst hit rate ever. They claim it’s because humans take so long to review stuff but I think it’s much more likely that it’s because of false positives, and also because not all vulnerabilities even matter. There are a lot of fairly trivial vulnerabilities in software. AI-assisted vuln hunting returns so much random bullshit -- some of it is real but no-stakes; some of it is false positives -- that a lot of developers have actually shut down the “bug bounty” services they ran, a longstanding practice where you could get rewarded for finding bugs in open-source code, because they were getting so much slop. (6) Cybersecurity has been underfunded for decades now. The internet is basically held together with glue by desperate flailing experts who have conspicuously not been called upon to help out right now. Americans are scammed out of $150 BILLION a year through the same methods that cybersecurity violations happen: social engineering, phishing, etc. The problem is horrible, but it’s mundane and unsexy and AI is only accelerating it. It’s stuff like ClickFix making links unsafe or whatever. (7) Ok so PHP: I’m seeing a lot of people say that PHP has always been pretty vulnerable. It’s the first language some people learn and it’s useful for web dev so it’s something that less-expert people are doing a lot of stuff with. It does seem like it’s quite serious, though also not necessarily this newsworthy if not for AI. this guy is very good on AI and cybersec and I think he gets at a lot of the heart of the matter here: https://cyberplace.social/@GossiTheDog/117490226992669637 The internet is entirely reliant on shit that doesn’t get funded. The whole ecosystem is fucked. This would take too much to get into but basically the megacorps are all reliant on open-source software that’s maintained by like fucking volunteers and tiny orgs who don’t get paid by anyone making the shit that Facebook uses in its backend. It’s the same problems we’ve had since Web 2.0 at least but they’re just worse now because there’s so much money and garbage flooding the zone (8) Ultimately vis-a-vis AI I see the problem here as the same problem there is with vibe coding: while hobbyist programming is really cool (I am a very very light hobbyist programmer; it’s super fun and I love making dinky little shit), engineering actual secure software is extremely fucking hard. Some people tell me otherwise but as best as I can tell AI code trends inevitably towards slop. It makes software devs less attentive and they don’t understand the codebase and it’s just a mess. I know multiple people who are trying to change careers entirely because their whole jobs have been attempting to parse completely inscrutable Claude nonsense. And this is only going to get worse. Linux just started accepting AI contributions because Linus Torvalds is a piece of shit and he’s just in charge of the Linux Foundation forever which is getting bankrolled by corporations like OpenAI. They’ve even stopped mandating that people mark which changes are AI made because they don’t want to draw attention to it when that AI contributions suck. I’ve already heard experts saying that it’s manifestly making things worse. The Linux kernel is 40,000,000 lines of code -- there’s simply no way that an LLM limited by context windows is going to be able to contribute to that structure in a way that strengthens it. Here’s the thing: 90% of the world’s servers run on Linux. Vulnerabilities in Linux have the capacity to lead to genuine cascading failures -- worse than PHP, worse than Crowdstrike 2024 levels. And everyone is doing this. Meta’s new agent is psychotic slop nonsense. This is very jargony but this whole thread is hilarious (https://neuromatch.social/@jonny/117478470249100713) it’s basically just like it’s so unfathomably profoundly insecure, and Meta explicitly uses Claude Code exclusively lol.

Bryan Thomas

What's the song at the end? Is that a Caleb production?

Alicia Ruth

Another absolute banger closing song

Mew M

The audio on this one keeps cutting off while Ben and Emil are talking, especially during the Ben bits. Was there some recording/encoding snafu?

Samantha

Thank you for writing this up, I appreciate your obsession. I have such a hard time keeping track of and comprehending this kind of thing, but my partner works in cyber security so I always just have him explain every cyber security-related news story to me. Everything you said about cyber security being so fucked in this country is what he’s been saying for years.

Kill The Computer

Thanks for letting me know, I think it was an export thing. In my DAW it sounded fine but I hear what you are saying. I re-exported it and it sounds fine so I'm uploading new copies now. I appreciate the feedback!

Bryan Thomas

So damn good, man! Nothing but bangers. We need the album.

James Rule

AI tech is a million dollar idea that people are trying to sell for trillions of dollars.

Theodora

yeah i love blabbing about this shit so it's no problem, sorry that it's so cobbled-together and poorly written lol anyway seems like php mightve been more of an AI-narrative-directed false alarm than i thought: https://stodon.swhack.com/@dpk/117495463314150312 . what a nightmarishly opaque information ecosystem. same as it ever was i guess

Ricardo Gutierrez

Came in to ask this! Caleb where can we find this banger?? 🔥🔥

ChigBungus

Seeing people genuinely believe that the new Tesla will be able to fly makes me feel insane. About to go full Sam Neil at the end of In the Mouth of Madness.