Intense Focus: Anthropic Breach and Investment Risks  🔗 **Source:** [Intense Focus: Anthropic Breach and Investment Risks](https://share.snipd.com/episode/f2bdeb4b-0e7e-4376-94b0-b6bae9e098c3) ## Highlights - **Claude Watermark Was Bypassed Within Hours** - Anthropic's Claude watermark was bypassed within hours, showing watermarking is fragile and quickly exploitable. - Developer Guillaume Mayer published a working override four hours after launch and the code got huge traction on X and GitHub. Transcript: Jaeden Schafer Let's talk about what's going on with Anthropic right now. Because there's a whole bunch of kind of changes that they were making, not just with the text model, but with basically all of the code, everything that was coming out of Claude, they're Saying, look, this is all going to be watermarked because of a new regulation that came out of the EU at their new EU kind of bill that the AI bill that passed. And because of this, they were going to watermark everything that came out of AI. Now, some of the controversy I've heard in the argument here is a lot of people are saying, look, I'm like trying to write an article, or I'm trying to write an email, or I'm trying to write Any sort of document. I'm sitting there and typing out the rough draft. I throw it in there and say, polish this a bit, right? It could be something very small, like, hey, just make sure the grammar is good. When Claude spits it out and it's like, hey, I tweaked the grammar for you or fixed it up a little bit. All of a sudden, the output is going to be watermarked, even if I mean, it's just lightly tweaking it. You go take this document and you submit it somewhere. And all of a sudden, people are concerned that you're going to get penalized, whether that's at your job or whether that's at school or anywhere else, because you used AI. So there's, there's a lot of that that's going on. But anyways, with this particular jailbreak, someone created, so I think a lot of people are excited about this because of all of those reasons. Developer Guillaume Mayer, he published a working watermark override. He did this four hours after Anthropoc announced that they were going to be adding the watermark. So it's not like this takes a very long time. The removal code has been bookmarked more than 20,000 times on X, and it has over 100 GitHub contributors. ([Time 0:01:32](https://share.snipd.com/snip/10ba4d3a-ed3f-413d-9e93-d7c887582e79)) - **EU Rules Force Fast Watermark Rollouts** - The EU AI Act mandates watermarking with tight deadlines and steep fines, pushing companies to implement imperfect detection quickly. - Penalties can reach 3% of annual turnover, forcing major players like Anthropic to comply by August/December deadlines. Transcript: Jaeden Schafer The EU AI Act basically mandates this watermarking for new models. And all this has to happen by August. So the deadline came and they had to have everything done by the end of the month and all of the existing models that currently exist. So the new models, everyone's got to have it by the end of August. But even the older models all need to have this by December. And the fines are not a joke. They're up to 3% of your annual turnover. So these companies making, you know, Anthropic is at $65 billion in annual run rate. They do not want to spend any percentage of that for fines because they don't have watermarking. ([Time 0:02:57](https://share.snipd.com/snip/da24935a-3d08-4dc0-8501-b0fc47f39ede)) - **Engineers Built Watermark Removals As A Hobby** - Jaeden describes engineers quickly building competing removal tools, including Eric Hughes who made one in 15 minutes using Claude. - Mayer, a native French speaker, says he edits his writing with Claude and Grammarly, illustrating everyday dependency on these tools. Transcript: Jaeden Schafer But the original tool that everyone that kind of went viral made by Myers, he's a native French speaker. He says he edits all of his writing with Claude and Grammarly. And I mean, if I was trying, I also speak French. I'm not very good. But if I'm, if I was, if I'm ever sending French to someone, a friend in France, I'm always going to throw it through Chachupetier or Claude to go fix the grammar or fix it. I might even voice to text it, kind of write it out. But like, I know I got the grammar wrong, whatever. So for anyone that English isn't their first language, this makes perfect sense. And even for me, where English is my first language, I'm still using AI tools to help with my grammar. And so just to have all of that automatically watermarked, right, especially when you think of a tool like Grammarly, or any of the others that are implementing an API, because this Is the watermark is through the API as well. So anyone that's like, hey, we'll help you with like writing or researching your school will help you with your grammar. Like any of these AI tools are using an API to cloud. They're all getting watermarked as well. ([Time 0:04:07](https://share.snipd.com/snip/a1bec097-9ff4-4c3f-9637-eec5cdd92d72)) - **Amazon Drops Alexa Plus Fee To Stay Competitive** - Amazon removed the $19.99 Alexa Plus fee and auto-upgraded Fire TV devices with the AI assistant to compete with Google and Roku. - Jaeden points to dramatically higher conversation volume when assistants provide true conversational AI versus legacy voice assistants. Transcript: Jaeden Schafer Amazon is dropping the $19.99 fee for Alexa Plus. So if you wanted to get their AI to help you with Alexa, you had to pay an extra $20 a month. And I think the reason for this is not shocking. It all comes down to competition. But for this, it's going to affect every current Fire TV stick, Fire TV Cube, and Alexa Plus enabled smart TV in the US is they're going to all get the AI assistant. It's all auto upgraded. You don't have to get an app. You don't have to get a sign on. You don't have to get Prime. You just get this automatically baked in. Amazon said that Alexa Plus users have nearly twice as many conversations on Fire TV as users had with the original Alexa. Why are they even announcing that? Because it is so blatantly obvious. ([Time 0:05:43](https://share.snipd.com/snip/f550c06d-b2ae-4944-b482-9d878733b66b)) - **Calendly Adds System Audio Note Taker And Cali** - Calendly launched an AI note taker plus an assistant called Cali that records, transcribes, summarizes, drafts follow-ups, and generates action items. - Instead of a meeting bot, Calendly listens to system audio to avoid in-meeting bot friction and possible platform limitations. Transcript: Jaeden Schafer Transcribe and summarize, draft and your follow up emails. And it's going to generate action items from meetings. So I mean, this is basically like Firefly or tons of these other tools where you have an AI bot inside the call. The difference though, is Calendly is trying to copy what Granola is doing. And instead of having like an extra bot inside of a Google Meet, for example, they're going to be listening to your system audio. I think there's a couple reasons why. One of them, which maybe I'm like pessimistic here, but if I was Calendly and I saw Google come in and basically kneecap my company by cloning our feature and then sabotaging all of our Signups, they look spammy, I would be very suspicious about building some sort of integration, which is just like another member bot added to like a meeting. And maybe there's like some way that Google could shut down bots and meetings, or opt in because they only want you to use the Gemini bot in the meeting. Like, you know, I imagine there could be something like this system audio. On the other hand, number one, it just seems way more intuitive. If you opt in on the system wide level, you never have to touch something. You never have to say like this person's meeting taker is in the meeting. You don't, no one has to accept it. Google can't shut it off. It's just literally listening to your system wide audio level. ([Time 0:08:52](https://share.snipd.com/snip/5fa3c064-a155-4872-9381-ceaff5a9c6ba)) - **Google's $12B Marvell Deal Escalates Chip Wars** - Google agreed to buy up to $12.2B of Marvell shares tied to AI chips, signaling intense hyperscaler leverage in chip supply deals. - The stake vests against purchase targets through Marvell's fiscal 2033, and Broadcom stock fell on the news. Transcript: Jaeden Schafer Marvel's shares and stock has jumped 6% on a Google deal. It's for up to $12.2 billion. This is a stake tied to their AI chips and chips are basically the hottest topic in AI right now. Everyone is trying to get more of them. No one can create enough or buy enough. Google is going to buy up to 58 million Marvel shares at $200 a piece. It's worth about $12.2 billion. The stake is going to vest against Google's purchasing targets, which are running through Marvel's fiscal 2033. So these are, you know, vesting over a long period of time, but this is a really impressive partnership and Marvel's, you know, I mean, they're up 6%. ([Time 0:11:11](https://share.snipd.com/snip/46341656-aec8-4b03-be5d-97b165d2216e)) - **Flock Uses Plate Co-Occurrences To Infer Associates** - Flock introduced AI features that infer driver associations by co-occurrence windows instead of face recognition, raising fresh privacy alarms. - The associate tool defaults to 0.75 confidence, returns up to 20 vehicles, and requires only a brief justification with a minimal case-number check. Transcript: Jaeden Schafer I recently listened to an interview with the Flock CEO, and maybe a lot of people did on the All In podcast. And my takeaway from the whole thing was I'm personally really just not a big fan of Flock and to maybe get like some level of, you know, not to get too political, but I think from a political Perspective on the right and the left, there's reasons why people in America don't like flock, right? These cameras that are basically on the side of the road and they take a picture of people's license plates as they go by to track where cars are driving. I think if you're probably, the argument I've heard from liberals is that they don't like, you know, that you could possibly use this. Maybe the government overreach would happen and they'd use this for like ICE to deport people and track them and stuff. You know, government surveillance, the argument that I've heard from conservatives, why they don't like it is because they're worried that, you know, let's say like another COVID Happens and they don't want to wear their mask or they don't want to comply with some sort of government mandates that they'll be tracked or it's going to turn into like the United Kingdom Where people get, you know, put in jail for posting memes that are against that are not politically correct. So I think from both political perspectives, people don't like this company. I think their CEO did a really good job on the All In podcast of trying to defend, you know, he's just trying to make the country a safer place and really drive home that it's a felony to chop These things down. Felt like there's a little bit of scare scare tactics At the end of the day, I think they have an uphill battle right now as far as privacy goes. People don't like being tracked. And these types of features where there's, you know, like a new AI feature on the floor, like cameras rolls out is what people are concerned about. The point I think he made that was really good. And as far as like messaging goes, you know, him being like, hey, look, people, different cities can decide how long they want to save their data for if that's only 48 hours, or if that's A week, I think in the past, it had been like defaulted to a year's worth of data was saved. And people were worried about it, you know, being abused by police to stalk people or, you know, different things. And there was a number of abusers that were like using it for not good stuff, right? A dirty cop just like trying to use it for whatever reason. And so, you know, give it to the Sopranos to go track someone that they're trying to hunt down, right? So like there could be a bad use of this technology. And I think what their messaging that was really good was just saying like, hey, look, people can decide how long they want to store it for. And we don't do any facial recognition. ([Time 0:12:28](https://share.snipd.com/snip/845b908e-dab4-451a-9729-d9c317d2acdc)) - **LLM Wikis Spawn A Personal Knowledge Base Wave** - Andrej Karpathy's LLM wiki idea sparked many personal knowledge-base projects that compile indexed markdown wikis from articles, notes, and images. - Users like Casey Newton built 1,400+ page wikis by converting newsletter archives into linked MD files and summaries. Transcript: Jaeden Schafer Okay, all of this stems from a tweet that came out in April of this year. Andre said, you know, the two was titled LLM knowledge bases. And he said, quote, something I'm finding very useful recently using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code and more into manipulating knowledge stored as markdown and images the latest llms Are quite good at it he said i index source documents so he'll put in a bunch of like articles papers repos data sets images etc into a raw directory and then i use an llm to incrementally Compile a wiki which is just a collection of MD files in a directory structure. The wiki includes summaries of all of the data in raw, all of the backlinks, and then it categorizes data into concepts. It writes articles for them and it links them all. It converts web articles into MD files. I like to use the Obsidian Web Clipper extension, whatever. So he kind of goes into this idea. This was an interesting idea. When he posted it, it got about 21 million views on the tweet. It had about 60,000 people that liked it and over 108,000 people bookmarked that. ([Time 0:17:23](https://share.snipd.com/snip/69f779c7-d3d8-49d0-87fc-59074b6a04ae))