Your AI is lying and you don't know it

Happy Sunday,

If you read Wednesday's newsletter you'd know that Claude just dropped the most powerful AI model on the entire planet. It wasn't even out for 48 hours and the US government has already shut it down. And most people have absolutely no idea why this is happening. I'm also breaking down when to use Claude Chat vs. Code vs. Cowork, my tool of the week that changed how I prompt AI, a pro tip for catching when AI starts lying to you, the Visa x ChatGPT integration that just changed shopping, and three Claude finance agents Anthropic just launched that you need to try.

Let's get into it.

🚨 THIS IS SERIOUS

The US Government Just Shut Down the Most Powerful AI Model Ever Made

I need you to actually read this one. Last Monday Anthropic released Claude Fable 5... a Mythos-class model and one of the most powerful AIs ever built and released to the public. I covered it in Wednesday's email. Three days later, Amazon's cybersecurity team found a jailbreak in the model that could potentially be used for cyberattacks. Amazon's CEO personally contacted the White House about it. And within hours the Commerce Department sent Anthropic a letter telling them to immediately cut off all access. Anthropic couldn't separate American users from foreign users fast enough... so they just turned everything off. For everyone.

The US government used export control laws... the same laws they use to restrict weapons and military technology... to shut down an AI model. That has never happened before. They've blocked chip exports. They've restricted semiconductor manufacturing. But this is the first time they've ever looked at an AI and said "this is too dangerous to exist." Anthropic is fighting it in court. There are also reports that a China-linked group may have already accessed Mythos, which could be the real reason behind all of this.

We are living through the moment where AI got so powerful that a government treated it like a weapon. Whether they overreacted or made the right call... the fact that we're even having this conversation tells you everything about where this is headed.

Read Anthropic's official statement →

🔬 AI PRO TIP

How to Catch When AI Is Lying to You

If you use ChatGPT or Claude a lot you've probably noticed that sometimes it starts off great and then slowly gets worse the longer your conversation goes. It doesn't just go from perfect to hallucinating overnight. It's gradual. First it starts ignoring small instructions you gave it. Then it starts making assumptions. Then all of a sudden it's giving you completely wrong information and you don't even realize it happened.

Here's why this happens. These models have something called a context window - basically how much of your conversation they can hold in memory at once. The longer your chat goes, the more it has to juggle. At some point it starts dropping things. It forgets instructions you gave it 20 messages ago. It starts prioritizing your most recent messages and losing the earlier context. That's when the quality drops and the hallucinations creep in. Most people just keep going and don't realize the session is cooked.

Peter Steinberger... the guy behind OpenClaw, one of the biggest open-source AI agent platforms out there... found a trick that basically stops this. It's called canaries. You know how coal miners used to bring canaries into mines? If the canary stopped singing, there was toxic gas and you needed to get out. Same idea. You add one tiny instruction telling ChatGPT or Claude to start every single response with your name. That's your canary. As long as the AI keeps using your name... you're good. The moment it stops? That's your signal the session is degrading and it's time to start a new chat before it starts making stuff up.

Add that to the start of your prompt or your custom instructions. It takes two seconds to set up and it'll save you from trusting an output that's completely off. This is one of the best pro tips I've shared in a while - seriously go try this in your next session.

Full guide on canaries and why they work →

💡 SAVE THIS

Claude Chat vs. Code vs. Cowork (When to Use Each)

One of the most common questions I get is "Claude has all these different products now... which one am I supposed to use?" Let me break it down really simply because once you get this you'll use Claude completely differently.

Claude Chat is your thinking partner. This is the one you already know - claude.ai, the app, the regular chat. Use it when you need to brainstorm ideas, write emails, understand complex topics, or plan something out. You upload a file and ask questions about it. You talk through a problem. Think of it like having a really smart friend you can bounce ideas off of.

Claude Code is for building. This one runs in your terminal and can read your entire codebase, edit files, run commands, and create pull requests. If you need to build an app, fix bugs, write documentation, or automate anything technical... this is it. It doesn't just tell you what to do. It actually does it.

Claude Cowork is for delegating. This is for anyone who wants Claude to actually do the work for you - not just talk about it. It runs on your desktop and can access your files, connect to your email and Slack, organize documents, create reports and presentations, and run tasks on a schedule. You give it a goal and it goes and does it. Pro tip - use the Google Chrome extension inside Claude Cowork so it can actually control your Chrome and automate tasks for you. (This is how you can automate your emails, schedule posts, fill out forms - definitely worth checking out.)

Don't just stick to using one. I use all three every single day for different things and they each do something the others can't.

Get the full Claude Chat vs. Code vs. Cowork guide →

🌟 TOOL OF THE WEEK

Wispr Flow: Talk to AI Instead of Typing

I've been telling people I don't really type anymore and they always give me this look like... what are you talking about? So let me explain. There's a tool called Wispr Flow that I've been using that instantly changed how I interact with AI. This section is not sponsored at all - I genuinely love this tool and think everyone should be using it. You just talk out loud and it turns your speech into clean, structured text. Not like Apple's built-in dictation that gives you every "um" and half-finished sentence. Wispr Flow uses AI to clean up what you're saying as you talk. Removes the filler, adds punctuation, restructures everything so it reads like you sat down and carefully typed it out.

You know that feeling where your thoughts are running through your head way faster than your fingers can keep up? You end up writing a shorter prompt because you don't feel like typing it all out. And then the AI gives you a mid answer because you gave it a mid prompt. With Wispr Flow you just talk and it handles the rest. You can give ChatGPT or Claude way more context in way less time because speaking is like 3x faster than typing. I use it for everything now - emails, Slack, prompts, even this newsletter.

Here's why I think this 10xs your AI outputs. Most people don't realize how these AI models actually process your prompts. They understand structured text way better than a big blob of words. When Wispr Flow cleans up your speech it doesn't just fix grammar... it structures your text with clear sections, bullet points, and organized thoughts. That makes a massive difference because now the AI can actually parse what you're asking for instead of trying to decode a wall of text. It's the difference between handing someone a messy pile of notes and handing them a clean brief. You can use it for free but I pay for it and it's worth every penny.

Try Wispr Flow →

💳 AI UPDATE

Visa x ChatGPT Just Changed Online Shopping Forever

Visa just made AI shopping real. They partnered with ChatGPT and embedded their entire payment network directly inside it. So instead of going to Google, clicking through a bunch of results, reading reviews, comparing prices across ten different tabs... you just tell ChatGPT what you need. It shops across every merchant that accepts Visa, finds the best option, compares prices, and actually buys it on your credit card (you have to approve it first). ChatGPT has taken extensive safety measures to make sure this is secure - your real card number is never shared with merchants and every transaction goes through Visa's tokenization.

I totally understand not wanting to give AI your credit card... I'm not saying you should yet. But at what point do we decide that this technology has gotten so good that it's going to make our lives that much easier and we can trust it? I don't think we're too far away.

If you're a business owner this should be a huge wake-up call. When people start shopping through AI instead of Google... the AI is choosing what to recommend. Businesses that are showing up inside these models are the ones the agent is going to choose over the ones that aren't. That's just how it works - the agent looks at what it knows about, and if your business isn't in there, it's not getting recommended. There are multiple ways you can fix this but I tell most people my go-to tool is Cytd AI. It essentially trains the AI on your business so you actually start showing up in these answers. If you haven't checked it out definitely go try it out because there's a good chance you're losing customers every single day to this.

💰 FREE RESOURCE

3 Claude Finance Agents That Will Change How You Look at Money

Anthropic just launched a set of finance agents inside Claude and these are incredible. These are plugins inside Claude Cowork and Claude Code you can start using right away. There are ten total but here are the three I've been using the most.

The Market Researcher

Tracks sector developments, synthesizes news, filings, and broker research - instead of spending an hour reading five different articles you get everything in one place in about 30 seconds.

The Model Builder

Creates and maintains financial models from filings and data feeds - ask it to model different scenarios and it adjusts everything automatically.

The Earnings Reviewer

Reads earnings transcripts, updates models, and flags the changes that actually matter - no more reading 40-page transcripts trying to figure out what the CEO was really saying.

If you want to install all of these I put together a guide on exactly how to set them up. Takes three seconds. Go check it out.

Get the full finance agents guide (free) →

Try those canaries in your next AI session and go test out Wispr Flow. Between the two your AI responses are going to get 10x better. I promise.

See you Wednesday.

Maverick

P.S. If you're a business owner looking to actually implement AI into your workflows, my team and I work with a small number of people and businesses to build custom AI systems, automations, and tools. If that sounds like you, apply here.

P.P.S. If you're new here or you ever skip one, I've got a newsletter archive of all my past emails you can catch up on anytime.