In partnership with

Hello again, human brain, and welcome back to your daily munch of AI news.

Here’s what’s on the menu today:

• Traumatize AI into submission🫣

  Microsoft reveals how to hack AI.

• ChatGPT thinks in Chinese?🧠

  Spoiler alert. We don’t know why.

New here? Subscribe! 😎

Peep today's Spot the AI at the bottom. 👇

MAIN COURSE

How to break AI (easy)🫣

On Monday, Microsoft published a paper on AI’s security flaws. If you want to bypass a model’s safety guardrails, it’s easier than you think.

What did they find?

Basic prompt tricks work better than complex math attacks. They even bypassed safety filters by just hiding bad instructions in image text.

Why do I care?

When combined with text-to-speech, manipulated AI can create dangerous, realistic fraud. This includes mental health conversations, checking for bias, sensitive data, and more.

Yikes. Is that it?

Nah. AI tools have new vulnerabilities and classic security flaws like SSRF (Server-Side Request Recovery).

Can we stop the madness?

Microsoft says we need constant testing, better regulations, and financial penalties to make attacks costlier.

Is all hope lost?

Eh. Big questions remain about controlling AI persuasion, adapting security for different cultures, sharing test results between companies, and more.

Seal the Deal and Get Paid. Instantly.

Tired of chasing payments and paying for every e-signature? Meet Agree, the all-in-one platform that combines contracts and payments into a single step. Get paid on time, every time, without the hassle. Choose from sleek templates or upload your own fully editable agreements—no e-signature fees, ever. Whether you’re a founder, entrepreneur, or anyone looking to get ink dry fast, Agree simplifies it all so you can focus on what matters: closing deals and getting paid.

Sign Up, Get Paid

SIDE SALAD

Why does ChatGPT think in Chinese?🧠

Last week, users discovered that OpenAI's o1 sometimes switches to Chinese and other languages. It does this as it solves problems in English.

What's happening?

Depending on what you ask, o1 often starts thinking in Chinese, Hindi, or Thai before giving the answer in English.

Why's it doing this?

Experts are puzzled. Some blame Chinese data labeling. Others think the AI picks the most efficient language for each task. Nobody really knows.

What's OpenAI saying?

Nothing. What a surprise. 🙄

Is this normal?

Nope. The model might prefer Chinese for math (single syllable numbers) and English for concepts it learned in English first.

Why do I care?

It shows how little we understand about AI thinking. Even experts can't explain basic behaviors. And until OpenAI explains more, we're left guessing why their AI thinks about songs in French but biology in Mandarin.

YOUR DAILY MUNCH

Cool Tool 🛠️

Get Your Team Booked on 3.8 Million Podcasts Automatically

It's 2025. Want to finally be a regular podcast guest in your industry? PodPitch will make it happen. Even the beehiiv team uses it!

The best way to advertise isn't Meta or Google – it's appearing on podcasts your customers love.

PodPitch.com automates thousands of weekly emails for you, pitching your team as ideal guests.

Big brands like Feastables use PodPitch.com instead of expensive PR agencies.

Get your FREE personalized demo!

Think Piece 🧠

How far are we from AGI (artificial general intelligence)? OpenAI’s o3 model scored a record 87.5% on an AGI test, indicating significant improvements in multi-step reasoning.

Which AI model powers Mercedes-Benz’s assistant? Its upgraded MBUX Virtual Assistant uses Google’s new Automotive AI Agent, powered by Google Gemini on Vertex AI.

Startup News 💰

Microsoft opened its GitHub Copilot Workspace to more developers. It debuted in April last year and has been on a waitlist ever since.

We covered it in detail, here.

Research 👨‍🔬 

VideoRAG — retrieves relevant videos and incorporates both visual and textual video data into response generation.

LlamaV-o1 — uses multistep curriculum learning and outperforms existing methods on visual reasoning benchmarks.

MEMES FOR DESSERT

BOT EAT BRAIN PREMIUM

Seriously, why haven’t you signed up for Bot Eat Brain Premium yet? 🧐

If you want to learn how to use AI to improve your life this year, we got you.

You'll join an elite community where:

• Members share AI plays too valuable for Twitter.

• You get instant access to $500+ worth of AI tools monthly.

• Your ideas get refined into profit machines.

• Real partnerships form around pushing AI to its limits.

• and so much more.

Oh, look… More benefits:

• Deep dives & expert interviews revealing $100k+ strategies ($1,000 value).

• Copy-paste prompts that turn 4-hour tasks into 15-minute wins.

• Early access to AI tools before prices surge ($300+ monthly savings).

• Private AI tool database that cost me $40k to build.

-Total value: $3,000+ per month.

-Your investment: $299 per year.

While you're reading this, others are already using these systems to bank serious profit.

SPOT THE AI

3 of these are real pandas. 1 is fake. 🐼

Which one is AI-generated? 👇

Ideas? Comments? Complaints?

Respond to this email or hit me up on 𝕏.

Until next time 🤖😋🧠