Welcome, humans.

Mary Meeker—the legendary tech analyst who called the internet boom—just dropped a 339-page AI report that's denser than a black hole and twice as mind-bending.

Here’s a few spoilers for ya: ChatGPT's growth (800M users in 17 months) makes Google look like a tortoise, and inference costs (how much it costs to “prompt” an AI) fell 99.7% in the past two years.

Translation? AI = accelerating like nothing before. We’ll break down the report in full for y’all tomorrow.

Now, ever wondered how influencers would handle the end of the world? The Dor Brothers, who use AI to make some pretty genius satirical videos, just made an “influenders” supercut with Google’s new video model Veo 3 that feels a LITTLE too real… (in more ways than one).

That’s not the only viral Veo 3 influencer video we saw this weekend—apparently, these “Historical / “Bible times” vids of famous Biblical “influencers” are going bonkers on TikTok.

This trend has gone so viral, REAL influencers are now pretending to be “100% AI Veo 3” themselves. Y’know, for the views.

How viral are we talking here? Sir Demis Hassabis (CEO of DeepMind, who is in fact knighted) said “millions of videos” have been made with Veo 3 over the past few days.

If you didn’t know, rn YouTube gets about ~2.6M videos uploaded daily, while TikTokers upload ~34M vids a day. Veo 3 is still under the radar for most ppl—but imagine how quickly human-generated content will be overrun by AI content if this trend continues… (it could be “the end of the world internet as we know it…”)

Here’s what you need to know about AI today:

• ChatGPT discovered a Linux security flaw cybersecurity experts missed.

• UMG, Warner, and Sony moved to license catalogs to AI music providers.

• Meta’s AI to replace nearly all of its product risk reviews.

• Researchers introduced an AI that can self-replicate itself.

AI just found its first real zero-day vulnerability—and security researchers are freaking out…

One of the biggest risks of widespread AI is the ability for anyone with a ChatGPT account (or worse, ChatGPT itself) to find and exploit code vulnerabilities at scale. Well, that day just arrived—and the implications are wild.

Security researcher and Oxford PhD Sean Heelan just revealed how OpenAI's o3 model discovered CVE-2025-37899, a previously unknown remote vulnerability in the Linux kernel's SMB implementation.

Quick translation: This is like finding a secret backdoor into millions of computers worldwide that lets hackers break in from anywhere on the internet—in the most secure part of the operating system that powers everything from your WiFi router to Amazon's servers.

Matt Johansen breaks it down brilliantly in his video analysis: “Remote zero day in the Linux kernel is a very serious string of words.”

Here's what makes this breakthrough so significant:

Sean wasn't just telling o3 to “go find bugs” when he found this issue. He was benchmarking the model using vulnerabilities he'd already discovered manually, when o3 surprised him by finding a completely different bug—one that even HE had missed.

It’s all a bit technical, but basically, o3 found a catastrophic flaw that only a few dozen people on Earth could spot:

• When given 3K lines of targeted code, o3 found known vulnerabilities 8% of the time.

• With 12K lines of code, success dropped to 1%—but that's when it found the previously unknown vulnerability.

• The catch? It generated false positives 28% of the time.

What's particularly fascinating is o3's bug report quality. Not only did it find the vulnerability, but it also correctly identified why Sean's proposed fix for a similar bug would've been insufficient—spotting a glitch (race condition) he missed.

Sean's takeaway says it all: 

Why this matters: AI could actually become cybersecurity researchers’ most powerful tool. Companies spending millions on bug bounties should take note: the economics of vulnerability discovery just changed dramatically.

The current signal-to-noise ratio (1:50) means you'll wade through false positives, but as Sean notes, “had I used o3 to find and fix the original vulnerability, I would have, in theory, done a better job than without it.”

The scary part? If researchers are using AI to find vulnerabilities, so are the bad guys. 

So what should regular people do? First, don’t panic: A patch to the Linux kernel has already been committed and merged into the official Linux kernel repository. Unless you’re managing Linux servers, this shouldn’t impact you directly.

As Matt concludes in his analysis, “Security is always an arms race, and we just got access to a stronger, more capable arm.” He's excited rather than scared—because the same AI helping find these bugs will help patch them faster too.

For vulnerability researchers, the message is clear: start integrating these tools now or risk being left behind. For the rest of us? Make sure automatic updates are turned on, and buckle up—the age of AI-powered cyber-warfare has begun.

FROM OUR PARTNERS

Build AI that actually works for your business 

There's a huge gap between flashy AI demos and real value. Enter Agents from Retool—the easiest way to leverage AI into meaningful action by connecting directly to your business systems.

With Retool, you get:

• Model flexibility: Use OpenAI, Anthropic, Azure, or custom models.

• True integration: Works with existing databases, APIs, and business tools.

• Enterprise security: Fine-grained RBAC and permissions built in.

• High-level control: Visual builder meets code flexibility.

Experience a proven platform trusted by 10,000+ companies, from startups to Fortune 500s. Ramp saved $8M and increased efficiency 20%.

Start building your agent workforce.

Prompt Tip of the Day

Want a prompt straight from a research scientist at DeepMind to help you learn through “Socratic tutoring”? Shared by Dwarkesh Patel (of the Dwarkesh pod), this prompt gets your AI to keep “asking you probing questions which reveal how superficial your understanding is.”

Just start a new chat with your favorite AI model, copy and paste this prompt in, then ask it what you want help learning!

Treats To Try.

1. In addition to running DeepSeek’s new R1 model on your own computer (at a lower quality, mind you), you can also run it on OpenRouter for free.

2. Google AI Edge is a platform enabling developers to build AI features that will run directly on mobile devices and web apps, while the Edge Gallery includes models you can download and run on your Android phone locally (so private and offline).

3. PrettyPrompt is a Chrome extension that optimizes your AI prompts to get more useful responses (like Grammarly, but for prompts).

4. OpenPaper makes it really easy to upload, read, and chat with research papers, and even cites where it got the information from in the original paper.

5. Duckbill uses human assistants and AI to handle your life admin like booking appointments, paying tickets, and finding gifts so you get your weekends back—paid only rn.

6. Imagen 3 (Google’s last gen image generator) lets you create a limited quota of free images on AI Studio; if you need images and hit your limit w/ GPT, try it out.

See our top 51 AI Tools for Business here!

Around the Horn.

• Major record labels (UMG, Warner, Sony) neared a settlement with AI music providers Suno and Udio, and are now in talks to license their catalogs to the AI.

• Meta planned to automate 90% of all risk assessments (like new safety features), although Meta says only “low risk decisions” will be automated.

• Anthropic, maker of Claude, hit $3B in annualized revenue (which is total monthly revenue over 12 months)—that’s up from $2B back in March.

• AI powered roll-ups is becoming a new trend—that is, buying “people-intensive” professional service firms (law firms, healthcare practices, customer service) and using AI to increase efficiency (i.e. the classic “PE haircut” of lowering headcount), raise margins from 10% to 40%, then use the gains to roll-up more.

• Three new AI studies to check out:

  1. AI optimized for user feedback learned to manipulatively target vulnerable users with behaviors that were hard to mitigate or detect, distorting their reasoning.

  2. Researchers showed state-of-the-art AI are easily manipulated by a universal jailbreak, fueling “dark LLM” risks.

  3. Researchers introduced the Darwin Gödel Machine, an AI system that can self-rewrite and improve its own code—and has surpassed other self-improving systems.

  4. For more, check out the top AI papers of last week here. 

FROM OUR PARTNERS

Are you building a rat’s nest?

Building your own data processing pipeline starts simple—but scaling it is another story. What begins as a few  scripts and connectors quickly turns into a tangled mess of never-ending fixes and updates. Unstructured replaces the DIY rat’s nest so you can focus on AI innovations.

Get started today!

Monday Meme

A Cat's Commentary.