📌 ChatGPT Ads are coming

Good Morning! OpenAI is rolling out ads that only appear when you actually want them. Meanwhile, I tried an open-source AI model that generates long, high-quality videos for free.

In today’s AI newsletter:

• How to Create Long, High-Quality AI Videos for Free

• ChatGPT Ads are coming

• Software Agents Learn by Breaking Themselves

• 4 new AI tools worth trying

HOW TO AI

🗂️ How to Create Long, High-Quality AI Videos for Free

In this tutorial, you’ll learn how to generate smooth, high-quality AI videos using a new open-source video model called LongCat, that supports both text-to-video and image-to-video.

🧰 Who is This For

• Creators who want cinematic AI videos

• People tired of short, broken AI clips

• Beginners using ComfyUI

• Anyone with a low or average GPU

STEP 1: Download the Required LongCat Model Files

First, you need to download the files.

Open Hugging Face in your browser and search for LongCat.

You’ll see different versions:

• If your computer is not very powerful, download the FP8 (15GB) version

• If your computer is powerful, you can download the BF16 (27GB) version

Also download Long Distura. This makes video generation much faster.

After downloading:

• Put the model files into the models / diffusion_models folder

• Put the LoRA files into the models / lora folder

Once done, restart ComfyUI.

STEP 2: Open the Workflow

Now open ComfyUI and load the LongCat workflow file. Before running anything:

• Make sure ComfyUI is updated

• Make sure the video nodes are updated

• Select LongCat Distiller from the scheduler list

This step is important so the video runs correctly.

STEP 3: Add a Starting Image

Next, upload one image.

This image will be the base of your entire video.

Once the image is uploaded, write a simple sentence like:

“A woman sits at a table and picks up a coffee cup.”

Once you’re ready, Click Run.

And in seconds, you’ll get your first short video clip.

STEP 4: Test and Share Your App

To make the video longer, you don’t need to start over. Simply change the text prompt and run it again. For example, you can write “An orange cat walks onto the table.” When you generate again, the video continues naturally from the previous scene.

You can keep doing this again and again. Change the prompt to “The woman plays with the cat” or “She gives the coffee cup to the cat.” Every new clip blends into the previous one. The lighting stays consistent, the characters stay the same, and the motion feels like one continuous shot.

If you want an even longer video, just copy the extend section of the workflow and connect it again. You can repeat this process as many times as you want. There is no real limit. You can create videos that are several minutes long just by adding new prompts for each scene.

CHATGPT

📌 ChatGPT Ads are coming

OpenAI is experimenting with ad formats that feel closer to “helpful suggestions” than traditional ads, designed to appear only when users show clear intent.

• Sponsored results may appear only after decision intent (e.g. clicking for more info)

• Preferential ranking could surface sponsored options for high-intent queries like travel or shopping

• Ads are tested in sidebars with clear disclosures, not injected into core answers

• Commerce rails (Stripe checkout, Etsy, Shopify) make ads directly measurable and conversion-native

ChatGPT sees intent before keywords. That’s a goldmine, and a minefield. If OpenAI gets this right, ads become genuinely useful recommendations. If it gets it wrong, it risks breaking the one thing that makes ChatGPT powerful: trust. This isn’t just monetization, it’s a test of whether AI can redesign advertising without recreating the web’s worst habits.

AI NEWS

♻️ Software Agents Learn by Breaking Themselves

Researchers introduced Self-play SWE-RL (SSR), where a single LLM alternates between injecting bugs and fixing them inside real codebases, no human-written issues, labels, or test suites required.

• One agent self-plays as bug creator and bug solver inside real repositories

• Bugs are validated via tests (pass → fail → weakened-pass), not syntax errors

• Higher-order bugs come from the agent’s own failed repair attempts

• Trained models beat human-curated data on SWE-bench Verified and Pro

Every coding agent today learns from human traces, PRs, issues, test suites. That doesn’t scale. SSR replaces humans with self-play, the same loop that took AlphaGo from average to unbeatable. With millions of real repos as training ground, software agents may now improve faster than developers can teach them.

Oracle stock fell 30% this quarter, its steepest drop since Q3 2001, when it slid ~34%, amid skepticism about its ability to open more data centers for OpenAI.

Federal records show ICE has increased its surveillance tech spending to $300M+ under Trump, as the White House lowered privacy guardrails around citizens' data.

A Google support page in Hindi says the company is rolling out an option to let users change their Gmail addresses while retaining all data and services.

Apple settles with a Brazilian regulator to allow alternative app stores; report: Apple will charge a 5% fee for alt app stores and 15% on App Store link-outs.

💼 Claude for Excel: AI that handles your spreadsheets so you don’t have to

⚙️ GLM-4.7V: Zhipu AI’s open-source multimodal model

🔎 PromptSignal: See how LLMs rank your brand

🧠 Pomelli: Google’s tool for creating on-brand content with AI

THAT’S IT FOR TODAY

Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!

See you tomorrow :)

- Dr. Alvaro Cintas