🎬 OpenAI shuts down Sora video app

Good Morning! OpenAI is officially killing its viral AI video app, Sora, just six months after launch to rein in massive computing costs. Plus, you’ll learn how to generate studio-quality product photos for free.

Plus, in today’s AI newsletter:

• OpenAI Shuts Down Sora Video App

• LiteLLM Compromised in Cascading Supply Chain Attack

• LeCun’s Team Cracks Open World Models

• How to Generate Studio-Quality Product Photos

• 4 new AI tools worth trying

AI NEWS

🎬 OpenAI Shuts Down Sora Video App

Six months after launching to massive hype and hitting a million downloads in under five days, OpenAI is abruptly pulling the plug on its Sora short-form video app. The move is part of a massive cost-cutting campaign as the company pivots away from expensive consumer experiments to focus purely on enterprise dominance.

• The app allowed users to generate, remix, and share AI videos in a social feed, but user retention dropped as the astronomical compute costs to render video at scale piled up.

• In the wake of the shutdown, Disney has officially backed out of its massive $1 billion investment deal that would have allowed Sora users to generate content with its copyrighted characters.

• OpenAI is pivoting aggressively toward high-productivity enterprise use cases to compete with Anthropic’s booming Claude business.

• Instead of maintaining a standalone video app, OpenAI plans to fold its core tech into a singular desktop "super app" featuring its web browser, ChatGPT, and Codex coding tools.

Viral consumer buzz doesn't equal a sustainable business model when the product burns through thousands of dollars of compute per hour. By killing a fan-favorite app and walking away from Disney's massive investment, OpenAI is showing just how intense the pressure is to achieve profitability and win the enterprise war before they go public.

AI NEWS

🚨 LiteLLM Compromised in Cascading Supply Chain Attack

Two versions of LiteLLM (v1.82.7 and v1.82.8) have been abruptly removed from the Python Package Index (PyPI) after attackers injected malicious code designed to steal developer credentials.

• The breach actually originated from Trivy, an open-source vulnerability scanner used in LiteLLM's CI/CD pipeline, which was compromised by a threat group known as TeamPCP.

• The attackers modified existing version tags in Trivy's GitHub Actions, allowing them to silently inject malicious code into organizations' workflows without changing the underlying release metadata.

• Using this backdoor, the attackers intercepted LiteLLM's PyPI publishing token and pushed new, poisoned packages containing a credential-stealing component file (litellm_init.pth).

• To cover their tracks, the attackers even spammed the project's GitHub vulnerability report with AI-generated "Thanks, that helped!" comments to bury legitimate warnings from security researchers.

This is a textbook, terrifying example of a cascading supply chain attack. A trusted security tool gets compromised, which is then used to backdoor a widely used AI framework, instantly exposing the credentials of any developer or enterprise that blindly installs the update. As AI development moves at breakneck speed, securing the CI/CD pipeline and pinning dependency versions is becoming just as critical as the code itself.

WORLD MODELS

🌍 LeCun’s Team Cracks Open World Models

Yann LeCun’s team just released LeWorldModel, a 15-million parameter model that successfully predicts physical reality. While LLMs only predict the next word, world models predict actual physics, objects moving, colliding, and falling, laying the foundation for robots and autonomous AI to plan and act in the real world.

• Solves the notorious "collapse" problem (where world models cheat by mapping every input to the same output) without needing complex duct-tape fixes like frozen encoders or stop-gradient hacks.

• Massively simplifies training by using just two loss terms (a prediction loss and a "SIGReg" regularizer to keep representations diverse), reducing the required hyperparameters from six down to just one.

• Incredibly efficient: Trains on a single GPU in a few hours, plans 48x faster than foundation-model alternatives, and uses roughly 200x fewer tokens.

• Proves that the JEPA architecture LeCun has been pushing since 2022 actually trains stably without breaking.

Two AI futures are competing right now: building bigger text-based LLMs, or building world models that actually understand cause, effect, and physics from raw pixels. LeWorldModel proves that the second path isn't just theory, it works, it's stable, and it just got drastically cheaper and more accessible to build.

HOW TO AI

🗂️ How to Generate Studio-Quality Product Photos for Free

In this tutorial, you will learn how to use Pomelli, an experimental AI marketing tool from Google Labs, to extract your brand's unique "Business DNA" and transform basic smartphone pictures into professional, studio-grade product photography without writing a single prompt.

🧰 Who is This For

• People selling products online (Shopify, Amazon, D2C brands)

• Dropshippers who need clean product visuals fast

• Small business owners without a studio setup

• Content creators shooting product promos

STEP 1: Extract Your Business DNA

Head over to Pomelli and sign in. Instead of manually uploading brand kits or writing complex style prompts, Pomelli automates the setup. Enter your business or product website URL and click continue. The AI will scan your site to extract your "Business DNA”, automatically identifying your brand's color palette, typography, tone of voice, and visual style. If your website blocks AI scrapers, you may need to ask your developer to adjust your site permissions first.

STEP 2: Access the Photoshoot Feature

Once your Business DNA is successfully generated, navigate to the "Photoshoot" tool. Upload a basic, unedited photo of your product, even a messy smartphone snapshot will work perfectly. Because Pomelli already understands your brand's visual identity from the previous step, it does not require any complex prompting or prompt engineering to get started.

STEP 3: Select Your Template and Generate

After uploading your image, Pomelli will analyze the product and recommend a variety of curated templates, such as contextual settings, lifestyle shots, or seasonal themes. Choose an aspect ratio optimized for your specific social feed (like 4:5 for an Instagram post, 9:16 for Stories, or 1:1 for square feeds) and click generate.

In just a few minutes, the AI will apply your Business DNA to the template, outputting studio-quality images that perfectly match your brand's exact aesthetic.

STEP 4: Turn Photos Into Full Campaigns

Once your images are generated, you can download them directly or save them back to your Business DNA for future use. To take it a step further, click on the campaign feature and type a simple goal, like "Father's Day 20% discount." Pomelli will stitch your newly generated product photos into fully designed, ready-to-post social media layouts with matching text and calls to action, saving you hours of manual editing.

Anthropic announces an “auto mode” that enables Claude Code to make permission-level decisions while preventing destructive actions like mass file deletion.

Arm unveils its own AI chip called the AGI CPU, a departure from its traditional role as a designer of chips for others; Meta and OpenAI will be early customers.

Epic cuts 1,000+ jobs, saying it is “spending significantly more” than it is making and that the layoffs and $500M in savings put it “in a more stable place”.

Munich-based Interloom, which aims to capture tacit knowledge for AI agents from businesses' operational records, raised a $16.5M seed led by DN Capital.

🦞 NemoClaw: Nvidia’s open-source security layer for AI agents

🤖 ASMR: Supermemory’s experimental agent memory system

🎨 Uni-1: Luma’s all-in-one model for text + images

🎥 Replit Animation: Turn text prompts into animated videos

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