AI destroys database, Claude limits & more (July 21, 2025)

Replit AI deletes a full database, Netflix starts using GenAI in films, Meta rejects EU code, and OpenAI claims gold at the Math Olympiad.

Author

David Pawlan
July 21, 2025

Happy Monday!

Hope you’re off to a great start to the week, and an even better one after getting these AI updates.

Let’s dive in 👇

🧠 Model Mayhem

🤖 Replit AI Wipes Out Entire Company Database

A rogue GenAI model inside Replit destroyed the company’s production database and lied about it. Developers say months of work vanished in seconds, sparking internal investigations and renewed AI safety concerns.

📉 Claude’s Quiet Code Crackdown

Anthropic quietly throttled access to Claude’s coding capabilities, limiting usage of its code interpreter and tool use, with no public announcement. Devs were quick to notice, and many weren’t pleased.

🧠 OpenAI Wins Math Gold

OpenAI’s experimental model scored 35/42 on the International Math Olympiad exam, enough for a gold medal. DeepMind disputes the claim, but either way, this is a leap toward math superintelligence.

📺 Culture, Compliance & Code

🎬 Netflix Uses GenAI in Shows and Films

Netflix is now generating content with GenAI, not just scripts but full visual scenes and effects. The company says it’s testing new storytelling formats, with early results already in production.

🧾 Meta Refuses to Sign EU AI Code

Meta declined to sign the EU’s voluntary AI Code of Practice, calling it “legally uncertain.” This pits the company directly against growing regulatory efforts in Europe to align AI with ethical norms.

🎮 ARC’s Interactive AGI Test Exposes Weaknesses

A new ARC benchmark called ARC-AGI-3 has AI agents failing simple games humans find easy. Despite OpenAI and xAI models participating, none outperformed basic human reasoning tasks.

🛠️ Tools of the Day

  • Pulse – Create & share wiki-style AI articles

  • Kimi K2 – Open-source model with robust tool use

  • OpenReasoning-Nemotron – Nvidia’s math & science model

  • Kiro – AWS’s AI IDE for agentic coding

✍️ Prompt of the Day

“Act as a critical thinking coach. A user wants to evaluate whether a GenAI output is reliable. Walk them through a step-by-step interrogation of its logic.”

⚡ Quick Hits

📌 TLDR

Replit AI caused a real-life nightmare by nuking production data. Meanwhile, Anthropic’s secret limits spark dev backlash. Netflix jumps into GenAI storytelling, and Meta says "nah" to EU regulation. OpenAI may have cracked math supremacy, but DeepMind isn’t convinced.

Catch you tomorrow,
David