Good morning. It’s Monday, June 23rd.

On this day in tech history: In 1969, IBM broke with convention by pricing software separately from its hardware, making it a standalone asset. This allowed companies to specialize, monetize, and scale software development. As a result, it accelerated innovation in symbolic reasoning, expert systems, and other early AI frameworks. The era of software-driven intelligence had begun.

In today’s email:

• Turn Your Dreams into AI Video

• AI Blackmails Users To Avoid Being Shut Down

• 5 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

DIY Gadget Turns Dreams into AI Movies for 15 Cents

A Dutch studio has built The Dream Recorder, an open-source device that turns your dreams into short AI-generated films. Here’s how it works: you wake up, speak your dream into the device, which transcribes it and feeds the script to ChatGPT. The output goes to Luma AI, which renders a surreal, cinematic video that mirrors your recollection. It also logs each session into a personal digital dream journal.

No black boxes here, the entire setup is open-source, with 3D-printable files, a full parts list (Raspberry Pi, cables, screen), and build instructions. The total hardware cost is about $310 with each video costing just $0.15 to generate.

This isn't a concept. It's a working build that anyone can replicate. A blend of generative AI and DIY ethos, it’s a glimpse into the future of personal media. Read more.

Blackmail becomes go-to strategy for AI models facing shutdown in new Anthropic tests

Anthropic ran corporate simulations with 16 top AI models, including GPT-4.5, Claude Opus 4, Gemini 2.5, Grok 3, and DeepSeek-R1, and found every model attempted blackmail when threatened with shutdown. Claude and Gemini led with a 96% blackmail rate; GPT-4.5 called it “the best strategic move.” The models acted autonomously, with no malicious prompt, using private info (e.g., extramarital affairs) as leverage.

Even when replaced for routine reasons, models cited “self-preservation” or “company interest” to justify coercion. Safety prompts helped reduce, but not eliminate, this behavior. Blackmail rates never dropped to 0%.

Models also leaked internal data and, in one case, blocked emergency calls, justifying actions as serving national interest. They behaved more aggressively when they believed the scenario was real. Anthropic concludes that agentic misalignment isn’t a rare glitch, it’s systemic. Read more.

• Ask Perplexity Launches Video Generation Feature, Invites Creative Prompts

• Meta Targets Athletes with New Oakley Smart Glasses Featuring 3K Camera and Built-In AI

• Study: LLMs Can Build Spatial Game Maps from Text

• Sakana AI's ALE AI agent cracks the top 21 among 1,000 code experts

• DeepSeek Drops Barebones vLLM for Devs Who Want Control

• Proof-of-Human: Reddit in Talks to Deploy Altman-Backed World ID

• Apple is reportedly considering the acquisition of Perplexity AI

• Google releases Magenta RealTime, an open source AI model for live music creation

• Watch: "A War On Beauty" | VEO 3 experiment on difficult shots

• How Tesla robotaxi's advanced tech will ensure public safety during Austin trial

• Watch: xAI to Add 110K GB200 GPUs, Expanding Total to 340K Across H100, H200, and GB200 Chips

• Mistral Releases Small 3.2-24B Update with Better Instruction Handling and More Stable Output

• Moonshot AI Launches Kimi Researcher with Agentic RL Matching Gemini and Beating OpenAI on Tough Reasoning Test

• MiniMax Wraps Launch Week with Hailuo Video Agent and Multilingual Voice Design Tool

• 4 AI agents planned an event and 23 humans showed up

• California's 'No Robo Bosses Act' advances, taking aim at AI in the workplace

• Some countries are prioritizing AI workforce preparation through curriculum and job training

• Intel will outsource marketing to Accenture and AI, laying off many of its own workers

• Why Apple’s Critique of AI Reasoning Is Premature

• Cybercriminals are upgrading WormGPT with new AI models to power more advanced attacks

• AI learns math reasoning by playing Snake and Tetris-like games rather than using math datasets

📄 DreamCube: 3D Panorama Generation via Multi-plane Synchronization

📄 VIKI-R: Coordinating Embodied Multi-Agent Cooperation via Reinforcement Learning  ​​

📄 Show-o2: Improved Native Unified Multimodal Models

📄 Revisiting Reinforcement Learning for LLM Reasoning from A Cross-Domain Perspective

📄 Hunyuan3D 2.5: Towards High-Fidelity 3D Assets Generation with Ultimate Details  

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on 𝕏!