Good morning. It’s Wednesday, Jun 4th.

On this day in tech history: 1998: Microsoft launched Windows 98, codenamed "Memphis," succeeding Windows 95 with improved hardware support and Internet Explorer 4.0 integration. This OS helped advance AI by providing a stable platform for tools like SNNS (Stuttgart Neural Network Simulator) and CLIPS (expert system framework), enabling researchers to develop early neural networks.

In today’s email:

• The AI Tools for Generating Video

• Apple Finally Joins the AI Race

• Deepmind’s Force Prompting

• 5 New AI Tools

• Latest AI Research Papers

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

The AI Tools for Generating Video: Between Concept and Creation

AI content tools are rapidly reshaping how images, video, and audio are created, edited, and monetized, shifting the focus from manual production to automated, scalable pipelines.

Topaz Labs’ Bloom takes on image resolution, offering an AI-powered upscaler that boosts output by up to 8× without sacrificing core details. Its five tuning modes, from Subtle to Max, give users granular control over stylization, generating four outputs per pass to accelerate visual iteration. In parallel, Captions’ Mirage Studio introduces digital actors with micro expressive realism, rendered in minutes from text, audio, or images via an omni-modal foundation model built for fluid performance. ManusAI extends that vision with a scene-planning video engine that maps, designs, and animates entire narratives from a single prompt, bringing storyboard-to-screen automation to a new level.

As video formats become more modular and monetizable, OpusClip’s OpusSearch automatically identifies and repurposes high-performing moments from existing footage, transforming passive archives into scalable content pipelines. Creatify’s AdMax tackles paid media at scale, using TikTok and Instagram trend data to reverse-engineer viral formats into AI-generated ad variations.

On the audio front, PlayAI’s PlayDiffusion replaces autoregressive editing with parallel speech inpainting, using a 20-step denoising loop to restore natural pacing and tone at lower computational cost. Rounding it out, Character.AI evolves chat into performance: animated avatars, real-time scene direction, and group streaming push beyond reactive dialogue toward scripted, persistent character arcs, soon joined by redesigned profiles and a social feed. Meanwhile, HeyGen’s AI Studio tightens control over avatar performance with tools like gesture mapping, speech mirroring, and voice-director modes, layering production polish onto a fully synthetic cast.

Apple reportedly tests AI models that match ChatGPT's capabilities in internal benchmarks

Apple plans to open its small on-device AI models, reportedly with around 3 billion parameters, to third-party developers at WWDC 2025, enabling basic features such as text summarization. These compact models run directly on Apple hardware but are far less capable than larger in-house models used internally.

Apple is internally testing cloud-based foundation models with 3, 7, 33, and 150 billion parameters; the largest matches ChatGPT’s performance in benchmarks, according to employees. Yet these advanced systems remain confined to the company’s “Playground” testbed, with no timeline for public release. Insiders cite executive hesitancy over accuracy and hallucinations, leaving Apple to quietly license OpenAI’s chatbot instead. Meanwhile, splashier projects, including a chatty new Siri, AI-enhanced Shortcuts, and a health assistant dubbed “Mulberry,” have slipped to 2026 or later. Read more.

Deepmind's "force prompting" lets AI create realistic video motion without physics engines

Researchers from DeepMind and Brown University introduced force prompting, a technique that enables generative video models to simulate realistic motion using directional force inputs, without relying on physics engines or 3D modeling. Users define global or local forces as vector fields., wind across a scene or a tap on a point, which the system processes through a ControlNet-integrated Transformer model built on CogVideoX-5B-I2V.

The model, trained for one day on four Nvidia A100 GPUs, uses synthetic datasets: 15,000 clips of wind-driven flags, 12,000 of rolling balls, and 11,000 of flower impacts. It generalizes well, preserving physical plausibility and style features, while learning intuitive concepts like mass. Despite outperforming text-based and motion-path baselines, and even PhysDreamer in realism and alignment with input forces, it still falters in complex scenarios. Force prompting offers a lightweight way to steer AI-generated motion through intuitive physical cues. Read more.

• Elon Musk says 𝕏Chat is rolling out to all

• Microsoft Bing Adds Free Sora Video Generator to Mobile App

• Anthropic Shows Off Claude’s Writing—But Keeps the Code on a Leash

• OpenAI Unveils Two Major Updates, Advances Global Expansion Play

• Meta collects by far the most data to train their models

• Eye-opening device: Self-powered AI synapse mimics human vision, achieves 82% accuracy

• Monaco entrepreneur launches world’s first real-time AI generated film

• Yoshua Bengio launches LawZero to develop safe AI systems free from commercial influence

• Supercomputer shows black hole cracking neutron star in final explosive seconds

• MIT AI digs through 1 million samples to find 19 materials that can replace cement

• World’s first mass-produced nuclear reactor to enter testing in US

• Fi’s AI-powered dog collar lets you monitor pet behavior via Apple Watch

• Washington Post Plans AI-Led Coaching for Non Professional Writers

• Meta becomes the latest big tech company turning to nuclear power for AI needs

• MIT Spinout Builds AI Tool to Detect Model Uncertainty

• Marc Andreessen: US Must Focus on Building Robots, Not Reviving Old Manufacturing Jobs

• Cohere and Arc Institute Launch BioReason AI, Predicting Disease Pathways with 97% Accuracy Using DNA and LLM Fusion

• ChatGPT has fun new incel problems

• Thomson Reuters is building AI agents with OpenAI

• AI-generated CUDA kernels outperform PyTorch in several GPU-heavy machine learning benchmarks

📄 Beyond the 80/20 Rule: High-Entropy Minority Tokens Drive Effective Reinforcement Learning for LLM Reasoning

📄 UniWorld: High-Resolution Semantic Encoders for Unified Visual Understanding and Generation ​​

📄 VS-Bench: Evaluating VLMs for Strategic Reasoning and Decision-Making in Multi-Agent Environments

📄 Time Blindness: Why Video-Language Models Can't See What Humans Can?

📄 Taming LLMs by Scaling Learning Rates with Gradient Grouping

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!