⏱ 4-min read
Today's stories share a quiet shift: from cloud → local, prompt → structure, text → multimodal. Google's Gemma 4 runs on a 16GB laptop. Reve 2.0 turns image prompts into editable layouts. HeyGen built a brand spec for video. Google Labs is generating personal stories from your inbox.
The era of "send a prompt and hope" is ending.
Let’s get into it. ↓
Google Just Made Multimodal AI Small Enough for Your Laptop
Google introduced Gemma 4 12B, a new open multimodal model built to run locally on everyday laptops. It handles text, vision, and native audio without separate multimodal encoders — cutting memory needs while keeping reasoning close to Google’s larger 26B MoE model.
Image source: Google
Highlights
• Native audio input for offline transcription, formatting, and translation
• Runs locally with 16GB VRAM or unified memory
• Benchmark performance near the larger 26B MoE model
• Apache 2.0 license with support for Hugging Face, Ollama, LM Studio, MLX, vLLM, llama.cpp, and more
• Multi-Token Prediction drafters to reduce latency
Why it matters: the next wave of AI apps may not start in the cloud. They may start on your laptop.
In Partnership with Cuey
Have you gotten burned by a confident-sounding AI answer yet?
Claude, ChatGPT, and Gemini are brilliant and for casual work, one answer is plenty. But for high stakes research and decisions, a confident-sounding answer is a risk most users can't afford to take.
Cuey sends one prompt to ChatGPT, Claude & Gemini in a single tab. Spot hallucinations. Cross-check reasoning side-by-side.
Medicine Was Looking at Obesity. The AI Looked at the Whole Body.
Researchers at Helmholtz Munich built MouseMapper, an AI that maps an entire body in 3D, cell by cell. When they used it to study obesity, it found unexpected damage to facial sensory nerves and inflammation across multiple organ systems.
The latest issue of AIHealthTech Insider: Issue #103 unpacks how MouseMapper saw what traditional analysis missed, why whole-body AI mapping could change disease research, and the bigger question: what else has medicine been missing because nobody could see all of it at once?
Want it weekly?
Bring OOH Into the Modern Marketing Stack
AdQuick makes Out Of Home advertising approachable, measurable, and performance-focused. Designed for marketers at startups and large brands alike, it combines digital efficiency with real-world reach—so your campaigns always hit the mark.
🔥Big News in AI
Image source: Ideogram
🔹Ideogram 4.0 Just Made AI Design More Precise
Ideogram 4.0 brings dense text rendering, native 2K images, transparent backgrounds, and tighter layout control. It was trained with bounding boxes tied to region descriptions, so the model understands where text, objects, and design elements belong. Logos, posters, typography, and product visuals just got closer to real design software.
🔹HeyGen Just Gave AI Video Its Own Brand Spec
HeyGen introduced frame.md, a new spec for branded videos and motion. design.md helped agents keep websites and decks consistent, but video needed different instructions. frame.md teaches agents how brand systems should move — including pacing, framing, transitions, scenes, and visual rhythm. Static brand guidelines just became motion-aware.
🔹Google Just Built a Personalized Story App for Your Life
Google Labs introduced Dreambeans, an experimental mobile app that connects to your Google apps using Personal Intelligence. Every day, it creates personalized story collections from things you might miss, plus topics you already care about. Your inbox, calendar, and digital life just became a daily narrative feed.
Your best prompts are the ones you'd never bother typing.
The detailed ones. The ones with examples and edge cases. Wispr Flow lets you speak them instead — clean, structured, ready to paste into any AI tool. Free on Mac, Windows, and iPhone.
Reve 2.0 Just Turned AI Images Into Editable Layouts
While most image models still treat prompts like magic spells, Reve is betting on structure.
Reve 2.0 generates 4K images using layouts, where every object, region, and design element is segmented, labeled, and editable. Images become more like code — addressable, movable, and controllable.
Image created with Reve
The real question: when every part of an AI image can be edited like a webpage, does prompt engineering become design engineering?
Codex Added Pets to Codex
OpenAI added pets. Tiny, animated creatures that live in the corner of the app, thinking when your code processes, celebrating when it compiles, hiding when it errors.
You can create a completely custom animated sprite with just one prompt. It takes only two minutes, and it will respond to your code indefinitely. The full guide — 5 steps, 8 ready-to-copy prompts — is on Medium.
AI Job Postings Are Quietly Breaking Away From the Rest of the Market.
The new Stanford HAI AI Index shows AI-related skills still appear in a small share of U.S. job postings but that small share is growing much faster than the broader market. Additional data from the Bipartisan Policy Center shows AI-skill postings grew at triple-digit rates in early 2026, while overall job postings grew in single digits.
This Sunday Special breaks down the Stanford findings, why AI roles are growing while hiring gets more automated, the $60K Google Cloud hackathon closing June 11, and the free courses worth your weekend.
Want it every Sunday?
New here? Get the next issue Sunday.
70,000+ AI builders read The AI Entrepreneurs every week.
