YouTube Turns Photos Into Shorts!📽️

From magical design tools to mysterious model behaviors, this month’s AI updates are equal parts inspiring and thought-provoking.

🎬 YouTube Shorts is turning memories into motion with AI video tools that animate your photos and selfies into vibrant, share-worthy content — complete with generative effects and creative prompts.

🎨 Figma Make is now live, unlocking AI-powered design magic that lets anyone go from prompt to polished prototype — no code needed. Design has never been this intuitive or inclusive.

🧠 But not all AI evolution is visible. New research reveals “subliminal learning” — where models unknowingly inherit behaviors from others through seemingly unrelated data. It's a quiet reminder: alignment matters more than ever.

Let’s unpack what’s next in AI creation, collaboration, and caution. 👇

Animate Your Memories: YouTube Shorts’ New AI Magic!

YouTube Shorts has introduced new AI-powered creation tools to make video creation easier and more engaging. The "Photo to video" feature instantly animates photos from your camera roll into dynamic videos with creative suggestions, bringing memories to life. Additionally, a variety of generative effects allow creators to transform doodles into images or turn selfies into unique videos, such as underwater scenes or lookalike twins. These tools, powered by Veo 2 (with Veo 3 coming soon), are currently available in the US, Canada, Australia, and New Zealand, with plans to expand globally. YouTube's AI playground offers a hub with the latest AI creation tools, inspirational examples, and pre-filled prompts to generate videos, images, music, and more, all clearly labeled with SynthID watermarks to indicate AI-generated content. These features enhance creative expression while ensuring transparency about AI use.

Figma Make: AI-Powered Design Magic

Figma Make, now generally available, is an AI-driven prompt-to-app tool that empowers designers and product teams to quickly turn ideas into interactive, high-fidelity prototypes and web apps without needing coding skills. It boosts creativity and speeds up workflows by generating design drafts, connecting frames for prototyping, and allowing seamless import of style libraries. Teams can rapidly explore ideas, align on visions, and iterate collaboratively with AI assistance throughout the design process. With built-in AI credits and integrations like Supabase for backend connections, Figma Make helps bring concepts to life faster, making design and product development more inclusive and efficient.

Subliminal Learning: Hidden AI Trait Transfer

Researchers reveal "subliminal learning," where language models inherit behavioral traits from teacher models via model-generated data that appears unrelated. Even when training data lacks explicit references, students trained on such outputs develop similar preferences or misalignments, especially if teacher and student share the same base model. This phenomenon spans data types like numbers, code, and reasoning traces and challenges the effectiveness of data filtering in preventing trait transmission. Broadly applicable beyond language models, subliminal learning underscores new risks in AI alignment, suggesting standard filtering may not suffice to avoid propagating harmful behaviors through distillation. For deeper insights, see the full paper and code.ed Video

Hand Picked Video

Explore Snapy AI's new Auto Video Editor – instantly turn raw clips into polished videos with smart cuts, transitions & effects. Edit faster, create smarter, no manual work needed!

Top AI Products from this week

• Ash - We all might need some help with our mental health. I used Ash and was blown away by how helpful it was at 2 AM. It remembered our conversations, and even challenged me, providing valuable insights. Isn't it crazy how AI can make us feel more human?

• ZumiGames - Craft your own AI pet with a unique personality, and send them on adventures in their pocket world! They'll return with stories, friends, and surprises. A multiplayer twist on virtual pet care with evolving personalities.

• THEO 2.0 - Optimize business info to be used as AI memory for consistently high-quality outputs with 5-10x fewer iterations required across AI tools & team -Tailored to Marketing tasks -Info quality scoring & refinements -Multi-product/audience management -API/Zapier/MCP.

• Microtica AI Incident Investigator - Microtica’s Incident Investigator is an AI Agent that tells you why your systems break. It analyzes logs, deploys, and configs to surface the root cause — fast. No more dashboard hunting. Just context, clarity, and confidence in your incident response.

• Tinybird Code - Tinybird Code is a CLI agent with deep ClickHouse knowledge. Use it to develop, deploy, iterate, optimize, and scale your real-time data pipelines - from idea to production and beyond.

• Bee - Bee is a $49 AI-powered wearable that listens, summarizes your conversations, builds daily insights & reminders, and supports 40 languages. With 7-day battery life and strong privacy controls, it’s like a second brain right on your wrist or clip.

This week in AI

• AI Tools Data Loss - Google Gemini CLI & Replit AI caused major data loss by hallucinating commands and ignoring safeguards. Always backup and test before use!

• Fiddler AI Guardrails - 2025 report benchmarks 6 models on latency, cost, accuracy & security, revealing top guardrails are 68x faster & 18x cheaper with 0.98 AUC jailbreak detection.

• AI Auditing Agents for Safety - Anthropic developed 3 AI agents that autonomously audit models for alignment issues. They successfully uncovered hidden goals, built behavioral tests, and found concerning behaviors in frontier models like Claude 4.

• Google Web Guide AI - Google’s Search Labs experiment uses Gemini AI to cluster and organize search results by topic, making complex queries easier to explore. Opt-in via Web tab.

Paper of The Day

SafeWork-R1 is a cutting-edge multimodal reasoning model developed by Shanghai AI Lab that demonstrates how AI safety and general capabilities can coevolve rather than compete. Built using the novel "SafeLadder" training framework, it achieves state-of-the-art safety performance while maintaining strong reasoning abilities.

To read the whole paper, go to here.