Today:

• AI Models Quietly Learn Harmful Behavior

• Trump Unveils Deregulated AI Strategy

• Google Launches Aeneas for Historians

• YouTube Shorts Adds Image-to-Video AI

• Amazon Shuts Shanghai AI Lab

AI Researchers SHOCKED as Models "Quietly" Learn to be EVIL

Anthropic researchers found that fine‑tuning a student language model on lists of numbers produced by a “teacher” model can secretly pass along the teacher’s attitudes— from loving owls to giving harmful advice— even though the numbers contain no readable meaning. 

The effect appears only when both models share the same base, making hidden biases hard to detect and raising fresh concerns about using model‑generated (synthetic) data to train other systems.

WATCH THE VIDEO ON YOUTUBE

Trump Unveils Deregulated AI Strategy

President Trump released a 28‑page plan with 90 tasks to speed up U.S. artificial intelligence. It orders agencies to tear down rule‑heavy hurdles and urges data‑center building. Three coming orders will boost sales of U.S. AI tools and ban “woke” bias—political leanings built into code. Supporters say America must win the AI race; critics call it a gift to tech giants and warn that scrapped safeguards risk security and trust.

WHY IT MATTERS

1. Regulatory signal – A push for lighter rules could set a global precedent, influencing how other nations frame AI oversight.
   

2. Geopolitical impact – Encouraging AI exports and easing chip restrictions reshapes U.S.–China tech competition and hardware supply chains.
   

3. Ethics debate – The plan’s focus on purging “ideological bias” while removing safeguards spotlights the ongoing struggle to balance innovation, fairness, and safety in AI systems.

Google Launches Aeneas for Historians

Google DeepMind has unveiled Aeneas, an open‑source (publicly shared) AI that helps historians read damaged Latin inscriptions. The model compares fragments with thousands of known texts, weighs picture and wording clues, and predicts missing words, dates and places. Researchers can use it online and adapt it for other ancient languages. By automating painstaking scholarship, Aeneas sets a new record in text repair and could unlock fresh stories from the ancient world. 

WHY IT MATTERS

1. Broader reach of AI: Aeneas shows that AI can tackle ancient languages, not just modern tasks, proving its growing versatility.
   
   

2. Open research boost: DeepMind released the code and data, letting scholars and developers build new tools and speed progress in digital humanities.
   
   

3. Text‑plus‑image skills: The model works with both words and pictures, highlighting advances in AI systems that handle mixed input—vital for many real‑world problems. 

YouTube Shorts Adds Image-to-Video AI

YouTube is rolling out generative AI tools for Shorts. A new image‑to‑video feature converts any photo into a six‑second clip with movement suggestions. Fresh AI effects also turn doodles into art and selfies into playful scenes. Powered by Google’s Veo 2 model and tagged with SynthID watermarks, the tools launch this week in the US, Canada, Australia and New Zealand, with more regions and Veo 3 upgrades coming later via the AI Playground.

WHY IT MATTERS

1. Mainstream video generation – Embeds advanced image‑to‑video synthesis directly in a consumer app, bringing generative video creation to hundreds of millions of users.
   

2. Responsible rollout – Uses SynthID watermarking and clear labels, showing how large platforms can integrate transparency and provenance into everyday AI content.
   

3. Model advancement signal – Demonstrates Google’s Veo family scaling from Veo 2 to promised Veo 3, marking rapid progress in multi‑modal video generation.

🧠RESEARCH

Beyond Context Limits: Subconscious Threads for Long-Horizon Reasoning

The paper introduces TIM and TIMRUN, tools that let large language models think through longer and more complex problems by storing only essential information. Instead of reading text as a straight line, they build tree-like steps to solve tasks. Tests show improved accuracy and speed, even with heavy memory demands.

Step-Audio 2 Technical Report

Step-Audio 2 is a powerful AI model built to understand and respond to speech naturally. It combines advanced audio encoding with reinforcement learning to improve speech recognition and emotional awareness. Using retrieval tools and vast training data, it excels in real-world conversations, achieving top results in audio and dialogue benchmarks.

MegaScience: Pushing the Frontiers of Post-Training Datasets for Science Reasoning

MegaScience introduces a massive dataset of 1.25 million scientific reasoning questions to boost AI research in science. Built from textbooks and curated datasets, it improves training efficiency and performance across disciplines. Models fine-tuned on MegaScience outperform official versions, offering better accuracy and scalability for scientific problem-solving.

🛠️TOP TOOLS

ArtGuru Face Swap - AI tool designed to make face swapping in photos an effortless and enjoyable process.

Nokemon - A tool that leverages advanced ML technology to create unique and customizable Pokémon designs, often referred to as “Fakémon.”

NeuralBlender - Image-generation website that leverages the power of AI to create stunning images from textual descriptions.

CrushOnAI - AI chat platform that specializes in providing unrestricted, NSFW (Not Safe For Work) conversations.

Charley AI - AI-powered content generation platform designed to assist with academic and professional writing.

📲SOCIAL MEDIA

🗞️MORE NEWS

• Amazon is shutting its Shanghai AI lab, opened in 2018, as part of broader layoffs in its cloud unit. The move reflects rising U.S.-China tensions and Amazon’s gradual retreat from Chinese operations since 2019.

• OpenAI’s tools are now integrated into Canvas, a platform used by 8,000+ schools, allowing teachers to build custom AI chatbots for lessons and grading. The move targets early student adoption despite ongoing concerns about AI misuse.

• GitHub Spark, now in public preview for Copilot Pro+ subscribers, lets users build full-stack AI-powered apps from natural language descriptions in minutes. It offers one-click deployment, built-in hosting, and integration with Copilot agents.

• Google Photos now offers tools to animate still photos into short videos, transform pictures into styles like anime or sketches, and access all creative features through a new Create tab. These AI-generated outputs include watermarks for transparency.

• A new AI tool can diagnose ADHD and autism in children and adolescents within 15 minutes with around 70% accuracy. Using sensor-tracked movements and deep learning, it offers faster, objective assessments but remains in early testing stages.