Today:

• LeCun Unveils JEPA AI Model

• OpenAI Offers To Buy Chrome

• Character AI Launches AvatarFX Tool

• Physical Intelligence Unveils π0.5 Robot

• Sam Altman Steps Down From Oklo

Meta's AI Chief: "I'm DONE with LLMs"

Yann LeCun says large language models (LLMs) won’t lead to true AI. Speaking at Nvidia's conference, he introduced an alternative: JEPA, a new model that learns how the world works through visual experience, not just text. 

Unlike LLMs, which rely on words or "tokens," JEPA builds internal mental models to reason and plan like humans. LeCun believes real intelligence needs this kind of deeper, more physical understanding.

WATCH THE VIDEO ON YOUTUBE

OpenAI Offers To Buy Chrome

OpenAI told a U.S. judge it would happily buy Chrome if Google is forced to sell. The firm says Google’s browser and search data block rivals, hurting ChatGPT’s own future search product and its chances of being pre‑installed on phones. OpenAI also offers Apple a cut of ChatGPT revenue to deepen Siri integration. The Justice Department argues that breaking Google’s grip and sharing data will spur fairer, faster artificial‑intelligence competition.

Why it matters

1. If the court makes Google share its search data, AI firms would gain richer training material and could build better tools.
   

2. Control of a major browser would let an AI company reach users directly, gather real‑time feedback, and shape how people search online.
   

3. OpenAI’s revenue‑sharing pitch to Apple hints at how AI services can make money while embedding themselves in everyday devices.

Character AI Launches AvatarFX Tool

AvatarFX lets anyone turn a single image and voice clip into a lifelike talking video. Its advanced diffusion‑based model syncs speech, movement, and facial emotion across long scenes, even when several characters take turns. Smart training tricks and fast inference keep visuals crisp and smooth. Built‑in audio tools offer many voices, so stories feel fresh and interactive. The result is an easy, real‑time platform for expressive, personalized storytelling.

Why it matters

1. Pushes video synthesis forward – Mixing cutting‑edge diffusion with audio conditioning shows a clear path toward longer, more coherent AI‑generated films.
   

2. Democratizes character animation – Users need only a picture and voice, lowering technical barriers for creators, educators, and marketers.
   

3. Highlights safety challenges – Realistic talking avatars raise new questions about deepfakes and consent, pressing the industry to refine trust‑and‑safety tools.

Physical Intelligence Unveils π0.5 Robot

π0.5 is a new robot brain that learns from mixed data—images, text, and motion—to see, read, and act in messy homes it has never visited. By first planning a step in language and then sending joint commands, it clears kitchens, makes beds, and obeys voice prompts. Co‑training with web images and data from many robot types lets it spot new objects and reuse skills, narrowing the gap to human flexibility.

Why it matters

1. Proves robots can adapt to real‑world homes, a key step toward useful household assistants.
   

2. Merges seeing, language understanding, and movement in one model, setting a template for embodied foundation AI.
   

3. Shows that adding web visuals and varied robot data greatly boosts generalization, guiding smarter data strategies for future systems.

🧠RESEARCH

Learning to Reason under Off-Policy Guidance

LUFFY trains big reasoning models using both their own trial‑and‑error steps and outside example solutions. A smart weighting method prevents blind copying while still encouraging fresh ideas. This mixed training boosts math and logic scores more than copying alone and helps the models solve new, unseen problems with better generalization.

Eagle 2.5: Boosting Long-Context Post-Training for Frontier Vision-Language Models

Eagle 2.5 is an advanced AI that understands both pictures and words across very long videos and large images. New training tricks keep detail and context while cutting computation. A fresh 110‑thousand‑video dataset improves learning. The 8‑billion‑parameter model matches top commercial systems in comprehension, but with far fewer resources and cost.

FlowReasoner: Reinforcing Query-Level Meta-Agents

FlowReasoner is a master AI that builds a fresh team of small helper bots for every question a user asks. It learns by trial‑and‑error rewards that weigh result quality, speed, and simplicity. Trained from DeepSeek R1 knowledge, it beats earlier systems by over 10%, proving tailored agent teams boost accuracy.

🛠️TOP TOOLS

Logo Redesign AI - AI-powered tool that transform and enhance your brand’s visual identity with ease and efficiency.

Copyleaks AI - Designed to identify and flag text generated by AI models.

Figurative Language Checker - Scans your text and ultimately identifies figurative language elements like similes, metaphors, and personification.

Neural Frames - Audioreactive AI animations for musicians, creatives and visual artists.

Augie AI - Designed to make video-first marketing accessible and intuitive for businesses of all sizes.

📲SOCIAL MEDIA

🗞️MORE NEWS

• OpenAI CEO Sam Altman quit as chairman of young nuclear (atom‑based) power firm Oklo, clearing conflict that blocked joint work. Oklo says the move eases future deals with ChatGPT maker; its shares fell sharply afterward.

• New Oscar rules say films using generative artificial intelligence can win awards. The Academy stresses human input matters. Generative means software that creates images, voices or scripts from text. Voters must watch every nominated movie.

• Two‑person startup Nari Labs released Dia, free software that turns written words into lifelike voices. The 1.6‑billion‑setting model, trained on Google TPU chips, handles emotions, speaker changes and directions, beating rivals like ElevenLabs in tests.

• Anthropic’s security chief says autonomous AI “virtual employees” with logins and duties may appear in workplaces within a year, ultimately forcing firms to redesign security, monitor AI accounts, and decide who answers for misbehavior.