• AI Emergence
  • Posts
  • The Ghibli Effect, $40B Moves & AI That Can Dunk 🍿

The Ghibli Effect, $40B Moves & AI That Can Dunk 🍿

Along with: Runway Gen-4 Is Here- and It’s Cinematic

Hi there đź‘‹

1 million people signed up on OpenAI within the first hour of launch. That’s wild!

What stood out again was OpenAI’s ability to move fast and put out products that are meaningful to the users before anyone else. This time, it started with GPT-4o’s image generation- especially the Ghibli-style artwork that flooded the internet with dreamy, hand-drawn visuals.

The speed of innovation- and how quickly it goes viral- is only getting crazier.

Fun fact: It seems most of the new signups came from India. The trend is real!

Let’s dive into this week’s updates.

What would be the format? Every week, we will break the newsletter into the following sections:

  • The Input - All about recent developments in AI

  • The Tools - Interesting finds and launches

  • The Algorithm - Resources for learning

  • The Output - Our reflection

Table of Contents

Amazon just introduced Nova Act, part of its Nova family of models- designed to carry out real tasks inside a web browser. Think filling forms, booking appointments, buying groceries, or planning events. It’s not just about generating responses- Nova Act can actually take action on your behalf.

At the core, Nova Act is built to handle complex multi-step tasks by breaking them down into smaller, verifiable browser actions. For example, it doesn’t just “buy a product,” it searches, selects, fills out forms, skips unnecessary add-ons, and completes the checkout. All in a reliable, step-by-step way.

It runs on Amazon’s Nova models (Micro, Lite, Pro) and plugs into Amazon Bedrock, so developers can scale from prototype to production quickly. The SDK is available now in research preview, and you can customize it using Python- adding specific instructions, logic, and even API calls or Playwright integrations.

What makes it stand out?

  • Higher success rates (90%+ on tricky UI tasks like dropdowns)

  • Designed for multi-step workflows

  • More customizable than most agent frameworks out there

And up to 75% cheaper than alternatives (source)

Gen-4 is designed to improve consistency across frames- aiming to deliver more stable characters, objects, and scenes without needing fine-tuning.

What’s new:

  • Scene Consistency: You can lock in a specific character, object, or visual style, and the model tries to maintain that look across different angles, lighting, and environments.

  • Image + Text Prompts: The model supports combining text instructions with reference images to guide the overall tone and style.

  • Shot Coverage: You provide a reference and describe the scene- Gen-4 attempts to generate multiple perspectives with built-in scene understanding.

  • Smoother Outputs: There’s a noticeable improvement in motion, object tracking, and frame-to-frame alignment, making it more viable for narrative and commercial use cases.

  • VFX Capabilities: It also includes generative visual effects designed to integrate better with live-action or animation workflows.

Under the hood:

Gen-4 focuses on performance and prompt responsiveness, while still giving creators room to experiment visually. It’s positioned as a tool for everything from product shoots to short films- and like with all generative models, the output will vary based on prompts and context.

If you’ve been exploring AI for storytelling or video workflows, it might be worth a try to see how it fits into your creative process. (source)

Google DeepMind has introduced Gemini Robotics, a new family of multimodal AI models designed to power real-world robots. These models allow robots to perform complex tasks- like folding origami, making salads, or slam dunking a basketball- with zero prior training on those specific actions or objects.

Built on Gemini 2.0 and fine-tuned with robot-specific data, the models bring together vision, language, and physical action in one system. With Gemini Robotics-ER (Embodied Reasoning), robots can recognize objects, reason about tasks, and even generate code to complete actions- all on the fly.

What makes this a leap forward:

  • Dexterous control: Smooth, multi-step motion across varied tasks.

  • Generalization: Trained on a wide range of tasks, not just one.

  • Adaptability: Works across robot types, from research bots like ALOHA to humanoids like Apollo.

According to DeepMind, this marks a major step toward robots that understand, reason, and act- just like humans do, making them useful in homes, factories, and everything in between. (source)

OpenAI Plans Open-Weight Language Model Release

OpenAI is set to release its first open-weight language model since GPT-2, focused on strong reasoning capabilities. Launch is expected in the coming months.

Before release, the model will undergo OpenAI’s Preparedness Framework evaluation, with extra steps taken to account for potential modifications post-release.

To shape the model’s design and usefulness, OpenAI is hosting developer feedback events- starting in San Francisco, then expanding to Europe and APAC. (source)

Zhipu AI just introduced AutoGLM Rumination, a free AI agent designed for web-based tasks like research, travel planning, and report writing- powered by its in-house GLM-Z1-Air and GLM-4-Air-0414 models. It’s entering a hot AI agent market and already turning heads.

How It Stands Out:

  • Matches DeepSeek R1 in reasoning performance

  • Runs 8x faster, with 1/30th the compute

  • Fully free, unlike competitors like Manus, which charges $199/month

Key Features of AutoGLM Rumination

  • Autonomous Web Agent: Can search, synthesize, and generate structured content across web environments.

  • Powered by Proprietary Models: Uses GLM-Z1-Air for speed + efficiency, and GLM-4 for high-performance outputs.

  • Full Access for Free: Available via Zhipu’s mobile app and model site with no paywall.

  • Strong Benchmark Performance: Claims GLM-4 outperforms GPT-4 on multiple benchmarks.

  • Backed by Major Investment: Recently secured 300M yuan ($41.5M) from the city of Chengdu, part of a flurry of government support.

With China’s AI agent race heating up, Zhipu’s free, fast, and capable offering puts pressure on paid alternatives. (source)

OpenAI has secured a $40 billion funding round, the largest private tech deal ever, valuing the company at $300 billion. Led by SoftBank ($30B) with support from Microsoft, Coatue, Altimeter, and Thrive, the round nearly triples the previous largest private raise.

A significant portion (~$18B) is expected to go toward Stargate, OpenAI’s compute infrastructure venture with SoftBank and Oracle. However, the full investment hinges on OpenAI restructuring into a for-profit entity by Dec 31, adding pressure amid legal challenges and regulatory hurdles.

OpenAI now trails only SpaceX in private tech valuation and claims 500M weekly ChatGPT users, with projected 2025 revenue of $12.7B. The deal signals strong confidence in OpenAI’s continued dominance as the AI arms race accelerates. (source)

xAI has officially acquired X in an all-stock transaction, valuing xAI at $80B and X at $33B (net of $12B debt).

Founded just two years ago, xAI has quickly risen as a top AI lab, building cutting-edge models and infrastructure at scale. X, with 600M+ active users, has evolved into a highly efficient digital platform for real-time information.

The merger combines xAI’s AI expertise with X’s massive distribution, aiming to deliver smarter, more impactful experiences and accelerate innovation. Together, they’ll unite data, compute, models, and talent to build a platform focused on truth, knowledge, and progress. (source)

Apple is revamping its Health app with a built-in AI coach that offers personalized health advice, according to Bloomberg’s Mark Gurman. The feature- tentatively called Health+- is expected to launch with iOS 19.4, possibly in spring or summer 2026.

The AI coach will use data from users' medical devices and may include food tracking. It's currently being trained on insights from staff physicians, with Apple planning to expand training using recorded videos from additional doctors. (source)

Turn Text into Visuals Instantly with Napkin

Napkin lets you turn plain text into clear, compelling visuals in seconds- perfect for anyone who wants to communicate ideas faster and more effectively. No design skills, no prompting- just paste, generate, and share.

How to Use:

  1. Paste Your Text – Just drop in your content; no prompt engineering required.

  2. Generate Visuals – Napkin auto-creates relevant visuals; pick the one that best fits your message.

  3. Edit & Customize – Tweak layout, style, or content to make it your own.

  4. Export Anywhere – Download as PNG, PDF, or SVG to use across slides, docs, or the web.

Use it for: Infographics, diagrams, flowcharts, and more- perfect for engaging presentations and clear communication.

  • OpenAI just launched Academy, a free learning hub aimed at making AI knowledge accessible to everyone- from curious beginners to seasoned engineers. It offers online courses, live workshops, and in-person events, covering everything from AI basics to advanced integrations. Built in collaboration with institutions like Georgia Tech and Miami Dade College, the platform is designed to help users confidently apply AI in their work, life, and communities.

  • Claude’s Thinking, Now Visible: Anthropic has built an AI microscope to see how Claude thinks, step by step. In these two papers(Paper1, Paper2), they show how replacing parts of Claude’s neural network with simpler, interpretable components reveals its reasoning process- marking a major step toward AI transparency and safety.

  • Last week, we launched more free AI & ML courses to help you dive deeper into agent workflows and retrieval systems:

    • Building Data Analyst AI Agent- Learn how to automate data analysis with AI agents. This course breaks down core frameworks like CrewAI and Autogen, then walks you through building a functional agent for real-world data tasks. Great for both beginners and working pros.

    • Building and Evaluating RAG Systems- Go beyond the basics of Retrieval-Augmented Generation. You’ll learn data ingestion, embedding, and retrieval, then level up with reranking techniques, evaluation methods, and scalable architectures for production-ready RAG systems.

It has been one of the most crucial weeks in AI in terms of humans adopting AI tools. 

Now the real question is- how many of them will go beyond generating pretty images and actually start using AI meaningfully?

What are your thoughts? 

Until next time..

Reply

or to participate.