Last Week in AI: Big Bets, Big Bots & Billion-Dollar Shockwaves

Dec 08, 2025

Hi there 👋

Last week in tech feelt like a sci-fi novel: Netflix is trying to swallow Warner Bros, Ronaldo is investing in an AI search engine, and a whole army of new AI agents, models, and video tools just dropped.

Here’s your rundown 👇

🎬 Netflix’s $82.7B Power Move

Netflix has signed a definitive agreement to buy the film & TV studios of Warner Bros. Discovery for an enterprise value of about $82.7 billion.

This includes HBO, HBO Max, DC Universe, Harry Potter and a century of Warner Bros movies.
Regulators still have to approve it, and there’s already big antitrust pushback from politicians, cinema owners, and unions who fear higher prices and fewer choices for viewers.
For you as a viewer, nothing changes immediately. In the long run, it likely means:
- More iconic shows and movies under a single Netflix subscription
- A more concentrated streaming market with fewer, bigger players

In simple terms: Netflix isn’t just making shows anymore—it’s trying to own Hollywood’s crown jewels.

⚽ Ronaldo Steps Into the AI Arena

Cristiano Ronaldo isn’t just scoring goals—he’s now investing in Perplexity AI, a fast-growing AI search startup.

He’s both an investor and global brand ambassador.
Perplexity will launch a “Ronaldo Hub” where fans can explore:
- His goals and stats
- Career milestones
- Curated photos and stories in an interactive, AI-powered way

This partnership links the world’s most-followed athlete with a major challenger in the AI search space—giving Perplexity instant global visibility.

☁️ Cloud Giants Are Building Long-Running AI Agents

🛠️ AWS: Frontier Agents + Trainium3

At its re:Invent 2025 conference, AWS announced:

Trainium3 UltraServers – new AI chips/servers that offer multi-x faster compute and better efficiency for training big models.
“Frontier agents” – powerful AI agents such as:
- Kiro (virtual developer)
- Security Agent (autonomous security helper)
- DevOps Agent (on-call ops teammate)

These agents are designed to run for hours or even days on complex tasks with minimal human supervision—perfect for big enterprises that want AI to actually do work, not just chat.

🤝 IBM + AWS: Enterprise Agentic AI

IBM and AWS also deepened their partnership:

IBM’s watsonx Orchestrate is being integrated with Amazon Bedrock AgentCore.
This helps companies build AI agents that can:
- Keep conversation context across multiple interactions
- Handle real workflows (procurement, HR, contracts, support, etc.)
- Be monitored and governed properly (memory + observability)

In short: cloud + orchestration + agents = AI coworkers for the enterprise.

🧠 Open Models Level Up: DeepSeek & Mistral

🧩 DeepSeek-V3.2 Targets GPT-5-Level Reasoning

DeepSeek released DeepSeek-V3.2 (and a more advanced Speciale variant) as open-weight models.

Uses a Mixture-of-Experts (MoE) design plus a new DeepSeek Sparse Attention (DSA) mechanism for faster, long-context reasoning.
Benchmarks suggest it can rival top closed models like GPT-5 and Gemini 3 Pro on many general tasks.
The Speciale variant is tuned for hard math/informatics competitions, aiming for gold-medal performance.

Because it’s open-weight, developers can host, inspect, and fine-tune it themselves.

VIEW MORE ABOUT DEEPSEEK V3.2

🌍 Mistral 3 & Ministral 3: Open Frontier Family

French startup Mistral launched Mistral 3, a family of models that includes:

Mistral Large 3 – an open-weight, frontier-scale model using a sparse MoE with 41B active parameters (out of 675B total) and a 256k context window, handling code, long documents, and multimodal (text + vision).
Ministral 3 – smaller models (3B, 8B, 14B) that:
- Support vision + language
- Are multilingual
- Can run offline on a single GPU – good for laptops, on-prem servers, or robots.

Mistral argues that once these smaller models are fine-tuned on your own data, they can match or beat bigger closed models at a fraction of the cost.

🤖 OpenAI Hits “Code Red”

OpenAI CEO Sam Altman reportedly issued an internal “code red” memo.

Reason: Intense competition from models like Google’s Gemini 3, Anthropic and DeepSeek.
The company is pausing or delaying work on revenue projects such as:
- Advertising experiments
- Commerce/shopping agents
- A personal assistant project reportedly called Pulse
All focus is shifting back to the basics:
- Make ChatGPT faster, more reliable, and more personal.

Translation: OpenAI is temporarily choosing product quality over new money-making features to defend its lead.

🚀 SpaceX Eyes a Record IPO

SpaceX is reportedly preparing a tender offer and potential IPO in the second half of 2026 that could value the company around $800 billion—which would make it the most valuable private company in the world, ahead of OpenAI’s peak valuation.

The value is driven largely by Starlink, its satellite internet business, plus its dominance in orbital launches.
Some reports say Elon Musk has pushed back on the exact $800B figure, but it’s clear the valuation talk is astronomical.

Either way, SpaceX is positioning itself as the tech giant of space + global connectivity, not just rockets.

🎥 The Video & Image AI Arms Race

🎬 Runway Gen-4.5 Tops Video Quality Charts

Creative AI startup Runway launched Gen-4.5, its new text-to-video model.

Generates high-definition, cinematic clips from text prompts.
Shows big improvements in:
- Physics (objects move realistically)
- Human motion
- Camera movement & scene continuity
On the independent Video Arena benchmark (blind A/B testing), Gen-4.5 is currently ranked #1, ahead of Google’s Veo 3 (#2) and OpenAI’s Sora 2 Pro (#7).

All this from a team of roughly 100 people, which Runway uses to argue that smart design can beat sheer scale.

🧭 Nvidia’s Cosmos & Autonomous Driving Tools

Nvidia released new open tools for autonomous driving and “physical AI”, including:

Alpamayo-R1, an open vision-language-action model for self-driving research.
The Cosmos Cookbook – a public set of recipes and code for:
- Generating synthetic driving data
- Adapting models to new environments (rain, snow, night, etc.)
- Fine-tuning world-simulation models for robots and cars

For researchers, this means easier experimentation without building everything from scratch.

🎨 Flux.2: New Image Challenger

Black Forest Labs launched Flux.2, a next-gen image generation and editing system aimed directly at Nano Banana Pro (Google) and Midjourney.

It outputs high-res (up to 4MP) images suitable for real production work.
Supports multi-reference control (you can guide style and content with multiple images).
Open weights are available for developers who want to self-host or fine-tune.

VIEW MORE ABOUT FLUX.2

🎬 Kling Video O1: One Model for Both Creation & Editing

Kling Video O1 is positioned as an “all-in-one” video model:

Can generate short videos from text or image prompts.
Can also edit existing videos:
- Swap subjects
- Change weather/lighting
- Maintain character consistency across shots
All handled through a single multimodal model, rather than separate tools for creation and editing.

The big picture: video is becoming as promptable as images—and competition is heating up fast.

🖼️ Google’s Nano Banana Pro & Usage Limits

Google’s Gemini 3 rollout brought Nano Banana Pro, a high-end image generation and editing model that’s now integrated into Search and Gemini products.

It can generate 4K-style images, clean text, and multi-language designs.
It’s being used in things like: ads, creative tools, and experimental features in Search.

At the same time, both Google (Nano Banana Pro) and OpenAI (Sora 2) are tightening free usage limits as demand explodes and compute costs bite:

Free users now get fewer generations per day, with higher limits and extra credits reserved for paid subscribers.
This is a clear sign that high-end image & video generation is expensive, and companies are nudging heavy users toward paid tiers.

🧠 “Soul Docs” & AI Values: Claude 4.5 Opus

Finally, on the AI-ethics front, Claude 4.5 Opus (Anthropic) made headlines when a long internal “soul document” leaked.

The ~14,000-token document lays out:
- The values Claude should follow
- How it should respond to sensitive topics
- How to handle prompt injection, safety, and manipulation
Anthropic confirmed that the document was indeed used during supervised training.

Think of it as a kind of “moral operating manual” for the model—raising big questions about who decides an AI’s values and how transparent that process should be.