- NATURAL 20
- Posts
- Meta Acquires Agent Startup
Meta Acquires Agent Startup
PLUS: My 2025 AI recap: the 5 moments that actually moved the world
Wes Roth
January 02, 2026
In partnership with
The Future of Shopping? AI + Actual Humans.
AI has changed how consumers shop by speeding up research. But one thing hasn’t changed: shoppers still trust people more than AI.
Levanta’s new Affiliate 3.0 Consumer Report reveals a major shift in how shoppers blend AI tools with human influence. Consumers use AI to explore options, but when it comes time to buy, they still turn to creators, communities, and real experiences to validate their decisions.
The data shows:
Only 10% of shoppers buy through AI-recommended links
87% discover products through creators, blogs, or communities they trust
Human sources like reviews and creators rank higher in trust than AI recommendations
The most effective brands are combining AI discovery with authentic human influence to drive measurable conversions.
Affiliate marketing isn’t being replaced by AI, it’s being amplified by it.
TODAY:
Meta just did the thing
Meta has acquired Manus AI, a fast-growing startup known for building AI agents that operate like remote workers on their own virtual Linux machines. These agents can execute complex tasks via command-line, such as 3D renders or interactive dashboards, without needing fancy interfaces. Manus topped the Remote Labor Index with a 2.5% task completion rate—better than GPT-5 and Gemini, though still far from human-level.
This signals Meta’s pivot toward powerful, scalable AI agents. Led by Alexander Wang, the move shows Meta’s new AGI strategy: less open source, more focus on practical capabilities. The AI race is heating up fast.
Hey!
Before we sprint into 2026, I wanted to hit pause and do a quick “what just happened?” on AI in 2025.
Not the noisy stuff. The structural stuff — the releases and rules that changed what’s possible (and what’s allowed). Here are the five stories I think mattered most, in the order I’d tell a friend over coffee.
August 7, 2025
Do you remember where you were in August? I do, because that’s when OpenAI finally dropped GPT-5. After all the rumors and the waiting, they delivered something that felt... different.
It wasn't just a chatbot anymore. They introduced this "unified system" concept that I found fascinating. Instead of you having to choose between a fast model or a smart one, GPT-5 just knew. It had this built-in router that could decide, "Okay, I need to think hard about this one," or "I can answer this instantly."
And the capabilities? It crushed the math benchmarks (nearly 97% on AIME!) and finally made coding feel seamless. I remember seeing a demo where it built a "Jumping Ball Runner" game in a single prompt—graphics, sound, and all. It felt like the moment AI stopped being a tool we fiddled with and started being a partner that just got it.
November 18, 2025
Just when we thought the year was winding down, Google came in swinging in November with Gemini 3.
If GPT-5 was about unity, Gemini 3 was about depth. They introduced "Deep Think," a mode that pushes reasoning to a level we hadn't seen before, scoring a massive 1501 Elo on the LMArena. But the part that really excited me as a builder was Google Antigravity.
They reimagined the whole developer experience, turning AI from a code-completer into a full-blown agent that can plan, execute, and validate software tasks. It felt like Google was telling us, "Go ahead, dream big. We'll handle the gravity." It was the perfect capstone to the year’s model wars.
September 29, 2025
Sandwiched right between those two giants was Anthropic’s release of Claude 4.5 Sonnet in late September.
I’ve always had a soft spot for Claude because of how "human" it feels, but this update was pure utility. They staked their claim as the best coding model in the world, and the benchmarks (like SWE-bench Verified) backed it up.
But the killer feature? Computer Use. Seeing Claude navigate a desktop, use a browser, and act like a human operator was wild. It hit 61.4% on the OSWorld benchmark, leaving everyone else in the dust. Plus, they rolled out "Imagine with Claude," which generated software on the fly. For developers, this was the tool that actually helped us get work done in 2025.
March 18, 2025
None of the above would have been possible without the hardware to run it. Back in the spring, NVIDIA unveiled the Blackwell Ultra platform, and Jensen Huang called it the pavement for the "Age of AI Reasoning."
It’s easy to overlook the chips when we’re dazzled by the chatbots, but this was a beast. The GB300 NVL72 rack-scale solution was designed specifically to handle the massive compute needed for "Agentic AI" and "Physical AI." Basically, NVIDIA gave the world the factory equipment needed to build the future. It was the moment we knew the infrastructure was ready for the software explosion that followed later in the year.
February 2025
Finally, we have to talk about the guardrails. Early in the year, the EU AI Act really started to bite.
In February, the first bans kicked in—things like social scoring and manipulative AI were officially outlawed. It wasn't the flashiest news of the year, but it was arguably the most important for our long-term safety. It established the world's first comprehensive legal framework for AI, categorizing tools by risk. It was a sobering reminder that while we’re building these god-like intelligences, we still need human laws to keep them grounded.
🧠RESEARCH
NVIDIA’s new model is designed to be small and efficient for tasks requiring reasoning. It combines two different designs—Mamba and Transformer—to process data quickly. By using a "Mixture-of-Experts" approach, it only uses the necessary parts of its "brain" for each task, making it fast enough for everyday devices.
Making high-definition AI video requires massive computing power. HiStream creates these videos faster by removing repetitive work and building the video in chunks. It creates a rough version first and then refines the details, allowing for high-quality, scalable video generation that doesn't bog down the computer.
This report outlines "Step-DeepResearch," a system built to handle complex research tasks. It is designed to act like a human researcher, capable of gathering, synthesizing, and reasoning over large amounts of information. The paper details the team's approach to building an AI that can perform deep, autonomous investigations.
🛠️TOP TOOLS
Each listing includes a hands-on tutorial so you can get started right away, whether you’re a beginner or a pro.
BigJPG – AI Image Upscaler -AI image upscaler that uses deep convolutional neural networks to enlarge images—especially anime/illustrations—while reducing noise and preserving edges
Bing Create – Image Creator - web-based AI tool that turns text prompts into images
Bingo AI – AI Game / AI Art Quiz - free, browser-based guessing game where you match AI‑generated images to the model that created them—typically DALL·E, Midjourney, or Stable Diffusion.
📲SOCIAL MEDIA
🗞️MORE NEWS
OpenAI Bets on Audio OpenAI is shifting its focus from screens to voice, betting that speaking will be the main way we use computers in the future. They are reorganizing their teams to build a new device that you talk to, which is expected to launch in 2026. This move is part of a bigger trend where tech companies want to reduce the time we spend staring at phones and monitors.
Nvidia Eyes AI21 Labs Nvidia is in talks to buy a startup called AI21 Labs for up to $3 billion, mainly to recruit their team of highly skilled researchers. This type of deal suggests the company is more interested in hiring the talented people who work there than just owning their products. It highlights how much big tech giants are willing to pay to secure the best brains in the industry.
xAI’s Massive Data Center Elon Musk’s AI company is buying a huge warehouse in Mississippi to build its third major computer center. They plan to use an enormous amount of electricity—roughly two gigawatts—to power the supercomputers needed to run their artificial intelligence. The facility has been given a name that pokes fun at Microsoft, signaling the intense rivalry between the two companies.
Alibaba’s Realistic Image Model Alibaba has released a new tool that creates images from text, and it is designed to make the pictures look much more lifelike. The update fixes common issues like "plastic" looking skin and does a better job of displaying text inside the images. It is free for anyone to use and modify, which helps it compete with other top image-creation tools.
Moonshot AI Funding A Chinese startup called Moonshot AI has raised $500 million from investors to help build its next generation of smart computer programs. This large sum of money will pay for the powerful chips and equipment needed to train their new "Kimi" system. The company is growing quickly and is now valued at over $4 billion.
Microsoft’s "Founder Mode" Microsoft CEO Satya Nadella is changing his management style to be much more hands-on, acting more like a startup founder than a corporate executive. He is shaking up the company’s leadership structure to ensure they can move fast enough to keep up with rivals like Google and Amazon. This shows that even the largest companies are feeling the pressure to speed up their development in the AI race.
What'd you think of today's edition? |
Reply