Why AI Won't Make You Rich & Spy Glasses

TLDR;

This week, we take a look at many relevant robotics news, an essay that will send chills down your spine if you’re an AI investor, new models from xAI and OpenAI, and even Meta’s uber-cool, privacy-invading AI/AR RayBan’s.

Enjoy!

ROBOTICS

Robot “Violence Tests”

In one of the most remarkable videos from AI in recent times, a group of Chinese researchers has showcased their “violence tests” on a humanoid, attempting to push it around and observe its reactions.

→ WATCH VIDEO

TheWhiteBox’s takeaway:

There’s a lot to say about this video.

Who is ahead in robotics? With the upcoming IPO from Unitree, the progress we’re seeing in robotics coming from China is simply incredible. US companies like FigureAI are also showing mind-blowing progress, but China has one thing the US hasn’t: the capacity to drop manufacturing prices for these robots tremendously (even though some US startups have very ambitious manufacturing goals in the US). They are selling in the single-digit thousands in some cases, while US robots are still prohibitive in terms of price (~$59k for a chores robot).

What are these Labs actually trying to build with a robot that has great motor capacity? I can only think of one place where they would need so much agility: war.

That said, I always mention the importance of not getting carried away by demos.

Nonetheless, you can watch this Dwarkesh interview with Sergey Levine, a robotics household name, and see how he’s much more cautious about real progress (in fact, he’s notoriously skeptical of humanoids and is building a totally different type of AI robots).

To him, the ‘ChatGPT moment’ for robotics is at least two years away, and a GPT-5 moment (where models start being truly useful) is more like towards the end of the decade.

So let’s celebrate progress, but if there’s one field where we must not get carried away with hype, it’s robotics.

AGENTS

Are we doing MCP wrong?

As you may or may not know, there is considerable hype around the Model Context Protocol, a way for agents to utilize tools (they need tools to execute actions) using natural language.

That is, if the model needs to use a calculator, it can use a calculator MCP server to execute calculations using English, such as sending the server the following command: “use calculator_tool(5,6)”.

This isn’t done ‘just because’ but because English is a universal language all Large Language Models (LLMs) speak, so it’s a way to standardize tool calling (what agents do to execute actions) across all tools.

However, Vercel, a billion-dollar cloud company, believes we are doing this all wrong.

The issue is that we currently perform MCPs at the API level (the actual interface agents use to connect to third-party software, with MCP sitting on top to make them “English accessible”). This means that if we have a task where the agent has to do four steps, it executes four actions, while we could design it so that it executes just one, and the…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.