Best of your X follows: robot data, GPTs, and builder instincts

Robotics and Embodied AI

Google DeepMind: robot data is becoming the robotics moat

• Signal: Google DeepMind said Apptronik’s expanded Robot Park will feed Apollo 2 humanoid data into Gemini Robotics training 1.
• Why it matters: Apptronik’s release frames Robot Park, Apollo 2, and the DeepMind partnership as one continuous data-collection loop, with fleets active in Austin and at customer or partner sites 2.
• Implication: The bottleneck is not only better robot models; it is repeatable real-world task data from logistics, manufacturing, retail, and similar environments.

Organizational AI and Product Surfaces

Ethan Mollick: GPTs were an unfinished bridge to agent skill libraries

• Signal: Ethan Mollick noted that many companies still have active efforts to build GPTs, even though OpenAI largely moved on from the format 3.
• Why it matters: His sharper point is that GPTs could have become organizational skill libraries for agents, not just one-off chat customizations.
• Implication: Enterprise AI adoption may depend less on a new chatbot surface and more on durable, reusable bundles of process knowledge.

Paul Graham: builders beat entrepreneurship theater

• Signal: Paul Graham argued that startup founders should learn how to build things, because the hard part is product judgment and execution, not “entrepreneurship” as a subject 4.
• Why it matters: In an AI tooling cycle, this is a useful filter: automation raises the ceiling for builders, but it does not decide what should exist.
• Implication: The durable skill is still picking a real product problem and shipping against it; AI just changes the leverage on that work.

AI Competition and Governance

Ethan Mollick: “China vs. US AI” is eight different arguments

• Signal: Mollick broke the China-versus-US AI frame into at least eight games: company profits, scientific prestige, open versus closed strategy, national stacks, security capabilities, access control, model personalities, and ASI timing 5.
• Why it matters: The post is useful because it stops a vague geopolitical phrase from doing too much work.
• Implication: Any serious policy or market read has to specify which game it is analyzing before claiming who is “ahead.”

Research Taste and Model Culture

François Chollet: modeling is the programming layer under reality

• Signal: François Chollet wrote that “all of reality is programmable” if you can figure out how to model it 6.
• Why it matters: Read literally, it is not a claim that everything is easy to automate; it is a reminder that useful software starts with a working model of a messy system.
• Implication: The line pairs well with today’s robotics item: better models need contact with the world they are supposed to represent.

François Chollet: model weights as cultural artifacts

• Signal: Chollet imagined future “Latent Space Archaeologists” studying 21st-century model weights to reconstruct an extinct culture 7.
• Why it matters: It is a compact way to say that models do not only store capabilities; they also encode the distribution, priorities, and blind spots of the culture that trained them.
• Implication: Long-term AI governance will have to treat training data and weights as cultural records, not just technical assets.

What To Skim First