Benny Jiang

• I am building Airbot, the deployment OS for customer-facing robots.
• I am an AI engineer at Meta, working on ads creative agents.
• I like building things at the intersection of AI and product.
• I earned my master's degree in Computational Data Science from Carnegie Mellon's Language Technology Institute, where I was advised by Professor Eric Nyberg.
• Before that, I studied Computer Science and Data Science at UC Berkeley.
• I live in San Mateo, the best city of the world.

• Airbot, the deployment OS for customer-facing robots.
• Qdrop, a quick file sharing tool.
• write-like-me, a small tool for learning and reproducing a writing style.
• Roamee, an AI tour guide for exploring places with local context.
• Skillmaxxing, self-evolving skills for coding agents.
• TwentyEyes, a Mac eye break reminder that only counts real screen time.

• AI agents that can actually use tools and memory well.
• Products that feel simple after the hard engineering is hidden.
• Small teams, fast iteration, and shipping things before they get too precious.
• Travel, maps, interfaces, and the strange ways people discover places.

• Meta and frontier models. When Mark hired Alexandr Wang and paid AI researchers crazy money, I thought Meta could not build a frontier model. As of October 2026 this is basically proved wrong: Meta now has a tier 1 model, not the best yet.
  • Why I was wrong: Meta has no track record in hardcore technology. It is great at consumer products, but not at hard tech compared to Google or Tesla, and its bets on phones and the metaverse have not paid off (yet). I work there, and the culture rewards metric movement and fast iteration over long bets on fundamental technology.
  • What I learned: do not just pattern match. Past failure does not predict future failure, and vice versa. Reason from first principles instead: the first principle of LLMs is the scaling law (large data, large compute, large model), and Meta has all three.
• How fast AI would automate coding. Using Cursor in 2024 and early 2025, I thought it would take a long time for AI to take over coding. After Claude Code, AI replaced all of my primitive coding work.
  • What I learned: be more imaginative. AI can probably do far more than coding. The first principle of AI coding is that the task has verifiable output, so AI can finish any job with verifiable output. Making things more verifiable means making them more automatable.
  • The interface matters a lot. Claude Code versus Cursor is more freedom for the model to take actions and a simpler interface. I am not sure yet whether this is a trend worth digging into to get more power out of AI models.

• Aside
• Orca, the ADE for working with coding agents.
• Instinct, a personal assistant that takes action across your apps and devices.

• "What is the most resilient parasite? Bacteria? A virus? An intestinal worm? An idea. Resilient, highly contagious. Once an idea has taken hold of the brain, it is almost impossible to eradicate." - Cobb, Inception

• GitHub
• X
• LinkedIn
• Email