Tesla's FSD is a masterclass in human-in-the-loop AI agent design
Alexander Embirikos · Inside OpenAI: 2026 is the year of agents, AI’s biggest bottleneck, and why compute isn’t the issue · Dec 14, 2025 · at 1:19:37
Alexander Embirikos is the product lead for OpenAI Codex. He discusses product design lessons for mixed-initiative AI agents.
0:00 / 0:41exact quote · 41.5s
▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark
“I find the Tesla software, like, quite inspiring. In particular, it has, like, the self-driving feature, and, you know, I, I've mentioned a few times, like, today, like, I think it's really interesting to think about how to build, like, mixed initiative software that makes you feel maximally empowered as a human, maximally in control, but yet you're getting a lot of help, and I think they did a really good job with enabling sort of the car to drive itself, but all these different ways that you can adjust what it's doing without turning off the self-driving. So like you can accelerate, you know, it'll like listen to that. You can turn a knob to change its speed. You can steer slightly. I think it's a, it's actually a master class in like building an agent that still leaves the human in control.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →
More from Alexander Embirikos
The best way for AI models to use computers is writing code
“It turns out the best way for models to use computers is simply to write code.”
Ubiquitous code generation will increase demand for human software engineering skills
“And as you make code even more ubiquitous, it's actually just going to be used for many more purposes. And so there's just going to be a ton more need for people with this, like humans with this competency.”
A dedicated AI browser provides superior context compared to desktop screenshots
“A lot of work is done in the web, and if we could build a browser, then we could be contextual for you, but in a much more first class way. We weren't hacking like other desktop software, which have like very varied support for like what content they're render…”
Human prompt writing and validation are the main limits to AI productivity
“I think that the current limiting factor, I mean, there's many, but I think a current underappreciated limiting factor is like literally human typing speed or human multitasking speed on like writing prompts. And like, you know, you were talking about, it's li…”
AI-driven developer productivity will hockey-stick in 2026
“I think starting next year, we're gonna see, like, early adopters, like, starting to, like, hockey stick their productivity. And then over the years that follow, we're gonna see larger and larger companies, like, hockey stick that productivity. And then somewh…”