Wojciech Zaremba: OpenAI Codex, GPT-3, Robotics, and the Future of AI | Lex Fridman Podcast #215

Zaremba suggests that systems like GPT learn by compressing reality (predicting text), and consciousness/self‑consciousness might emerge when a powerful compressor starts modeling and compressing itself—analogous to Gödelian self‑reference.

He likens the human “story of self” to a GPT prompt: the narrative we prepend determines how we behave, just as a well‑crafted prefix (“You are Elon Musk…”) steers GPT into different personas and capabilities.

By training on code and text, Codex can translate human instructions into working code (e.g., GitHub Copilot), and more broadly into plugin calls for tools like calendars, documents, or creative software, effectively shrinking the gap between intent and execution.

While architectural innovations (e.g., transformers, dropout) matter, internal OpenAI analyses attribute most recent gains to scale; Zaremba sees three multiplicative levers—data, compute, and algorithms—with compute having yielded the largest returns to date.

Their Rubik’s Cube hand project required massive simulation with domain randomization and still struggled with real‑world latency and hardware maintenance; for a commercial robotics firm today, he’d start with tele‑operation to amass supervised data before autonomy.

He believes language‑based systems, augmented by multimodal perception and continual learning, can become extraordinarily good at understanding personalities and emotions, but insists they should be tuned to maximize human wellbeing rather than attention or addiction.

Zaremba models love as partly “optimizing another’s reward function as your own,” and sees meaning as arising from a structured set of internal rewards (curiosity, connection, creation) that can be understood scientifically yet remain experientially profound.