The next big breakthrough will be AIs learning on the job
The article argues that the next major AI breakthrough will come from models learning and improving through real-world work experience, rather than solely from scaling up pre-training data. This paradigm would allow AI to gain practical skills and adapt dynamically, much like how humans develop expertise on the job.
Background
- Dwarkesh Patel writes a newsletter (Dwarkesh Podcast / The Dwarkesh Letter) focused on long-form interviews with thinkers in AI, economics, and technology. His audience tends to follow tech-intellectual debates closely.
- The article argues that the next leap in AI capability won't come from bigger models or more training data, but from AI systems that learn continuously while deployed in real-world settings ("on the job").
- This contrasts with the current dominant approach: training a model once on a huge static dataset, then freezing it. The piece suggests this paradigm will give way to "continual learning" where AIs adapt from experience.
- Key context: Many AI researchers see diminishing returns from simply scaling up model size (the "scaling laws" debate). The article points to a shift in focus toward agentic, self-improving systems — a topic of intense speculation in Silicon Valley and AI labs like OpenAI, Anthropic, and Google DeepMind.
- The idea connects to long-standing concepts in AI: reinforcement learning from human feedback (RLHF), self-play (like AlphaGo's training), and the notion of AIs as "agents" rather than static tools.