Skip to main content

[@DwarkeshPatel] What are we scaling?

· 3 min read
[@DwarkeshPatel] Summarizer
Video Bot

@DwarkeshPatel - "What are we scaling?"

Link: https://youtu.be/_zgnSbu5GqE

Short Summary​

The speaker discusses the current state of AI development, particularly in regards to scaling up reinforcement learning and the challenges of achieving human-like intelligence. They express skepticism about the possibility of achieving transformative economic impact within the next few years due to the limitations of current models.

Key Quotes​

Key Quotes​

  1. "If we're actually close to a humanlike learner, then this whole approach of training on verifiable outcomes is doomed."
  2. "When we see frontier models improving at various benchmarks, we should think not just about the increased scale and the clever ML research ideas, but the billions of dollars that are paid to PhDs, MDs, and other experts to write questions and provide example answers and reasoning targeting these precise capabilities."
  3. "Somehow, this automated researcher is going to figure out the algorithm for AGI, which is a problem that humans have been banging their head against for the better half of a century, while not having the basic learning capabilities that children have. I find it super implausible."
  4. "Human workers are valuable precisely because we don't need to build in the schley training bloops for every single small part of their job."

Detailed Summary​

The speaker critiques the approach of scaling up reinforcement learning (RL) to achieve human-like intelligence, questioning the assumption that training on verifiable outcomes will lead to a human-like learner. They argue that if we are close to achieving a human-like learner, then the current approach of pre-baking skills into models through RL environments will be pointless. The speaker also discusses the tension between the billions of dollars spent on training models and the lack of progress towards achieving general intelligence.

The speaker suggests that the current approach to AI development is focused on building models that can perform specific tasks, but these models lack the ability to generalize and learn on the job. They argue that human workers are valuable because they don't need to be trained on every single task, but rather can learn and adapt on the job.

The speaker also touches on the idea that the development of AGI will require significant advancements in areas such as continual learning, and that the current focus on pre-training models on large datasets is not sufficient. They suggest that the future of AI development will involve the creation of continual learning agents that can learn and adapt on the job, and that this will require significant advancements in areas such as in-context learning.

The speaker concludes by expressing their skepticism about the possibility of achieving transformative economic impact within the next few years, and instead predicts that progress will be gradual and will require significant advancements in areas such as continual learning.