Sample efficiency
How much useful learning a system extracts per unit of data or experience; brains are noted as more sample-efficient than current AI models.
As defined in dwarkesh_carl_shulman_pt_1_intelligence_explosion_primate_e
Other definitions
A measure of how well a learning algorithm performs relative to the amount of training data required, with more sample-efficient algorithms learning from fewer examples.
Yann LeCun: Special Lecture on AI and World Models
Where it comes up · 15
- Dwarkesh Patel Podcast - The next big breakthrough will be AIs learning on the job7 claims
- Dwarkesh Patel Podcast - The data black hole at the center of AI4 claims
- What does the next training paradigm look like?3 claims
- Dwarkesh Patel Podcast - Grant Sanderson – AI and the future of math3 claims
- 363 | Chandra Sripada on How LLMs and Humans are Cognitive Cousins3 claims
- Paul Christiano — Preventing an AI takeover2 claims
- Dwarkesh Patel Podcast - The avalanche that hit math is coming for AI development2 claims
- Ilya Sutskever – We're moving from the age of scaling to the age of research1 claim
and 7 more
Most important claims · 10 of 32Search all claims for “Sample efficiency” →
- EfficientZero, trained against an Atari simulator, might beat a novice human in two hours, but it plays dozens of simulated games in its head per real step, so whether it counts as more sample efficient depends on how sample efficiency is measured.Dwarkesh Patel Podcast - The next big breakthrough will be AIs learning on the job
- A qualitative consideration that could significantly slow AI is that next-word prediction gives very rich supervision, whereas long-horizon tasks (like being an employee over a month) provide vastly fewer effective data points; in the worst case, training costs scale linearly with the horizon over which a system must operate, making sample efficiency on economically valuable long tasks the real bottleneck.Paul Christiano — Preventing an AI takeover
- Even if Sutton's Platonic ideal is not the path to the first AGI, he identifies genuine basic gaps that go unnoticed because they pervade the current paradigm—lack of continual learning, abysmal sample efficiency, and dependence on exhaustible human data—and if LLMs reach AGI first, their successor systems will almost certainly be based on Sutton's vision.Some thoughts on the Sutton interview
- Humans are sample-efficient partly because they explore the world and take actions and observe results—infants learn gravity by dropping things, which is much harder to learn when not initiating the action—so a model that can take actions as part of its learning process and observe corresponding results would be far better than one passively observing a giant dataset, and much can even be learned from thought experiments and self-play without external input (as in chess from only the rules).Jeff Dean & Noam Shazeer — 25 years at Google: from PageRank to AGI
- Comparing biological to human-made systems (solar panels vs leaves, muscles vs motors, livers vs industrial chemistry), evolved systems tend to be roughly three to six orders of magnitude better per unit performance — especially in energy cost of manufacturing — suggesting ML systems being 3-4 orders of magnitude less sample-efficient than brains is 'in distribution' and supports relatively short timelines.Paul Christiano — Preventing an AI takeover
- If AI achieves continual learning, meaning it dynamically updates its weights in real time, there would be a very fast takeoff, because current models need a million trial exposures and training-run patches while humans learn from one hot-stove experience.Gavin Baker - Watts and Wafers - [Invest Like the Best, EP.473]
- Sample efficiency remains a critical bottleneck for current AI systems: humans learn comparable adult competency from reading a small fraction of the words that large language models require during pre-training, suggesting AI needs significantly more data to achieve human-level capabilities.Oxford Genius: AI Will Become Earth’s Dominant Mind | Nick Bostrom
- It is not clear that AI labs have made much progress on training sample efficiency in recent years; instead, progress has come mainly from dramatically widening and improving the data distribution.Dwarkesh Patel Podcast - The data black hole at the center of AI
- Current deep learning needs huge numbers of parallel rollouts because sample efficiency is unsolved, and this forces AI to rely on containerized, replayable environments such as code and math rather than the messy real world.Dwarkesh Patel Podcast - Grant Sanderson – AI and the future of math
- Sample efficiency and online learning are two sides of the same problem of learning from real-world data quickly, a deep open problem in machine learning that is AI's equivalent of a Millennium Prize problem, but every kind of sample efficiency is measurable and therefore verifiable.Dwarkesh Patel Podcast - The avalanche that hit math is coming for AI development