A bitter lesson.

By N. Whitman. 1999

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Reinforcement Learning

Citations in the corpus

Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer (2019, background): “…research argues that general methods that can leverage additional computation ultimately win out against methods that rely on human expertise [Sutton, 2019; Hestness et al., 2017; Shazeer et al., 2017; Jozefowicz et al., 2016; Mahajan et al., 2018; Shazeer et al., 2018, 2017; Huang et al.,…”
Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity (2021, background): “Simple architectures— backed by a generous computational budget, data set size and parameter count—surpass more complicated algorithms (Sutton, 2019).”
Perceiver: General Perception with Iterative Attention (2021, background): “This is in accord with the idea that greater generality follows from making as much of a system learnable as possible (Sutton, 2019).”
UNICORN on RAINBOW: A Universal Commonsense Reasoning Model on a New Multitask Benchmark (2021, background): “Scaling Laws In contemporary machine learning, simple methods that scale often outperform complex ones (Sutton 2019).”
What Do NLP Researchers Believe? Results of the NLP Community Metasurvey (2022, methodology): “While the success of deep learning and large language models may be taken as supporting evidence for the Bitter Lesson, we find that the community has bought into the Lesson far less than it thinks it has.”
Measuring the Algorithmic Efficiency of Neural Networks (2020, methodology): “Within those domains, our desiderata were: 1. task of sufﬁcient difﬁculty to demonstrate that improvements work at scale [Sutton, 2019] 2. benchmark of high interest over long horizon in which there’s general agreement we’ve observed large progress in capabilities.”
On the Binding Problem in Artificial Neural Networks (2020, background): “Here we have proposed to address this problem by incorporating a small set of inductive biases to enable neural networks to process information more symbolically, while also preserving the crucial benefits of end-to-end learning (Sutton, 2019).”
Towards Continual Reinforcement Learning: A Review and Perspectives (2020, background): “(Sutton, 2019).”
The hardware lottery (2020, background): “It is an ongoing, open debate within the machine learning community about how much future algorithms will diﬀer from models like deep neural networks (Sutton, 2019; Welling, 2019).”
Towards the Systematic Reporting of the Energy and Carbon Footprints of Machine Learning (2020, background): “…models (Wu et al., 2019; Zhou et al., 2020; Reddi et al., 2020; Lu et al., 2018; Coleman et al., 2019; Jiang et al., 2019), yet another line of work has focused on the opposite: building larger models with more parameters to tackle more complex tasks (Amodei and Hernandez, 2018; Sutton, 2019).”
Inductive biases for deep learning of higher-level cognition (2020, background): “Finally, we would also like to mention that inductive biases are not the only way to bridge the gap to high-level human cognition: we may gain by improving our optimization algorithms, by scaling up neural networks (Sutton, 2019) and by moving to other frameworks that better capture uncertainty abou”
Sora: A Review on Background, Technology, Limitations, and Opportunities of Large Vision Models (2024, methodology): “The training approach of Sora aligns with the core tenet of Richard Sutton’s T HE B ITTER L ESSON [30], which states that leveraging computation over human-designed features leads to more effective and flexible AI systems.”
AI-GAs: AI-generating algorithms, an alternate paradigm for producing general artificial intelligence (2019, background): “As Richard Sutton has recently pointed out, history has shown that the algorithms that tend to win over the long haul are simple ones that can take advantage of massive amounts of computing [178].”
Process Reinforcement through Implicit Rewards (2025, unknown): “As the fuel of LLMs, data, will be depleted in the near future, we are entering a new era of experience, which is exemplified by RL (Sutton, 2019).”
Using cognitive psychology to understand GPT-3 (2022, background): “Thus, GPT-3 can be thought of as an experiment in massively scaling up known algorithms 11 .”