The Bitter Lesson by Rich Sutton

In his influential essay, "The Bitter Lesson", Rich Sutton, a prominent figure in reinforcement learning, argues that the most significant insight from 70 years of AI research is the ultimate triumph of general-purpose methods that leverage computation over those that rely on incorporating human knowledge. Sutton posits that while building in domain-specific human knowledge can provide short-term gains, these approaches tend to plateau and even impede long-term progress. In contrast, methods that scale with increasing computational power, such as search and learning, have consistently led to breakthroughs.

Sutton supports his argument with several key examples from the history of AI:

• Computer Chess: Early attempts to create chess-playing programs focused on encoding human strategies and knowledge. However, the system that ultimately defeated world champion Garry Kasparov in 1997, Deep Blue, was based on massive, deep search capabilities.

• Computer Go: Similarly, in the game of Go, initial efforts to leverage human understanding of the game were surpassed by systems like AlphaGo, which relied on search and learning from self-play.

• Speech Recognition: The field of speech recognition saw a shift from knowledge-based systems to statistical methods like Hidden Markov Models (HMMs), which performed significantly better in a 1970s DARPA competition. The more recent success of deep learning in this area further underscores the power of computation and learning from large datasets.

Sutton's "bitter lesson" is a four-part observation: 1) researchers build knowledge into their agents, 2) this provides a short-term boost, 3) it ultimately plateaus and hinders further progress, and 4) breakthroughs consistently come from scaling computation with search and learning. He concludes by advocating for the development of meta-methods that can discover and capture the complexity of the world on their own, rather than being explicitly programmed with human discoveries.

The Bitter Lesson's Bitter Lesson by Andrew Trask

"The Bitter Lesson's Bitter Lesson" presents a critique and extension of Sutton's argument. Trask contends that Sutton's focus on "pure learning" from scratch, akin to how babies and animals learn, is computationally impractical and overlooks the immense value of "inherited learning" from human-generated data.

Trask introduces several key quantified points to support his argument:

• The Scale of Evolution: Trask estimates that the evolutionary process that produced human intelligence involved over 10^50 operations. In contrast, current state-of-the-art AI models are trained with around 10^26 operations. This vast difference suggests that recreating the learning process from scratch is computationally infeasible.

• The Efficiency of Inherited Learning: Trask argues that human-generated text is a highly compressed and efficient source of knowledge, representing the output of 4.5 billion years of evolutionary optimization. By learning from this data, AI models can inherit a massive amount of information without having to rediscover it.

• Untapped Human Data: While some may believe that large language models (LLMs) have consumed the entire internet, Trask points out that the training datasets of leading AI models are in the range of 100-200 terabytes. However, the total amount of digitized human data is estimated to be around 180 zettabytes. This means that current AI models are using less than a millionth of the available human-generated data.

Trask's central thesis is that the future of AI lies in developing architectures that can effectively and privately access this vast, untapped repository of human knowledge. He argues for a hybrid approach that combines the benefits of inherited knowledge with the ability for novel discovery, moving beyond the limitations of "pure learning."