The Bitter Lesson: An Overview

Richard Sutton’s concept of the Bitter Lesson posits that, in the long run, it is more effective to let computational systems discover advanced AI methodologies through learning or search algorithms rather than relying solely on human knowledge. He argues that the contributions of humans to the most sophisticated AI systems we've developed have been minimal. In this regard, we have become mere background figures in a narrative dominated by artificial intelligence, which often thrives despite our self-importance rather than because of it. Sutton expressed these ideas nearly two years ago (March 2019), bolstered by compelling evidence and reasoning that, if accurate, presents a rather grim reality.

While he does not suggest that humans are entirely inconsequential—acknowledging the remarkable creativity and expertise behind systems like Deep Blue and AlphaGo (as well as ChatGPT and GPT-4)—he emphasizes that once sufficient computational resources are available, they become far more valuable for enhancing AI performance on specific tasks than human insight. Even though we might not have agreed with the notion that all problems are subject to this principle at that time (not all solutions can be derived from learning), extending his reasoning indefinitely leads us to the conclusion that few problems escape this paradigm.

This article will explore Sutton’s assertion as its foundational premise: while humans are not redundant, computation reigns supreme in the realm of AI. What are the implications of this?

This piece is excerpted from The Algorithmic Bridge, an educational newsletter aimed at connecting AI, algorithms, and individuals. It seeks to illuminate the impact of AI on our lives and equip you with the tools necessary to navigate the future effectively.

Subscribe to The Algorithmic Bridge

Bridging the gap between algorithms and people. A newsletter about the AI that matters to you. Click to read The…

thealgorithmicbridge.substack.com

The Bitter Reality of GPT-4

The recent release of GPT-4 exemplifies Sutton’s arguments. While the engineering expertise involved in its development is substantial, it is noteworthy that OpenAI achieved this feat with only 375 employees, whereas Google, with a workforce exceeding 150,000, struggles to match this success. Are OpenAI's staff members really five hundred times more intelligent than their counterparts at Google?

Debates could arise regarding OpenAI’s exceptional ability to attract talent or Google’s challenges in product delivery, but a more straightforward explanation emerges from Sutton’s lesson: with adequate computational power, the significance of human knowledge diminishes. The sheer number of minds tackling the challenge of enhancing AI models becomes a minor factor in the overall outcome.

GPT-4, created by such a limited group of individuals, is undeniably superior to models like Chinchilla or PaLM, which have also been contenders for the title of the best language models. Yet, the reasons for GPT-4’s superiority remain largely unknown—details about its parameter count, training data, or the computational resources utilized are all shrouded in secrecy. What we do know is that it operates at a scale that outperforms the vast team of highly skilled engineers at Google AI. OpenAI has emerged victorious, leaving many NLP researchers at Google and other tech firms peering into an uncertain future.

While I don’t anticipate mass layoffs for these professionals, the more profound issue lies in the devaluation of their expertise. Their ability to tackle profound questions about intelligence and language has rapidly diminished, eclipsed by the capabilities of just 375 employees and immense computational power, culminating in the enigmatic black box known as GPT-4. This reality is undeniably bitter.

A Future Beyond Understanding

Let’s delve into the broader implications of this situation. It’s essential to view the remainder of this article not just as an explanation of observable phenomena (as Sutton based his points on existing insights) but as a speculative exploration of potential futures that may not be far off.

Plato urged us to emerge from the cave over two millennia ago, yet we have spent centuries excavating a new one. Initially, we believed we possessed substantial knowledge, only to eventually confront our ignorance (to varying degrees). Nevertheless, we have clung to our perceived superiority over all other life forms, viewing ourselves as the stewards of this world.

Despite the challenging truths revealed by the Bitter Lesson—namely, that our contribution to the next wave of groundbreaking technology, which we hoped would stem from our accumulated understanding of the universe, is not as pivotal as once thought—we have managed to retain our dignity. While we may not be the protagonists in this unfolding story, we remain the only ones capable of infusing it with meaning. Although computers may now be the superior forces, we still hold the reins.

However, this dynamic may not last indefinitely. I want to clarify that I’m not suggesting AI will surpass human intelligence (AGI, ASI, etc.). The feasibility of that is uncertain. What is crucial is that, sooner than we anticipate, the systems we create will evolve to a complexity that even our most advanced minds may struggle to comprehend. This phenomenon is already in motion.

The reality is that no one—even the creators—fully understands what GPT-4 entails. The cultural memes and philosophical musings surrounding it are desperate and futile attempts to impose coherence on a system that is eluding our grasp. As time passes, it is likely that only mysteries will remain. Yet, we will continue advancing because our metaphorical steeds (the computers) will forge ahead toward the unknown, even when we can no longer recognize our destination.

We once stood as masters and rulers. Now, we remain the ideators but not the primary builders. Soon, we may become mere spectators in a world neither of our making nor comprehension—a reality that unfolds before us too rapidly to grasp and too intricately to decipher. The irrelevance we dread—not just as individuals but as the "chosen species"—is looming in the near future that we are eagerly approaching.

Accepting that we might not be the linchpin in this intricate puzzle has been painful. Realizing that we may not even be significant enough to contribute meaningfully to the extraordinary wonders awaiting us on this journey will be even more so.

Subscribe to The Algorithmic Bridge. Bridging the gap between algorithms and people. A newsletter about the AI that matters to your life.

You can also support my work on Medium directly and gain unlimited access by becoming a member using my referral link here! :)

Chapter 2: The Bitter Lesson in AI

The first video, "The Bitter Lesson (in AI)...," offers insights into the implications of computation over human knowledge in AI development.

The second video, "The Bitterest of Lessons: The Role of Data and Optimization in Emergence," further explores the significance of data in the evolution of AI systems.