One of my favorite quotes comes from Jurassic Park, when Dr. Ian Malcolm speaks to John Hammond, who created the first theme park with live dinosaurs:
“I’ll tell you the problem with the scientific power that you’re using here. It didn’t require any discipline to attain it. You know, you read what others had done and you took the next step. You didn’t earn the knowledge for yourselves, so you don’t take any responsibility for it.”
I’m sure I’m not the only person who, as they watch this classic for the umpteenth time, compares it to what each person has at their fingertips now with AI. Of course, with Claude writing production-level code and millions of people using GPT for their professional tasks, it took a long time for AI development to get to this mature state. To that point, the knowledge was most definitely earned, with researchers pouring thousands of hours and millions of dollars of compute to achieve even the slightest of efficiencies in training LLMs.
But for the average person who is merely a consumer of AI, the movie quote serves as a warning. Document work that used to take hours now becomes fully generated in seconds, and the trend of “vibe-coding” has proliferated across engineering circles, where people exclusively use AI to plan, code, and debug any errors that arise. The upgrade in efficiency is undeniable. Teams are shipping faster and accomplishing roadmaps in a fraction of their planned times, and we are constrained only by our own willingness to throw our problems into this new incredible black box.
When I try to envision the future of AI, I always try to compare it to the trajectory of chess engines. At first, engines were poor, and the world’s best grandmasters were able to beat them with ease. These engines used heuristics to evaluate positions, imitating human intuition and pattern recognition. However, they failed to play good chess, because it was impossible to constrain chess to a structured set of rules. Putting your rook on an open file might be a good move in 80% of cases, but there’s no guarantee it’s the best move in the rest. Even if one tries to compartmentalize different scenarios, there’s no way to know out of all the possible positions available, moving a rook to the open file is the best move in that position, because there’s just too many possible positions to consider.
As compute power improved, these engines evolved into something more powerful: calculation machines. There was no longer a need to evaluate moving a rook to an open file as +0.05; if the position was better after 20 moves as a result of playing that move then it was already taken into consideration. Now, engines can beat humans with moves we would consider impossible to find, purely because they look through every single branch possible up to a certain depth.
As we view the future of LLMs in the context of chess engines, we see parallels. At first, LLMs and early NLP methods tried to create structured rules to better understand complex abstractions. As compute improved and the price of tokens decreased, LLMs were able to train on more data to generate more accurate output.
But in the age of powerful chess engines, do we stop watching humans play chess? In fact, the opposite has happened — the popularity of chess has exploded since Covid. If chess is “solved”, why do we still value human mastery?
As someone who played competitive chess for years, I can offer a few perspectives. For one, chess is so complex that humans have spent decades trying to master it. When a chess engine appears that can calculate 50 moves in advance, the interest and curiosity regarding the game persists. Rather than competition, the chess community sees engines as tools to improve. In situations where the optimal move is difficult to find for a human, they can use the engine as a source of truth.
Secondly, chess engines have illuminated the beauty of the game. Every chess player has experienced moments where the engine finds a genius move in the most unassuming of positions. The immediate question for every chess player then is “how?”. This leads to a moment of understanding: trying to figure out how this move is the best leads to learning. Would we have gained that understanding without engines? Maybe given enough time and study, but there’s no guarantee we would get to the right answer unless we exhaust every variation which is time consuming for humans. So engines, by “solving” chess, have given humans access to knowledge and a newfound appreciation of the game.
Thirdly and most importantly, chess engines are banned from use during competitions. Every player can use the engine during practice, but ultimately they are responsible for winning over the board. This is an important distinction from LLM use, where they are now integrated with workflows both professional and casual. While there is the obvious reason that chess engines are banned because it would ruin the spirit of the competition, another is that they are more reliable than LLMs in providing an answer. An LLM could be integrated into a workflow alongside multiple safety checks and guardrails but there’s always a chance the output is not correct. There could be hallucinations or the LLM could have been trained on the wrong information, thereby producing an incorrect answer. This is different to chess engines, which instead rely on deterministic branches and act with complete information about the game state.
Given this context, it’s difficult to imagine how LLMs would follow in the footsteps of chess engines. People use LLMs for tasks requiring little interest, and are encouraged to use them to build production code. LLMs find innovative ways to solve tasks humans wouldn’t consider, but it is unclear how much learning is being done to understand why. In many cases, people use agents to review the work of other agents.
Even in the age of AI, I believe that human mastery is beautiful. There is meaning in struggle, where it ends in learning. And there are others who are eager to learn new skills and improve themselves, and they will thrive in the age of generative AI. Many have already created learning courses using LLMs, either having them check work or selecting areas of improvement and generating subsequent quizzes and lessons to target weaknesses. For the first time in history, there is no longer a dependence on others to create tools that can help you specifically. In the past, a student only had a curriculum built by teachers based on what they thought was important — today, each person can construct a full course for themselves in almost anything.
The golden age of learning is here, and we can take a page from chess engines on how we can better use them to earn the knowledge for ourselves.