Lest we become overly enamored with AI, we remind ourselves:
AI is not a substitute for human understanding. We do not want a chatbot where you can ask a question and get an opaque answer. The answer isn’t the point. The meaning of the answer — its implications, its nuances, its accuracy, its effect on your future assumptions and decisions — is the point. This means an agent should show its work and you should understand its derivation. Large language models are great at generating code; Observable’s job is to make that code more interpretable and malleable. That’s why we care about the computational medium, where you can see the state of a program as its runs, where you can instantly modify running programs, and quickly add interaction or visualization. The medium is where humans and computers work together.
AI is not a substitute for human collaboration. We don’t want agents to replace human relationships. Observable is designed for collaboration — to quickly see what others are working on, to contribute to others’ work or riff on it, to reuse code, to ask for or receive feedback — and that extends to chats, too.
AI is not a substitute for human creativity. Observable grew out of the nascent community of D3 users, and we continue to be inspired by the creativity and generosity of this community. As magical as it is to have an agent quickly sketch for you, an agent is not a community. A community does not merely suggest code; it provides support, empathy, and compassion. A community builds relationships that go beyond whatever tool you use. Read How D3 moved us for a few of those stories (seriously).
For some reason, this reminds me of the Navier–Stokes problem and the surrounding controversy. The primary goal of mathematics (and many other scientific fields) has always been human understanding of ideas. Solving a problem (say in math) allows one to explain to others the new ideas that led to a new proof / solution. On the contrary, the AI companies seem to be using AI to just “get the answer” rather than aligning with the aforementioned goals of mathematics. It’s a game to see whose AI is the “smartest.”
AI is not a substitute for human understanding and creativity… at least not yet.
Fascinating, and helps me with some thinking I’ve been having lately.
I did a little more research on this topic, which resulted in the following thoughts (from Perplexity). Teachable and extensible - great systems are built from building blocks, and it seems like abstractions, simple data structures, and architecture still have their place if we want to build extensible systems.
“Getting the answer” versus insight
Here is a useful analogy.
Suppose an AI gives you a circuit layout that passes simulation, design-rule checks, and manufacturing tests—but it is so complex that nobody can tell why its control loop is stable, which components are safety-critical, or how it will behave when you change the load. You possess a successful artifact, but not necessarily a transferable design principle.
Likewise, a formally checked proof might establish:
The theorem is derivable from these definitions and axioms.
Human mathematical understanding aims for additional things:
Here is the key invariant.
Here is the mechanism that prevents—or causes—singularity formation.
Here is why this estimate is the right one.
Here is how this changes our view of related PDEs.
The latter is what makes a result teachable, extensible, and useful beyond one theorem.
“The most important problem is unsolved,” says Luis Silvestre, a mathematician at the University of Chicago. “The Clay problem is settled, but the main problem for the Navier-Stokes equations is not.”
Furthermore, last Thursday, three mathematicians posted a proof of their own that showed that OpenAI’s method can never be extended to solve the full problem. In other words, the loophole will never be closed, barring some completely new idea.
In other words, the LLM’s result does not—and will never—answer the Navier-Stokes problem that mathematicians really care about.
It did, however, unambiguously solve the problem according to the Clay Institute’s original formulation. The official problem statement, penned in 2000 by mathematician Charles Fefferman, offers an option called “C,” in which solutions are allowed to use an external force like OpenAI’s.
This might even be good news for “team humanity.” LLMs are great at finding blowups that exist—at searching the infinite landscape of fluid scenarios and plucking out the precise situation that breaks the equations. But proving that blow-up is impossible is a kind of math AI still struggles with. “We may be at less of a disadvantage, or maybe an advantage, compared to LLMs,” says Cao-Labora. “LLMs are especially good at constructing things that are very explicit and not as good—for now—in making new theory.”
I have a feeling that AI is simply taking another path to get smarter. While us mere mortals care about emotions and elegance, the AI is just robotically marching toward minimizing the loss function. In the limit, these two paths may converge on indistinguishable output, but I think we are years away from true “AI creativity.”