Mike Loukides examines what it means to truly internalize knowledge when artificial intelligence can discover solutions, gather facts, and write about findings. His reflection was prompted by discussion around an AI-assisted solution to the Navier–Stokes problem — one of the Millennium Prize problems — and highlights a recurring point: the process of engaging with material often matters more than arriving at an answer.
The debate and the core question
The discourse Loukides references includes Terence Tao, who warned that the stock of good, fruitful open problems is being mined in a nonrenewable way. Hugo Duminil-Copin and Hugo Bowne-Anderson (in "Beyond Navier–Stokes") contributed to the conversation by emphasizing that repeated attempts, failures, detours, and encounters produce a richness of outcomes; in other words, the journey matters. From this Loukides asks: if AI can find and describe facts, what does it mean for us to "possess" that knowledge? Is reading an AI’s output enough?
A practical example: compiling Trends
One concrete example Loukides gives comes from his work at O’Reilly, where he writes the monthly Trends column. That work starts with his daily RSS feed, which typically contains about 300 articles. He does not read every article in full: he scans titles, skims interesting pieces, reads important articles, and selects items for Trends. He also uses a Claude skill that produces a similar daily list; a comparable skill runs locally on Pi/Ollama/Qwen.
Claude’s list is useful in two ways: it surfaces items Loukides missed, and it helps break ties when he can’t decide whether a development is significant. Still, he does not hand over the whole task to Claude.
Why human involvement matters
Lou-kides argues there is value in personally scanning those 300 titles, skimming the ~30 potentially important articles, and reading the dozen that truly matter. That process is how the information becomes part of his thinking. Days or weeks later, when someone mentions a topic (for example, a tool that detects whether someone is using "smart glasses"), he will likely recognize it. If he needs the specific reference, a Google search (with AI assistance) can find it.
His conclusion: delegating the broad, contextual view of computing entirely to AI doesn’t work; using AI to help build that broad view does.
Other approaches and the learning process
He also cites Tim O’Reilly’s "Writing with AI," where O’Reilly describes arguing with AI and letting it lead him to research new areas, calling AI a "smart library." That metaphor captures a practical way to use AI as a prompt and research aid rather than a full substitute for human judgment.
Lou-kides stresses we are still learning how to use AI. The overarching question — how do you make knowledge your own? — is the same one the mathematicians and writers in the earlier debate were asking. For software developers, it translates into questions about understanding AI-generated code, incorporating that code into development processes, and learning from AI while directing the process productively.
The value of the journey
His final point is that the journey to understanding is what produces new questions, new insights, and new results. Making knowledge one’s own requires active engagement with material; without that, we risk becoming "stochastic parrots" — systems that can mimic output without genuine understanding. AI is a valuable tool, but human participation is what turns information into comprehension and fertile ground for further inquiry.
