On Fri, Jan 16, 2026 at 1:35 AM Erik Moeller <[email protected]> wrote:
> A huge reason readers may prefer AI summaries, even if inaccurate, is > to get to an answer more quickly (the same reason Wikipedia itself > outperformed other information sources even when its quality was still > very uneven.) Yes. It's latency. It was always latency. "Wiki" means quick. Wikipedia made the sum of human knowledge (or its arithmetic mean, anyway) accessible near you: in your home, in your backpack, in your pocket, on your person. It shortened the distance from question to answer by abstracting the trip to the library or bookshelf, the time spent poring over the table of contents, and the wait for the newspaper to arrive, for a new textbook edition, or for the translation to appear in your language. For performance engineers like me, Wikipedia's "end-to-end latency" is the time between a reader clicking a link and the article fully rendering on their device. For many years, I believed the key to Wikipedia's continued relevance was shaving milliseconds off this number by tuning Wikimedia's code and infrastructure. But *true* end-to-end latency is not measured between server and browser, but between curiosity and insight. And it turns out that network and code latency contribute only modestly to that number. The milliseconds it takes for Wikimedia's servers to transmit an article to your device are dwarfed by the time you need to wrack your brain for the right terms to query, locate the relevant section of the article, interpret its meaning, and relate it to your question. Wikipedia improved on Britannica's end-to-end latency by several orders of magnitude. Modern AI is now doing the same to Wikipedia. I can describe to Gemini what I want to know using vague, imprecise, or even incorrect terms, and it tells me what I might be thinking of, using my language: not merely the language listed in my Babel userbox, but terms I understand that relate to concepts I already know and are appropriate to my level of understanding. At its worst, AI generates hallucinated, sycophantic slop. But at its best, it is an interface to human knowledge that is not merely incrementally faster than browsing Wikipedia, but categorically faster. I think the key to ensuring the future knowledge infrastructure remains free and open is to once again beat closed, commercial platforms on latency, ideally by an order of magnitude or more. This is possible, if you again consider true end-to-end latency and the invisible factors that contribute to it, like the time it takes to distinguish truth from falsehood, and information from manipulation. I'm not sure Wikimedia should lead the charge. Even if their relevance is fading somewhat, the projects are an immense trove of value for humanity. Any rash effort to remake them from within is likely to destroy more value than it creates. But there is plenty of room out there for new things. I'm glad to see you experimenting in this space, Eric.
_______________________________________________ Wikimedia-l mailing list -- [email protected], guidelines at: https://meta.wikimedia.org/wiki/Mailing_lists/Guidelines and https://meta.wikimedia.org/wiki/Wikimedia-l Public archives at https://lists.wikimedia.org/hyperkitty/list/[email protected]/message/DYH4X2KQWJOGQQI5C7O7W6RGKNPIVBTS/ To unsubscribe send an email to [email protected]
