5 significant areas where I've been wrong + right in the past 20 years
Wed, 12 Nov 2025 15:23:23 GMT
1. CONVERSATIONAL SEARCH ENGINES: In middle school into high school, I started developing a "conversational search engine" called GNUAsk. I applied both to undergrad and grad school with the idea of prototyping a conversational search engine. I knew neural nets would play a role, I didn't know exactly how. I was using hidden markov models to unreliably generate natural language, translating natural language queries into a set of machine queries, collecting gigabytes of natural language with bots via chat room logs ( AOL Instant Messenger (AIM), IRC, etc), and tuning bots to attempt to answer questions in these chats with varied levels of success.
2. SEARCH EVERYWHERE: I thought conversational search engine should be integrated (as a background secretary) into the fabric of our Operating Systems (e.g. in the terminal, in our chats). This wasn't a big leap -- Alan Kay and Smalltalk showed us what an operating system (Plan9) could be when every program could talk to each other. I was taught an interesting take in undergrad that "AI is Search" -- i.e. AI is solving search problems -- and this stuck with me.
3. THE UNIVERSAL MESSENGER: I hoped (against what seemed like inevitable market-cornerning) that communication applications would converge on standards or interoperable APIs that would allow their participants to funnel every correspondence into a single system. I thought the solution may be primarily "multi-auth" (kind of like Mint.com) but I think the more important advancement is LLMs being able to bundle and triage messages, summarize, cut through the noise.
4. THE INTERNET, A SEMANTIC GRAPH: I believed we'd move in the direction of graph-based note-taking with (bidirectional) back-links (and prototyped graph.global ~2016, a precursor to Roam Research). I think Wikidata has helped a lot, but I think LLMs have shown that "figuring it out on the fly" is perhaps more tractable as an approach to keep up with the world (and saving tags / associations along the way as a bi-product) than tagging everything.
5. Similar to point [4], I thought that we'd have much better universal learning systems, which -- instead of being domain specific -- are sensitive to what each individual knows. I figured this wasn't much different than the targeted advertising that is already done to/on/for us. Danny Hillis, in a previous life, had many talks on Personalized Learning: https://www.youtube.com/watch?v=cz7wN904rvs -- we have MOOCs and LLMs are very useful as tutors, however the world seems to be headed in a direction of increased digital artificial scarcity in many ways (including library ownership).
Tags: Roam Research, AOL Instant Messenger (AIM), IRC, "AI is Search", Wikidata, Plan9, Alan Kay, MOOC, LLM, Smalltalk, conversational search engine should be integrated (as a background secretary) into the fabric of our Operating Systems, GNUAsk, graph.global, Danny Hillis