# Bottleneck of AI-Human Interaction Loop > Notes on what I experienced regarding the human bottleneck in AI-human interaction between 2023 and 2025. Notes on what I experienced regarding the human bottleneck in AI-human interaction between 2023 and 2025. ## Notes ### November 14, 2025 [Speculative tree search](https://wiki.g15e.com/en/pages/Speculative%20tree%20search.txt) ### August 9, 2025 started supporting 2,000 free requests per day. It feels like I should make use of it somehow. Needs thought. For now, I've set things up so that my iPhone and the laptop at home stay connected all the time (Blink App + mosh + tmux). What would be good to do on a narrow screen? Open a branch and tell it to do a big chunk of coding work? (Review on a big screen.) I guess I'll need to make something like a priority queue. ### July 12, 2025 It's hard even to write down exactly what I'm thinking, and even if I write it exactly, the result is different from what I expected. A sentence that pinpoints the problem with the current method of instructing AI agents in natural language: > What you write down doesn't mean exactly what you think it means. And when it does, it doesn't have the consequences you expected. --, [Software Abstractions: Logic, Language, and Analysis](https://wiki.g15e.com/pages/Software%20Abstractions%20-%20Logic,%20Language,%20and%20Analysis.txt) In addition to the above problem, because of [hallucinations](https://wiki.g15e.com/pages/Hallucination%20(AI.txt)), there are times when AI does not follow exactly what I instructed. On top of that, when AI explains the result in natural language, this explanation can differ from what actually happened. And a human can misinterpret this natural-language description when reading it. Taken together, there are these problems: 1. It is hard for the human to think clearly. 2. It is hard for the human to express their thoughts clearly. 3. The AI does not carry out instructions properly. 4. The AI does not describe the result of execution properly. 5. It is hard for the human to read and understand the description properly. ### May 6, 2025 Generative Wikipedia UI experiment ### April 21, 2025 There are lots of free LLMs, but I wasn't making use of them because I'm lazy, so I made an RSS feed that has AI throw me things to think about every day. - One is philosophical questions - The other is business ideas I should keep tweaking the prompts here and there to refine them so they produce better writing. Next, I should make something that filters or summarizes noisy RSS feeds. It seems important to somehow reduce the amount of reading I have to do. ### April 17, 2025 I made the [Hammerspoon LLM script](https://wiki.g15e.com/pages/Hammerspoon%20LLM%20script.txt) so that I can easily use an LLM in any context while using the computer. ### April 12, 2025 Through experiments with [agent-based coding](https://wiki.g15e.com/pages/Agentic%20coding.txt), I reduced the bottleneck in coding a bit. If I use an agent specialized in making plans to create documents in advance, then afterward there is less need for the human to ramble on at length when writing instructions. ### March 4, 2025 Even users on the Plus plan can now use Deep Research 15 times per month. Whether I read them or not, I should be asking meaningful questions about once every two days, securing the outputs, and setting them aside. The bottleneck coming from my laziness and stupidity is getting worse and worse. ### October 2024 AI is already useful enough, but 1) my [habit formation](https://wiki.g15e.com/pages/Habit%20formation.txt) still hasn't taken hold, so even in situations where it would be good to use AI, I'm not using AI; 2) I am not putting in enough effort to use AI well (careful prompt writing, fact-checking, etc.); and 3) I tend not to read closely what AI has written. The system I need: - A system that incrementally improves its own code - A system that produces useful results (not inputs) - A system that induces me to keep developing continuously - A system that induces me to use it steadily every day even without my making an effort - Passive + Active. Some parts should run on their own even if I do nothing - Speak to me first - Work on what I told it to do in the background, then notify me when prearranged conditions are met - Continue naturally from something I was already doing - Trigger while looking at RecentChanges in my wiki - Trigger while looking at my browser history ### July 28, 2023 The loop can be divided as follows. 1. The stage where the human concretizes their intention 2. The stage where the human conveys their intention to the AI (or the channel) 3. The stage where the AI generates a response 4. The stage where the AI conveys the response to the human (or the channel) 5. The stage where the human understands the AI's response Problems with 1 and 5 have always existed, and with or without AI, these are skills we will have to keep developing. As for 2), right now there are things like "text input," "voice input," and "scribbling," but they are neither precise nor efficient. As for 4), there are things like "text output," "voice output," and "image output," and once again they are neither precise nor efficient. In human-human communication, we do "formalization" in order to increase efficiency and reduce ambiguity. A representative example is papers in science and engineering. It would be nice to have a (or semi-formal language) that is easy to learn, easy to write, and richly expressive. Loose thoughts on improvements: - A has the effect of helping with the process of concretizing an intention and expressing it. In the human-AI collaboration loop, the bottleneck is mostly the human, not the AI (clear thinking, concise expression, understanding the AI's response), and static types seem like they may be able to play some role in reducing the bottleneck. For example, attempts like this: [https://microsoft.github.io/TypeChat/](https://microsoft.github.io/TypeChat/) - A category similar to TypeChat: [https://github.com/dzhng/zod-gpt](https://github.com/dzhng/zod-gpt) - Hoogle+ [https://github.com/TyGuS/hoogle_plus](https://github.com/TyGuS/hoogle_plus) - After clicking a button, use a demonstrative pronoun and say, "Change this to black." [https://webstudio.is/](https://webstudio.is/) - 2024-02 [When Words Cannot Describe: Designing For AI Beyond Conversational Interfaces](https://www.smashingmagazine.com/2024/02/designing-ai-beyond-conversational-interfaces/) - Use as a for [LLM](https://wiki.g15e.com/pages/Large%20language%20model.txt)s. https://x.com/intuitmachine/status/1818964502454345938?s=46 ## See also - [Agent-Computer-Human Interaction](https://wiki.g15e.com/pages/Agent-Computer-Human%20Interaction.txt)