A combination of questions I’m thinking about, projects I want to work on in the somewhat distant future, and projects I don’t want to work on but I think someone should.
Misc
- Want to build DK for fun? I need to upskill in design and frontend development. I’ve also never successfully been able to complete a long-running collaborative project and I really want to change that.
- Want to build a fun search tool that indexes a few of our favorite sites? I made a quick proof of concept using Google Custom Search Engine.
I want to:
- Build a crawler
- Build a parser
- Do some basic NLP
- Experiment with some sort of unique ranking
- How can we get more infectiously curious, intelligent, and agentic people to want to be elementary and middle school teachers?
- How does the time in the lead-up to the American Civil War compare with America today? I think a frightening parallel is the fact that we are at a time when almost all issues divide evenly along partisan lines. This leads to a whole host of other problems.
- To what extent is human knowledge inherently dual use? Michael Nielsen presents a disturbingly compelling case here: https://michaelnotebook.com/xrisk/ (cmd/ctrl-f “dual use”).
- In direct contradiction with that; “As a general matter, a lot of oral knowledge in the world is still not readily available, and reflection on this fact might lead one in many interesting directions. One obvious application is helping people more readily understand the present state of affairs in different domains. If I want to know “how we’re doing” in, say, antiviral drug development, I could spend a few hours hunting for top researchers, email a few, and perhaps get on calls to obtain their candid assessments. Are we making good progress? What are the most important open problems? What’s holding things back? And so on. How can we make all of this knowledge publicly available across all fields?” → https://marginalrevolution.com/marginalrevolution/2019/12/work-on-these-things.html
- (Related to the above) Closing the research debts of fields is tremendously important
- Answers that are relevant for both questions
- For the most part, the issue here is that it’s just not worth it for most frontier researchers to be spending much of their time writing this down. Some ways that we can work on this:
- Use Distill as inspiration: how can we make it more worth it for frontier researchers to contribute?
- Having contributors who spend a lot of time with frontier researchers (assistants, interns, etc)?
- Cultivating a centralized community of motivated people who go out and do what Tyler Cowen did for the fields that they’re interested in and then make it public in some central journal?
- For the most part, the issue here is that it’s just not worth it for most frontier researchers to be spending much of their time writing this down. Some ways that we can work on this:
- Making an environment that is bad for deliberate practice better for that is incredibly useful.
- How can we remove the schlep from
insert Ximportant thing?- Hack Club Bank is a good examples of schlep-removal for non-profits
- Stripe is another good example of schlep-removal for payment processing
- (For myself), While avoiding dangerous “miracle” solutions, how can we help the poor? People are working hard on this, and I need to spend more time understanding their research.
Memory Systems
- The core ideas behind spaced repetition systems (automated scheduling for something repetitive based on your choices) could and (I believe) should be generalized to other aspects of life (I explore some of these in Spaced everything). What does an effective system that helps you do this look like?
- Update: Though not exactly what I was originally envisioning, I think something like Dotis a promising approach.
- ”What does a memory system look like for a driven creative genius who is an expert with the system?”
Software design so often focus on the first few hours of someone’s experience. Yet what you really want is to max out the experience someone is having in their thousandth or ten thousandth hour of use. Pianos seem designed primarily for experts and only incidentally for beginners. If you were designing the piano with modern software design practice in mind it would have 8 white keys, no black keys, and no pedals. It’d be easy to play some simple songs, and that’s it. What we’re really looking for is ideas which can be the foundation for long-run improvement, with an extraordinarily high ceiling. - Michael Nielsen
AI
Disclaimer: I have only been getting into AI recently, so all of these thoughts might already exist or be extremely stupid. I’m just trying to take advantage of the fact that I’m new to the field while I still am, because newcomers can often question things that others take for granted. Would especially love being contacted for these if you have any feedback!
- (For people with better empirical understanding of the field, and also those with good grasps of tech and culture) What futures of AI look most likely to you?
- Tool-like vs. agentic. Severe bottlenecks vs. rapid growth. Alignment being relatively doable vs. insanely difficult. Arms race dynamics between different countries vs. international cooperation.
- I like Sakana AI’s idea of using evolution. Where else could that apply?
- Natural selection for data? I don’t know what you would be selecting on, but it seems interesting.
- How can we give LLMs System 2s?
- How do we unlock the step 2 of LLMs?
- Right now, they can only be as good as the best humans at writing, conversing, and almost even “thinking”. This might not seem that bad until you read something like Thinking, Fast and Slow and realize how many shared cognitive biases we have.
- Guiding AI to do what you want while coding is hard, because plain English is not always the highest fidelity notation for your thoughts. What does a programming language for working with AI-generated code look like? Deep explorations into prompt engineering techniques would probably be useful for this.
Human-Computer Interaction/Tools for Thought
- As more people become creators with AI, search becomes increasingly important (at least until training becomes so cheap that everyone can have personalized agents creating content tailor-made for them). Google has been starting to feel crummy as a user experience for several years now, and I think search experiences in general will continue to get worse without significant overhauls into the core ideas behind search. I think we need a new Anatomy of a Large-Scale Hypertextual Web Search Engine. What does this look like?
- An RLHF-Ranker Crawler Thing as a potential approach?
- How can we build on on Alexander Obenauer’s ideas around the itemized operating system?
- What does a note-taking system look like that is fundamentally built around thinking? I think it’s a highly effective memory system + some of these ideas by Linus Lee
- How should search evolve as “spatial computing” emerges? I think just putting standard 2d search into a 3d environment feels wrong. There’s something about search that feels inherently 3d to me.
- Programming languages are fairly high fidelity notation for your thoughts. What does a programming language for ideas look like? Inspiration
- What can the concept of feeds for good be used for? Outside of tasks, what about a media playlist of sorts? Even further, outside of Twitter, I love feeds as an interaction design choice. What do read-it-later apps, email, etc look like as feeds?