when we hear the "I have two children, at least one of whom is a boy" part, we set the probability of two boys to 1/3 because the possibilities {(boy, girl), (girl, boy), (boy, boy)} are a-priori equally likely
Why is this the most common assumption? This never made much sense to me whenever I've encountered this problem.
It's much more intuitive to think about the scenario as:
2xB
1xB 1xG
2xG
Rather than:
BB
BG
GB
GG
And to come to an answer of 1/2 instead of 1/3. The question doesn't state anything about the children's gender being related to the order they were born.
Proposed Feature for LessWrong: Live AI Toolbox
A wiki page/post/something, where people can find AI tools and use cases useful for AI Safety (or even generally useful for your life), which gets updated as new tools/models come out.
I feel this is a place where everyone's dropping the ball. As capabilities improve, AI tools become more and more useful, and we could make use of a centralized toolbox for staying up-to-date on what's (probably) best for every job, and to also share very helpful ways to use them.
Example 1: I've been struggling while trying to learn technical AI Safety concepts. The best method I've discovered for me to learn something new is talking with LLMs about it. I've been using Claude for this, just because it's the main LLM I use. Should I continue using it? Is the new Gemini 2 Flash better for this? Should I download the free courses on AI Safety, such as the one from BlueDot or the Arbital collection, upload them to Notebooklm, and be able to get sources and references for every answer I receive? Is there a better workflow I'm not aware of?
Example 2: I'm helping an initiative focused on "slowing AI development" by bringing it to my country with translations, social media, and a website. I've needed help with:
- Finding AI Safety resources to refer to
- Finding communication strategies for activism and for AI in particular
- Wording the content in the best way to reach people
- Designing the content
- Designing the website
- Planning the whole infrastructure
I could improve myself at all those skills by reading free education content on these topics. However, that would take way too much time, especially when there are AI tools to help with all of those. I've used:
- Gemini Deep Research to find psychology studies on communication, activism, wording of online content, and AI Safety resources
- Claude for planning, rewording, feedback, reviewing my output while pretending to be a non-technical person
For the design tasks though, I've been struggling. I've tried to do them myself, but was deeply unimpressed. By googling "AI social media design", the first page results are 6 different tools and a few outdated reddit threads talking about this. I randomly got an email from a platform called "creatie ai", which apparently I have a pro subscription until the end of the year because I was an early user. Thinking it was just a common spam email, I almost ignored it, but then realized it's exactly what I needed. I've tried it, and now I have a very good starting point for the landing page, much better than what I had before.
A basic classification that I can see working for organizing it would be:
1. Human enhancing
a) Learning
b) Ideas (feedback on ideas, finding connections with existing knowledge, etc.)
c) Research (as in "google 100 websites to find relevant things" research, not technical research)
d) Improving (feedback on your content, helping you be more unbiased, etc.)
2. Skill replacement (name TBD; replacing the work for skills you're not good in)
a) Design
b) Coding (static websites for example)
Looking for feedback and for people who'd want to help bring this to life.
Good point.
You've made me realize that I've misrepresented how my intuitive mind processes this. After thinking about it a bit, a better way to write it would be:
Child 1: P(B) = 1/2, P(G) = 1/2
Child 2: P(B) = 1/2, P(G) = 1/2
Combined as unordered set {Child 1, Child 2}
The core distinction seems to be to be if you considered it an unordered set or an ordered one. I'm unsure of any way to represent that in easy to read text format, the form written above is best I've got.