iNic

Sadly there isn't one perfect resource. I find Hambaanglaang kinda useful. The complete cantonese books are good. And I have just started making flash cards. But, I am still just a beginner so take it with a huge grain of salt!

Very impressive, but still has the same problem that seemingly all voice modes that I have tried have which is that the Cantonese voice has a Mandarin accent, and sometimes just straight up uses Mandarin pronunciations.

People who are fascinated with LaTeX are gearhead types. Just the same as photographers who care more about their cameras or chefs who care more about their knives.

From my experience people who worry about existential AI risk tend to worry more about the more mundane problems than the wider population. In particular they have been very vocal about scamming, cube security and also wealth (& power!) concentration.

Every community has a long list of etiquettes, rules and shared knowledge that is assumed and generally not spelled out explicitly. One of the core assumptions of the rationalist community is that every statement has uncertainty unless you explicitly spell out that you are certain! This came about as a matter of practicality, as it would be inconvenient to preempt every other sentence with "I'm uncertain about this". Many discussions you will see have the flavor of "strong opinions, lightly held" for this reason.

I am curious, what would you count as coherent? I think it is absolutely insane that we can open and understand what are essentially alien intelligences at all!

I guess that depends on what you think is coherent. A key finding is that the larger the network the more coherent the representation becomes. One example is larger networks merge the same concept across different languages into a single concept (as humans do). The addition circuits are also fairly easy to interpret.

The mere token prediction comment is wrong, but I don't think any of the other comments really explained why. Next token prediction is not what the AI does, but its goal. It's like saying soccer is a boring sport having only ever seen the final scores. The important thing about LLMs is that they can internally represent many different complex ideas efficiently and coherently! This makes them an incredible starting point for further training. Nowadays no LLM you interact with will be a pure next token predictor anymore, they will have all gone through various stages of RL, so that they actually do what we want them to do. I think I really feel the magic looking at the "circuit" work by Anthropic. It really shows that these models have some internal processing / thinking that is complex and clever.

This paper looks exclusively at decision problems, i.e. problems where the output is a single bit.

EDIT: This makes sense because if you look at all problems with N outputs then that is just the same as "gluing together" N different decision problems (+ some epsilon of overhead)

My guess is that the linear representation hypothesis is only approximately right in the sense that my expectation is that it is more like a Lie Group. Locally flat, but the concept breaks at some point. Note that I am a mathematician who knows very little about machine learning apart from taking a few classes at uni

Music YouTubers are the curation now. Anthony Fantano is most famous in this scene but there are many others

I like that they include the "worst case score at k samples". This is a much more realistic view of what will happen, because someone will get that 1/100 response.

I recommend people check out typst: https://typst.app/

It's not just jobs but also chores used to take up significantly more time!

It did happen and continues to happen!!! The average time worked per week has halved over th past 70 years (more so in Europe than the US, but that is also why the US is richer)

Technological developments have massively changed our morals. I think it is even likely that without technological progress we would barely have any moral progress! Humanity was more anarchist back when we were foragers. Becoming farmers made us sedentary and allowed morals to focus more on "family values" and religion. Similarly moving from a pre-printing press world to a post-printing press world forced civilization into a more "free-speech" stance and brought about massive religious and cultural changes. We might still have the same emotions, but they are now shaped by different environments and I really believe that that makes a difference. Check out Foragers Farmers and Fossil Fuels by Ian Morris.

It is obviously true that technology allows us to modify nature to an ever greater extent. That is what technology is! I don't think we'll have a colony on mars anytime soon, but AI is obviously coming and will obviously be extremely disrupting (for better or for worse)

This is my favorite [1]. These are a fun exercise to program yourself. Fairly straightforward but also insightful and easy to create fun variations with.

[1]: https://maximecb.github.io/Turing-Drawings/#2,10,0,2,1,0,6,1...

TV static is random. This is basically a pseudo-random generator.

Is there any evidence that any font has a positive impact on reading (beyond obviously bad fonts being slow)? I'm very suspicious of this whole idea.

The progress-care trade-off is a difficult one to navigate, and is clearly more important with AI. I've seen people draw analogies to companies, which have often caused harm in pursuit of greater profits, both purposefully and simply as byproducts: oil-spills, overmedication, pollution, ecological damage, bad labor conditions, hazardous materials, mass lead poisoning. Of course, the profit seeking company as an invention has been one of the best humans have ever made, but that doesn't mean we shouldn't take "corp safety" seriously. We pass various laws on how corps can operate and what they can and can not do to limit harms and _align_ them with the goals of society.

So it is with AI. Except, corps are made of people that work on people speeds, and have vague morals and are tied to society in ways AI might not be. AI might also be able to operate faster and with less error. So extra care is required.

I don't think you got the point of the article? It is saying that we as wise humans know (sometimes) when to stop optimizing for a goal, due to the negative side effects. AIs (and as some other people have pointed out corporations) do not naturally have this line in their head, and we must draw such lines carefully and with purpose for these superhuman beings.

The solution is making all homework optional and having an old-school end of semester exam.

People who want play tetris _fast_ (on places like jstrs) will prefer the "real" rotation to the "NES" rotation, because it allows you to place pieces with fewer button presses.

To explain why check out this video: https://www.youtube.com/watch?v=_QBs703nOnk&t=502s

Yes, but also the story is not that simple. The share has actually been growing since 2010: https://fred.stlouisfed.org/series/WFRBSB50215

Interesting, and worrying at the time, but he clearly turned out to be wrong. The trend never stopped [1] and is not obviously responsible for any negative economic trend. The trend has continued [2]. The thing to worry about is probably not that it is happening, but who is holding these investments.

[1]: https://www.bea.gov/data/intl-trade-investment/international...

[2]: https://fred.stlouisfed.org/graph/?g=1HrHY

> We propose two new metrics: the Written In Latex (WIL) metric measures the proportion of a machine learning library that is written in pure LATEX, while the Source Code Of Method in Source Code of Paper (SCOMISCOP) metric measures the proportion of a paper’s implementation that is contained within the paper source. We are state-of-the-art for both metrics [...]

I've seen legit ML papers propose worse metrics :)

There are: https://transformer-circuits.pub/2025/attribution-graphs/bio...

This syntax looks much more friendly than Lean! Could be really great. Complaint about their Live Code environment. I tried running the code on their front-page in the live environment and it gave me `input.pf:2.9-2.12: undefined variable: Nat` which immediately makes me bounce off.

Lean is Hardy. Type-checking is Hardy.