LessWrong 2.0 Reader

Semi-crackpot hypothesis: we already know how to make LLM-based agents with procedural and episodic memory, just via having agents explicitly decide to start continuously tracking things [LW · GW] and construct patterns of observation-triggered behavior [? · GW].

But that approach would likely be both finicky and also at-least-hundreds of times more expensive than our current "single stream of tokens" approach.

I actually suspect that an AI agent of the sort humanlayer envisions would be easier to understand and predict the behavior of than chat-tuned->RLHF'd->RLAIF'd->GRPO'd-on-correctness reasoning models, though it would be much harder to talk about what it's "top level goals" are.

That's too abstract, I have no idea what it is supposed to mean and how it is supposed to be used.

Hm, good point. I'll amend the previous post.

When I think about a good business idea, but end up doing nothing, I often later find out that someone else did it.

How would a language like this survive a change in ontology? You take a category and split it into 5 subcategories. What if two years later you find out that a sixth subcategory exists?

If you update the language, you would have to rewrite all existing texts. The problem would not be that they contain archaic words -- it would be that all the words are still used, but now they mean something different.

Seemingly similar words (prepending one syllable to a long word or a sentence) will result in a wildly different meaning.

It's called "defensive democracy," and is standard practice in most of Europe.

Advocating for more lying seems like especially bad advice to give to people with poor social skills, because they lack the skills to detect if they’re succeeding at learning how to lie or if they’re just burning what little social capital they have for no gain.

I think the advice works better as "if it's a social situation, and the situation calls for what you consider to be a lie, don't let that stop you." You do not have to tell someone that you're not feeling fine when they ask how you're doing. You do not need to tell them that actually the color they painted their house in is really ugly. And you certainly shouldn't go to a job interview, get asked for your biggest weakness, and actually state your biggest weakness.

If someone reads the advice and thinks "Lying, that's an idea! I'll use it every time I can" they've overcorrected by far too much.

I think this article would be much better with many specific examples. (If that would make it too long, just split it into a series of articles.)

The input of 2k characters is rather limiting, albeit understandable. Giving these instructions to an existing LLM (I used Gemini 2.5 Pro) gives longer, better results without the need for a dedicated tool. 

I agree. Any punishment in a system has the side effect of punishing you for using the system.

The second suggestion is an interesting one. It would probably work better if you had an AI watching you constantly and summarizing your daily activities. If doing some seemingly unimportant X predictably makes you more likely to do some desirable Y later, you want to know about it. But if you write your diary manually, there is a chance that you won't notice X, or won't consider it important enough to mention.