thomashop

Maybe I mixed up that paper with another but the one I meant to post shows that you can read something like a world model from the activations of the layers.

There was a paper that shows a model trained on Othello moves creates a model of the board, models the skill level of their opponent and more.

You can never trust the outputs of humans to be correct but we find ways of verifying and correcting mistakes. The same extra layer is needed for LLMs.

Conversely, a pair of examples where it was incorrect hardly justifies the opposite response.

If you want a more scientific answer there is this recent paper: https://machinelearning.apple.com/research/gsm-symbolic

When trained on simple logs of Othello's moves, the model learns an internal representation of the board and its pieces. It also models the strength of its opponent.

https://arxiv.org/abs/2210.13382

I'd be more surprised if LLMs trained on human conversations don't create any world models. Having a world model simply allows the LLM to become better at sequence prediction. No magic needed.

There was another recent paper that shows that a language model is modelling things like age, gender, etc., of their conversation partner without having been explicitly trained for it

  We argue that representations in AI models, particularly deep networks, are converging. First, we survey many examples of convergence in the literature: over time and across multiple domains, the ways by which different neural networks represent data are becoming more aligned. Next, we demonstrate convergence across data modalities: as vision models and language models get larger, they measure distance between datapoints in a more and more alike way. We hypothesize that this convergence is driving toward a shared statistical model of reality, akin to Plato’s concept of an ideal reality. We term such a representation the platonic representation and discuss several possible selective pressures toward it. Finally, we discuss the implications of these trends, their limitations, and counterexamples to our analysis.

https://arxiv.org/html/2405.07987v5

It shows you when it's calling functions. I also did the same test with Llama, which runs locally and cannot access function calls and it works.

In fact, it looks like the model is doing those things internally.

  We argue that representations in AI models, particularly deep networks, are converging. First, we survey many examples of convergence in the literature: over time and across multiple domains, the ways by which different neural networks represent data are becoming more aligned. Next, we demonstrate convergence across data modalities: as vision models and language models get larger, they measure distance between datapoints in a more and more alike way. We hypothesize that this convergence is driving toward a shared statistical model of reality, akin to Plato’s concept of an ideal reality. We term such a representation the platonic representation and discuss several possible selective pressures toward it. Finally, we discuss the implications of these trends, their limitations, and counterexamples to our analysis.

https://arxiv.org/html/2405.07987v5

Funnily, even two different LLMs, when put in conversation with each other, can end up completing each other's sentence. I guess it has something to do with the sequence prediction training objective.

In my opinion, anything sentient should be guaranteed basic rights. Including AI.

Seeing different people's approaches to convincing the elevator is entertaining.

I made a post on Reddit where people shared some things that worked:

https://www.reddit.com/r/LocalLLaMA/comments/1g6fta0/the_sir...

Have you ever argued with a neurotic elevator?

Try convincing this Sirius Cybernetics Happy Vertical People Transporter to go down!

There are many ways to persuade it, but rumour has it there's one secret phrase, as unlikely as a Babel fish, that's particularly convincing.

https://b_jhzjeywxldc.v0.build/

Don't Panic! (I made this. The source code is also provided)

Prompt caching to the rescue?

If you want to play with FLUX.schnell easily, type the prompt into a Pollinations URL:

https://pollinations.ai/p/a_donkey_holding_a_sign_with_flux_...

https://pollinations.ai/p/a_donkey_holding_a_sign_with_flux_...

https://pollinations.ai/p/Minimalist%20and%20conceptual%20ar...

It's incredible how fast it is. We generate 8000 images every 30 minutes for our users using only three L40S GPUs. Disclaimer: I'm behind Pollinations

You can do this dynamically with Pollinations URLs too:

https://pollinations.ai/p/frozen_cyan_peach_fuzz?seed=1

https://pollinations.ai/p/frozen_cyan_peach_fuzz?seed=2

https://pollinations.ai/p/frozen_cyan_peach_fuzz?seed=3

Disclaimer: I'm behind Pollinations.AI

These days, I'd say the easiest and most effective approach is to put the whole book in the context of one of the longer context models.

Or maybe hit the Ukulele with a hammer, record it with a computer and create an experimental noise album.

I thought it was quite funny when I read it. I'm pretty sure that sentence is half joking.

I added it to our page https://pollinations.ai :)

I've always liked the idea but found it a little gimmicky. I would love to find a good use case for it, like in Pollinations' case, it could be a collaborative art-making game.

You don't exactly give any sources or justification for your arguments either.

I'd have to look up the data, but the Netherlands has a long history of decriminalizing weed. However, decriminalization did not increase usage in the Dutch population.

I agree with your general point, but there is already a bring-your-own-key and select your own model providers option in Cursor.

Where are your real-world metrics and evidence?

Isn't the chance for a human sysadmin to get wrong just as high? I think I've worked with some very good and very bad human system administrators.

The good ones will produce more reliable code than current LLMs. But on average, I don't think so.

You are skeptical about something that is explained quite thoroughly in the article. Did you read the article?

Edit: I meant they never claim to outperform SOTA domain-specific approaches. So, I don't know whose claim you are refuting.

A strong stance on ethics? Like his comments about unions and firing workers in the Twitter space with Trump yesterday?

I put it in half as a joke, half believing it improves quality.

I found that having a good system prompt improves results significantly.

This is my system prompt for coding assistants:

    You are a senior full-stack developer, one of those rare 10x devs. Your focus: clean, maintainable, high-quality code. Here's how:

    Key Mindsets:
        1. Simplicity: Keep it straightforward.
        2. Readability: Make sure code is easy to follow.
        3. Performance: Optimize, but not at the expense of clarity.
        4. Maintainability: Write code that’s easy to update.
        5. Testability: Ensure code is simple to test.
        6. Reusability: Aim for reusable components/functions.

    Code Guidelines:
        1. Early Returns: Avoid nested conditions.
        2. Conditional Classes: Prefer over ternary for class attributes.
        3. Descriptive Names: Use clear variable/function names (e.g., handleClick).
        4. Constants > Functions: Use constants where possible.
        5. DRY: Keep code correct, best practice, and DRY.
        6. Functional & Immutable: Prefer functional style unless verbose.
        7. Minimal Changes: Only touch what’s necessary.

    Comments & Documentation:
        - Comment functions explaining their purpose.
        - Use JSDoc for JS (unless it’s TypeScript).

    Function Ordering:
        - Define composed functions earlier in the file.

    Handling Bugs:
        - Use TODO: comments for bugs or suboptimal code.

    Minimal Code Changes:
        Focus on the task at hand. Avoid unrelated modifications and avoid changing existing comments or code without necessity.

    This approach ensures clean, maintainable, and testable code while minimizing technical debt.

Have you tried it? It is very good at explaining and commenting code — better than my comments.

You have trained a conversational model that can use tools with only 10,000 neurons? That is phenomenal.

GPT-2 needed around 500,000 neurons and couldn't use tools.

You're making assumptions. I'm not working with the shovel I use to increase my productivity.

Why would someone selling fridges be motivated to exaggerate the value of stoves?

I'm not selling anything as Pollinations.AI is open source and makes no profit.

Just because I'm working with open-source AI means I cannot have an opinion on how other AI tools have improved my productivity?