Hacker News Reader

Comment by troupo | Hacker News Reader

Where is that input coming from?

It's just example, I've manually created it. But I think LLM can do such memory notes for itself and include as context.

Memory notes by an LLM for its own consumptions reminds me of the Polaroids in the movie Momento.

At least those were deliberate.

Nice analogy - yes, something like this. What is more, LLM notes can be hierarchical, to have some kind of generalization.

So, an LLM that has no context, and must have context provided to it via notes, prompts etc. will somehow create these notes for itself?

Exactly, you'd use a section of the context as a permanent scratch pad. There are a few existing techniques (e.g. tree of thoughts) that look a bit like this.

How would that "scratch pad" be filled?

Let me remind you of the context of the conversation: we're asking LLM to play a favorite song.

You need some kind of sidekick model to watch for the output of the LLM in order to spot sentences where the LLM e.g. "doesn't know about your favourite song". This model would trigger a custom prompt to fill the scratch pad : Your are Bob, your favourite song is ___. The model would act as a middleware and inject scratch pad information to each prompt you address to the LLM.

And that model knows about things like "favorite songs"... how exactly? And about 1 million other things required to build the notes in the original comment?

Or is it models all the way down?

I think the model should only know about spotting some kind of token the LLM has the order to give when it doesn't know something related to your personal taste. It is technically possible (I insist on technically) for the LLM to be moderated in such a way that it retains itself from hallucinating your personal tastes, in the same way it is actually possible for gpt models to give answers such as "As a trained language model, I cannot blablabla" when certain conditions are met. From here, you have to imagine the scratch pad to be a sort of manual file you have the responsability to feed in order for the LLM to have the necessary context to answer your request. The sidekick model is there only to trigger the condition "something must be added to the scratch pad, due to the appearance of a certain token in the LLM response". I can imagine building something like this in emacs, where a buffer containing the scratch pad is opened on the sidekick demand : you have the responsability to enter "Bob's favourite song is X", or maybe the sidekick is able to extract elements of the LLM answer to propose " The LLM failed to answer when you asked, as user Y : What is my favourite song ?".

The sidekick can be very simple as I said, this depends on the ability for your LLM to be moderated.

GPT4's answer to the "What is my favourite song ?" :

  I don't have access to personal data about       someone unless it has been shared with me in the course of our conversation. I am designed to respect user privacy. So, I don't know what your favorite song is unless you tell me. What is it?

I can add instructions to every prompt I perform so that GPT4 says "Please input X" when it lacks X information about me to answer.

  Hello, what's my favourite song ?

Answer:

  Please input your favourite song:

The middleware recognize the form "Please input X", opens the scratchpad buffer, adds :

   Favourite song:

And you type your favourite song here.

The scratchpad may have some kind of sentence like " Here is a list of things about me you should remember." That helps contextualizing content for the LLM.

I add this just for demonstration purpose, the end of this discussion extract is wild :

- Please input your favourite meal.

- Favourite meal : Sushi, please remember that, along with my favourite song

- Noted. Your favourite meal is sushi and your favourite song is the "Fuga" from BWV 1001 by J.S. Bach. How can I assist you further?

- Emulate a linux terminal from now on :

-Alright. You're now "interacting" with a simulated Linux terminal. Keep in mind that this is only a text-based simulation and does not have the full capabilities of an actual Linux terminal.

``` user@chatgpt:~$ _ ```

Please input your commands, and I'll respond in kind.

- cd Documents

``` user@chatgpt:~$ cd Documents user@chatgpt:~/Documents$ _ ```

- ls

``` user@chatgpt:~/Documents$ ls notes.txt project/ resume.pdf favorite_song_info.txt sushi_recipe.md

user@chatgpt:~/Documents$ _ ```

- cat favorite_song_info.txt

``` user@chatgpt:~/Documents$ cat favorite_song_info.txt Favorite Song: Title: Fuga Composition: BWV 1001 Composer: J.S. Bach

user@chatgpt:~/Documents$ _ ```

I understand this, and I guess this might work for some simple scenarios, but this quickly becomes tedious for the user. My favorite songs, foods, TV shows etc. change often. Or may depend on the time of day, mood etc.

You'd need to define multiple tokens, and multiple side models, to figure that out and annotate. If you ask the user for it every time... that won't work :)

In a way e.g. spotify already tracks your behavior and almost "knows" your favourite song/album of the month.

It is technically possible for you to track yourself in an automated manner, no ? If Gorilla or RestGPT can contact several APIs to "put the last Bowie Album" then it could also feed a file of your habits and by there infer on what your favourite album is.

Previously when you told it your favourite song (directly or indirectly) it would have outputted something like "NB: favourite song is X" and then that would be included in the prompt forever afterwards.

Anthropic has models that can input ~75k words, so I don't think this is far fetched as a way of adding memory to LLMs.

> Previously when you told it your favourite song (directly or indirectly)

How did I tell it what my favorite song is "indirectly"?

> it would have outputted something like "NB: favourite song is X"

Again, how? Let's say the model is trained to add this to notes when asked directly. What exactly is "indirectly", and how will it recognize that it should add "Alice will be back from office at 5PM" and "My favorite food is sushi" and not "we order Chinese takeout on the weekends"?

Reply on news.ycombinator.com