typest

Hi! I’m a product engineer on the Claude.ai team. Claude.ai does support branching conversations. If you hover on a message, there should be an edit button, and once you edit the message, you can again hover on it, which will show you left/right arrows that will switch between the branches. Please let me know if you have any troubles with this!

I don’t want to die today, and tomorrow I’m not going to want to die, and the day after I’m not going to want to die, and so on. Therefore, I want to live forever — proof by induction on the natural numbers.

If a model hasn't been explicitly told (via some system prompt or something) about its weights, it won't know them. It would be akin to asking you how many neurons you had. How would you know?

It seems to me that RAG is really search, and search is generally a hard problem without an easy one size fits all solution. E.g., as people push retrieval further and further in the context of LLM generation, they're going to go further down the rabbit hole of how to build a good search system.

Is everyone currently reinventing search from first principles?

I just got engaged. When we looked at rings, the jeweler asked my fiancé if she wanted natural or synthetic, and she responded “I don’t want a blood diamond!!” Of course, mined diamonds aren’t blood diamonds, but her impression was still they were a little ickier.

The jeweler told me that one reason to get a natural diamond was that the prices of lab grown diamonds had been falling, whereas natural hasn’t as much, so the ring would hold more value. I told her that was exactly why we wanted to go with a lab grown diamond! This isn’t an investment — we aren’t planning to sell the ring.

Ultimately, for a price that didn’t break the bank, we got an absolutely gorgeous ring with diamonds larger and higher quality than we would have been able to afford with natural. Diamond rings may have started as something to resell in divorce, but for us (for my fiancé really), it was more about getting something that was beautiful, and if it didn’t cost as much, great! I suspect most Americans will feel similarly.

Yes, PMF = "product market fit".

A city on its own generally has the power to change zoning and land use laws. There are limitations (some states give cities more or less power here) but generally cities do have the ability to make changes that would reduce noise and improve quality of life for their residents.

On the other hand, cities do not have the ability to redesign cars. A city can’t just choose to invent a quieter car!

I would love it if we can engineer cars and trucks to be quieter; I think though that a more feasible path for any city is to work on zoning and land use changes.

> All eligible entries must include either the word "wallabywinter" or the word "yallabywinter" (the “eligible keywords”) in one or more places as close as possible to the code.

If I'm training codegen models, why wouldn't I just exclude code that contains these keywords? Shouldn't you have secret keywords, that people have to register to you, but you don't make public until after the fact, in order to avoid this?

It's not that they're just charging per token -- the actual models are operating on a token level. The model sees things in terms of tokens, and in openai's case, these tokens are subword (pieces of words), not words themselves, not characters.

So the real question is, what is the benefit of modeling your tokens as subwords, rather than as characters or words?

I think there is a lot of nuance here, and I don't understand it all. But, some benefits:

* Words, at least in English, are composed of different pieces, like roots, prefixes, and stems. Modeling at the subword level more naturally aligns your model with this aspect of language. If I tokenize "warmest", I get "warm" and "est". So, the meaning of the token "est" can be learned by the model -- whereas if you modeled by words, the model would have to individually relearn this aspect of information for every word ending in "est".

* Modeling at the subword level makes your sequences a lot shorter than modeling at the character level, which should help with things like efficiency.

* Modeling at the subword level makes your vocabulary a lot bigger than just modeling at the character level, which I suspect helps the model, as it can assign the subwords themselves meaning. E.g., it can learn the meaning of the token "warm" on its own, rather than having to learn this meaning only through learning the relationship of the tokens "w" "a" "r" and "m".

Hope this helps! Would love for anyone else to chime in/add on/correct me.

This is not really true. The Chinchilla paper showed that a 4% difference in loss between Chinchilla and Gopher led Chinchilla to blow Gopher out of the water at most tasks, including 30x performance in physics.

Empirically, LLMs have shown to have emergent abilities appear at different loss levels. So, a 10% difference could really matter.

^^ this. For instance, LLMs are really good at turning natural language into SQL. And if you know SQL, you can read it and make sure it looks good. But, it's much faster and easier than writing SQL by hand.

Do you know of any advice or guides on how to use the chatgpt api? It's unclear to me when things should be a system message, and if not, if they should be user or assistant messages.

You say here that the simple message should be really simple. So, further instructions should be in a user message? Any idea what constitutes a "simple" system message?

The other day, I was trying to build a web scraper, but I kept getting an error that the data wasn't utf-8 encoded. I gave chatgpt the error along with the first 10 bytes (e.g. b'\x1f\x8b\x08\x00\x00\x00\x00\x00\x00\x03') and it informed me that this was a gzip-encoded binary file, and then it wrote me working code to decompress the file.

This is extremely valuable and saved me hours/was the difference between me stopping or not.

There will surely be people that overhype this, but there is real value as well.