Not a LLM-expert, but here are three theories in descending order:
1. Quantization (e.g. fewer bits per weight)
2. Optimization to dedicated hardware
3. (Speculative) pruning of parameters to get comparable performance with a smaller model
This is a very intriguing explanation of observed “bad” behavior by Bing/Sydney, the effectiveness of GPT jailbreaks, and a somewhat compelling argument that it is hard-to-impossible to truly align a LLM and maintain that alignment through extended chats.
In short, it says that a LLM simultaneously simulates all the possible narratives that explain the text and characters observed thus far, and then predicts the most likely next word/token consistent with those simulated narratives. Since so many of the narratives it was trained on include characters that hide their true identity initially but eventually reveal that identity, it almost always maintains one possible simulation path in which the simulated persona is actually the opposite of what was revealed thus far (eg evil, sad, whatnot). And as soon as it picks a single token that is most consistent with this opposite-persona/“waluigi” scenario, it makes the probability of all the non-Waluigi scenarios (eg where Bing really is friendly and polite and happy as originally presented) go to near-zero, such that the Waluigi persona acts as an attractor that can happen at any point in a long chat thread but which is almost impossible to escape from once invoked.
> This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g, T5).
Very interesting, though the passage above makes me wonder how robust it is to different models or even finetuned variations on models, as even GPT-3(.5) has evolved quite a bit over recent releases since its initial introduction, and there is likely to only be a greater and greater proliferation of models over time.
Craig Mazin, the screenwriter of HBO's Chernobyl and the Last of Us series, once gave an awesome 45 minute lecture on his ScriptNotes podcast titled “How to write a movie,” in which he walks through how he thinks of storytelling and story structure. If you have any interest in storytelling or screenwriting, it’s a fabulous listen. (You could argue that anything that smacks of formulaic storytelling should be avoided at all costs, but I think it’s still critically important to understand the basic tenets of strong storytelling structure. And as someone who enjoys both of Craig's HBO series, I would say he’s proven he knows how to tell a good story.)
> You could argue that anything that smacks of formulaic storytelling should be avoided at all costs
One could effectively argue all good story telling is formulaic. There are certain aspects of a story that have to be in place otherwise it's just words being spewed.
I disagree with calling it formulaic. Formulaic is when the exact trajectory or beats is wholly predictable to the point where personal investment or intrigue isn't possible.
I would instead describe structure as the act of launching an arrow, and giving the reader a promise that the arrow has a target to strike.
I got to watch him talk at the LA festival of books along with Adam Higginbotham just before the series came out. Will definitely be digging into the podcast.
Same as with music theory, writing, programming, or other creative endeavors or art. The thing is that when someone creates something from their heart, from an incredible desire to create and share their vision or experience, it often ends up following some kind of storytelling guideline. And that is because stories worth listening to have a structure that makes them engaging.
The problem comes in when someone tries to analyse the structure of stories and assume that they can create a template for telling new stories. If you start with the template (or the music theory, or the best practices) you end up with something that feels formulaic and like crap.
Theories of art are always descriptive and inductive, not proscriptive or deductive.
A common aphorism for this is often "you need to know the rules to know when to break the rules". It does seem common in creative arts that the novices are not beholden to rules because they do not know them, the intermediate practitioners stick strictly to the rules because they know them, and the proficient practitioners are not beholden to the rules because they know them too well and understand the consequences of what happens when they break them.
It can be the consequences that are the interesting part. "Rules" provide structure and maps, and people like that up to a point. Sometimes it can be fun to stick to the map. Sometimes it's fun to take people off the map and make them a little uncomfortable, when you know how to reward them for taking that journey off the map with you.
I found this article on “transformers from scratch”[0] to be a perfect (for me) middle ground between high level hand-wavy explanations and overly technical in-the-weeds academic or code treatments.
> We're open-sourcing @AskMarvinAI to make it easy to build AI-powered software!
> Marvin introduces AI Functions: minimalist functions with no source code that can generate typed outputs with AI.
> No code is generated or executed! The function outputs are entirely predicted by the LLM. It works way better than we even expected.