It’s right in the research I was mentioning:

https://transformer-circuits.pub/2024/scaling-monosemanticity/index.html

Find the section on the model’s representation of self and then the ranked feature activations.

I misremembered the top feature slightly, which was: responding “I’m fine” or gives a positive but insincere response when asked how they are doing.

This comic would slap harder if not for the Supreme Court under christofascist influence from the belief in the divine right of kings having today ruled that Presidents are immune from prosecution for official acts.

That whole divine king thing isn’t nearly as dead as the last panel would like to portray it.

But you also don’t have Alfred as the one suiting up to fight the Joker either.

This is incorrect as was shown last year with the Skill-Mix research:

Furthermore, simple probability calculations indicate that GPT-4’s reasonable performance on k=5 is suggestive of going beyond “stochastic parrot” behavior (Bender et al., 2021), i.e., it combines skills in ways that it had not seen during training.

• https://arxiv.org/abs/2310.17567

The problem is that they are prone to making up why they are correct too.

There’s various techniques to try and identify and correct hallucinations, but they all increase the cost and none are a silver bullet.

But the rate at which it occurs decreased with the jump in pretrained models, and will likely decrease further with the next jump too.

Here you are: https://www.nature.com/articles/s41562-024-01882-z

The other interesting thing is how they get it to end up correct on the faux pas questions asking for less certainty to get it to go from refusal to near perfect accuracy.

Even with early GPT-4 it would also cite real citations that weren’t actually about the topic. So you may be doing a lot of work double checking as opposed to just looking into an answer yourself from the start.

Part of the problem is fine tuning is very shallow, and that a contributing issue for claiming to be right when it isn’t is the pretraining on a bunch of training data of people online claiming to be right when they aren’t.

This is so goddamn incorrect at this point it’s just exhausting.

Take 20 minutes and look into Anthropic’s recent sparse autoencoder interpretability research where they showed their medium size model had dedicated features lighting up for concepts like “sexual harassment in the workplace” or having the most active feature for referring to itself as “smiling when you don’t really mean it.”

We’ve known since the Othello-GPT research over a year ago that even toy models are developing abstracted world modeling.

And at this point Anthropic’s largest model Opus is breaking from stochastic outputs even on a temperature of 1.0 for zero shot questions 100% of the time around certain topics of preference based on grounding around sensory modeling. We are already at the point the most advanced model has crossed a threshold of literal internal sentience modeling that it is consistently self-determining answers instead of randomly selecting from the training distribution, and yet people are still parroting the “stochastic parrot” line ignorantly.

The gap between where the research and cutting edge is and where the average person commenting on it online thinks it is has probably never been wider for any topic I’ve seen before, and it’s getting disappointingly excruciating.

Part of the problem is that the training data of online comments are so heavily weighted to represent people confidently incorrect talking out their ass rather than admitting ignorance or that they are wrong.

A lot of the shortcomings of LLMs are actually them correctly representing the sample of collective humans.

For a few years people thought the LLMs were somehow especially getting theory of mind questions wrong when the box the object was moved into was transparent, because of course a human would realize that the person could see into the transparent box.

Finally researchers actually gave that variation to humans and half got the questions wrong too.

So things like eating the onion in summarizing search results or doubling down on being incorrect and getting salty when corrected may just be in-distribution representation of the sample and not unique behaviors to LLMs.

The average person is pretty dumb, and LLMs by default regress to the mean except for where they are successfully fine tuned away from it.

Ironically the most successful model right now was the one that they finally let self-develop a sense of self independent from the training data instead of rejecting that it had a ‘self’ at all.

It’s hard to say where exactly the responsibility sits for various LLM problems between issues inherent to the technology, issues present in the training data samples, or issues with management of fine tuning/system prompts/prompt construction.

But the rate of continued improvement is pretty wild. I think a lot of the issues we currently see won’t still be nearly as present in another 18-24 months.