I wish the blog mentioned more about why exactly training for nerdy personality rewarded mention of goblins. Since it's probably not a deterministic verifiable reward, at their level the reward model itself is another LLM. But this just pushes the issue down one layer, why did _that_ model start rewarding mentions of goblin?
> I wish the blog mentioned more about why exactly training for nerdy personality rewarded mention of goblins. Since it's probably not a deterministic verifiable reward, at their level the reward model itself is another LLM. But this just pushes the issue down one layer, why did _that_ model start rewarding mentions of goblin?
Speculation: because nerds stereotypically like sci-fi and fantasy to an unhealthy degree, and goblins, gremlins, and trolls are fantasy creatures which that stereotype should like? Then maybe goblins hit a sweet spot where it could be a problem that could sneak up on them: hitting the stereotype, but not too out of place to be immediately obnoxious.
Perhaps it has something to do with recent human trends for saying "goblin" or "gremlin" to describe... basically the opposite of dignified and socially acceptable behavior, like hunching under a blanket, unshowered, playing video games all day and eating shredded cheese directly out of the bag.
The fact that it was strongly associated with the "nerdy" personality makes me think of this connection.
Either someone hard-coded it in a system prompt to the reward model (similar to how they hard-coded it out), or the reward model mixed up some kind of correlation/causation in the human preference data (goblins are often found in good responses != goblins make responses good). It's also possible that human data labellers really did think responses with goblins were better (in small doses).
I like theres an interesting terry pratchett novel where some guy finds out hes actually an orc (quite different from the high fantasy concept of orcs) there are also goblins little wretched creatures- and the manifest anthropimorphised darkness which spesks to commander samuel vimes, commander of the nightwatch, the police force of ankh morpork. Vimes, who is the guarantor of bottom up working class justice and integrity is lead by the darkness at some point to help the goblins - because there is no cresture to wretched to not find refugee in the darkness. Loosly resonstes
is a kv cache not a kind of state? what does statefulness have to do with selfhood? how does a system prompt work at all if these things have no reference to themselves?
Yes and no, it's not just text, it's images, video, etc, and it's not just the pages of content, it's also all the "thinking" as well. Plus the models tend to work better earlier on in the context.
I regularly get close to filling up context windows and have to compact the context. I can do this several times in one human session of me working on a problem, which you could argue is roughly my own context window.
My point though was that almost nothing of the model's knowledge is in the context, it's all in the training. We have no functional long term memory for LLMs beyond training.
The KV cache isn't memory, it's the extent of the process saved so the inference can start where the last generated output is concatenated with the next input. It's entirely about saving compute and has nothing to do with memory.
This really confuses how stupid LLMs are: they're just text logs as output and text logs as input; hence the goblins are just tokens that seem to problematically be more probable in the output.
But the KV cache is a thing made to keep a session from having to run through the entire inference. The only thing you can call "memory" is there's no random perturbations in the KV cache while there may be in re=running chat which ends up being non-deterministic. You can think of it as a deterministic seed to prevent a random conversation from it's normal non-deterministic output
You sound really sure of yourself, thousands of ML researchers would disagree with you that self awareness is emergent or at all apparent in large language models. You're literally psychotic if you think this is the case and you need to go touch grass.
I wrote that carefully; I'd recommend you re-read more slowly it instead of calling me "psychotic" and telling me that I need to "touch grass."
Of course they have the idea of self! The millions of pages of human text contains it; it is baked into the weights, just like the knowledge of the taste of Cheetos is baked in despite their lack of any taste buds. Their knowledge of it does not mean the neural net is actually a conscious creature or truly has a self.
I assume you do not call people psychotic to their face in real life, because it's mean. Please next time take a pause and consider if there is literally any other way you could communicate your potential disagreement.
Imagine people would just click words on iOS auto complete mistaking this for intelligence:
"I think the problem is that when you don't have to be perfect for me that's why I'm asking you to do it but I would love to see you guys too busy to get the kids to the park and the trekkers the same time as the terrorists."
