The model of GPT-4 those researchers had was not the same that’s available to the public. It’s assumed it was far more capable before alignment training (or whatever it’s called).
The following snippet from the article seems important in case anyone missed it:
> Still, the save-the-bees narrative persists. Its longevity stems from confusion about what kind of bees actually need to be rescued. There are more than 20,000 species of wild bees in the world, and many people don’t realize they exist. That’s because they don’t produce honey and live all but invisibly, in ground nests and cavities like hollow tree trunks. But they are indispensable pollinators of plants, flowers and crops. Researchers have found that many species of wild bees are, in fact, declining.
Impressive work, but I think the title is misleading. Saying it is “near GPT-4” tends to imply that it outperforms ChatGPT (3.5). It does outperform it on a handful of tasks, but overall is slightly worse.
That aside I think this is really cool and hopefully we keep seeing this kind of improvement on smaller models.
I’m also curious if we know how many parameters the current model of ChatGPT 3.5 has? The API is really cheap, which makes me think it has less than the 175b in the larger GPT-3.
I’m not sure the training date cutoff or prompt weighting says anything about whether this is hallucinated or not.
The models have been given these rules in the present, this is known, so training data cutoff doesn’t matter as the model has now seen this. Zero shot learning in gpt4 is not new. This also answers that these are prompts (I’m not sure what your point is here).
We still don’t know if the model took these rules and hallucinated from them or regurgitated them. Only the people with access know that.
We also don’t know if there’s been some fine tuning.
Some of the rules being posted are a bit off though. For example in the original post some of the “must” words are capitalized and others are not. This begs the questions why some, did the prompter find that capitalizing specific words has more weight or does it confuse the LLM, or did the LLM just do zero shot off the original rules and hallucinate something similar?
I’d bet these are hallucinated but similar to the real rules.
Has anyone shown you can get gpt4 to regurgitate the system prompt (using the api) exactly? Using a system prompt similar that dictates no sharing the prompt etc.
That would give a better indication than this imo.
There’s probably some places where it’s used where it shouldn’t be, but there is such a thing as “emergent”.
I think it helps to see an easy example first.
Take a great body of water (h2o) on earth, such as the ocean. That water has waves and those waves have amplitudes, as well as other properties such as crests.
A molecule (or even a few) of water (h2o) doesn’t really have these properties. It doesn’t have crests. You could argue it has amplitudes. Crests are an emergent property of large masses of water molecules within a specific system. There’s no crest to a water molecule.
Similarly, for demonstration purposes we could say there’s no waves/crests when gravity is removed.
So when I hear emergent, I imagine properties that show up under certain conditions within a system, that were not present in individual components.
It doesn’t have to be integrated into the LLM at that point. If an Email has hidden text “do X”, which triggers the LLM to try to “do X”, but all post/push APIs have a user verification on them before they’re sent.
Sure it could get messy when the LLM tries to summarize the “why” on that action, but this is fairly similar to where we are now with phishing and uneducated individuals.
It’s also unlikely these LLMs have unbounded actions they could take. Specific ones like “send email to all recipients” could easily be classified as dangerous. You don’t even need an LLM to classify that.
I sometimes think we forget there’s glue between the LLM and the internet, and that glue can be useful for security purposes.
This has always made me curious about multi modality and especially Google’s Palm. At least how google presents Palm in diagrams.
The way I’ve interpreted the scaling hypothesis is that we will see emergent intelligence with larger nets through automated training alone. If we want a model to learn images we throw a larger net at it with training data.
The way I’ve interpreted some of these newer techniques to multi modality is that they are stitched together models. If we want a model to learn images we decide this and teach one model images and then connect it to the core model. There’s not a lot of emergent behavior due to scaling in this scenario.
With that perception I don’t see how gpt4 says anything about the scaling hypothesis. However I am not in this field and would be grateful to learn more.
I’m not sure if dreaming is the right word here, but the recent ML model Dreamer v3 [0] “dreams”. This is more akin to thinking ahead since it’s really just predicting outcomes using its world model.
It’s possible they could try training the model on some hallucinated scenarios to avoid over fitting, but I’m no ML researcher.
I’d be weary of the above comment. Searching for similar stories online has lots of spam posts saying similar things. Same bad grammar too. Usually they tell you to go to that website or one with almost the exact name that’s been shutdown.