julianjm

I think all researchers and science communicators have a duty to present science in a way which educates and edifies, and doesn't mislead. It's not just that they're successful, but that their publicity gives them a prominent role as science communicators. Science is all about and questioning your assumptions, and acknowledging limitations. They claim the public interest in their charter. I think it's reasonable to demand integrity from them, at least as much as it is from any other researcher, if not more. And I think OpenAI would agree with me on that point.

For a specific example of how I think their framing is unhelpful: in the LAMBADA evaluation (sec. 3.1), they suggest that one-shot performance is low "perhaps...because all models still require several examples to recognize the pattern." This may be the first thing you'd think of for a few-shot learner, but then why is zero-shot performance higher than one-shot? If you remember that you're working with a language model, there's another possible explanation: the model probably models the last paragraph is a narrative continuation of the previous ones, and gets confused by the incongruity or distractors. (The biggest model is able to catch on to the incongruity, but only when it's seen it before, i.e., with >1 example.) Of course, this is just one possible explanation, and it's arguable, but the point is I think it's more useful to think of this as a language model being used for few-shot learning, not a few-shot learner where language modeling is an implementation detail.

I am not a fan of this trend of "Language Models Are X" in recent work particularly out of OpenAI. I think it's a rhetorical sleight of hand which hurts the discourse.

Like, the exact same paper could have instead been titled "Few-Shot Learning with a Large-Scale Language Model" or similar. But instead there seems to be this extremely strong desire to see certain ineffable qualities in neural networks. Like, it's a language model. It does language modeling. Turns out you can use it for few-shot learning and do amazingly well. Beyond that, what does it mean to say it "is" a few-shot learner?

On one hand, it's literally the same claim in a strict sense. On the other hand, it implies something much broader and more sweeping, that language modeling / unsupervised learning as a task over long contexts inherently implies meta-learning ability — which is a statement that is very difficult to properly formulate, let alone back up. But that's the argument that I feel is being slipped under the table by these titles. (And indeed it's very close to what they suggest in the text, though with no more than a wave of the hands.)

Don't get me wrong: their intuition is reasonable, it's super cool that they got this to work, and the results are very impressive on lots of tasks (though there are clear gaps). But as a VERY publicly watched lab, they have a serious duty (which I think they're neglecting) to frame their results more carefully. In particular, there's a sort of religion that if you train a big enough model on big enough data with self-supervision, it will somehow become AGI and/or learn to solve arbitrary problems. Claims like "Language Models are Few-Shot Learners" are clearly designed to fit into that worldview, even though the research doesn't point at it any more than a more conservative interpretation like "Lots of NLP Tasks are Learned in the Course of Language Modeling and can be Queried by Example." They touch on this limitation in their discussion section but I guess flashy titles are more important. I wish they would use their status to set a better example.