The art of fact-checking

[ny_quant]

CLAIM: Greta Thunberg recently deleted a tweet she posted in 2018 in which she said that the world will end in 2023.

AP’S ASSESSMENT: Missing context. While Thunberg did delete a 2018 tweet about the urgency of addressing climate change, she did not say the world would end in 2023. Her tweet included a quote from an article that said an influential scientist warned climate change “will wipe out all humanity” unless fossil fuel use was ended “over the next five years.” Further complicating the issue, that article incorrectly summarized the scientist’s speech. He never made such comments.

...Thunberg never said the world was set to end in 2023. The young activist was quoting an article that was paraphrasing a speech by a Harvard University professor of atmospheric chemistry. The scientist said the world had limited time to act to reverse the disappearance of floating ice volume in the Arctic or there would be drastic consequences, not that the world would end in five years.

The article Thunberg was quoting has since been deleted, but it was published by the site Grit Post in February 2018 with the headline: “Top Climate Scientist: Humans Will Go Extinct if We Don’t Fix Climate Change by 2023,” according to an archived version of the web page. The first line of the story stated: “A top climate scientist is warning that climate change will wipe out all of humanity unless we stop using fossil fuels over the next five years.”

Forbes quoted Anderson as saying: “The chance that there will be any permanent ice left in the Arctic after 2022 is essentially zero.”

У снопса есть скриншот того твита и сходный анализ.

Comments

[anhinga-anhinga.livejournal.com]

But also the current models have those wonderful properties (like attention-layers computing some large step gradients, which greatly helps the few-shot learning).

So, presumably, better models need to make sure to provide an equivalent of that too... And there might be other properties of today's LLMs we would want to capture (or further upgrade) in better models.

[misha-b.livejournal.com]

Well, I am not a believer in neural architecture search. It seems too much like graduate student descent to me. We have to go back to first principles informed by the successes of neural networks, at least that is what I have been trying to do.

I agree, attention seems really important. Backpropagation not so much, perhaps.

[anhinga-anhinga.livejournal.com]

The class of machines I have been playing with is, essentially, a class of flexible attention machines (in the sense that a linear combination of vectors is actually the quintessential form of attention, so it is indeed fundamental).

But the magic property of **attention layers** performing gradient steps on the fly (without backpropagation) is something else; it does seem to be responsible for many of the magic properties of the modern models. And there are several papers approaching this phenomenon from different angles, but it still remains rather mysterious (the few-shot learning and all the GPT-3 magic would be quite unlikely without it).

[misha-b.livejournal.com]

I am reasonably confident now that we understand fully connected networks. They also have a form of attention, in fact (Expected Gradient Outer Product, essentially, supervised PCA).

I suspect something similar happens in Transformer architectures but it is speculation at this point.

[anhinga-anhinga.livejournal.com]

Fully connected feed-forward multi-layer ones, or fully-connected recurrent single-layer ones?

(I do think it makes sense to replace scalar flows by vector flows on the level of individual neurons; this change immediately turns any classical neural net into an attention machine.)

[misha-b.livejournal.com]

I mean fully connected multi-layer networks. Not sure about recurrent ones.

[anhinga-anhinga.livejournal.com]

> I am reasonably confident now that we understand fully connected networks. They also have a form of attention, in fact (Expected Gradient Outer Product, essentially, supervised PCA).

Are you publishing something about this form of attention?

[misha-b.livejournal.com]

I sent a link by a deleted comment.

[anhinga-anhinga.livejournal.com]

Спасибо, очень любопытно.

[misha-b.livejournal.com]

Btw, it also explains grokking. I remember you asked me a couple of years ago about it, but I don't think my answer was correct.

[anhinga-anhinga.livejournal.com]

А, да, все основные загадки того, почему и как происходит Grokking, решил Neel Nanda в августе. Замечательная работа, с подробностями и деталями.

Я сделал небольшие заметки на эту тему, и в этих заметках есть ссылки на подробности:

https://dmm.dreamwidth.org/64571.html

https://github.com/anhinga/2022-notes/tree/main/Grokking-is-solved

[misha-b.livejournal.com]

Thanks, interesting work. From a quick look, I think the main idea is right. However the general mechanism was not clarified. EGOP makes clear what feature learning means in general.