China develops photonic AI chip | Hacker News Reader

China develops photonic AI chip

science.org

21 pointsBeijinger19 comments

@dang can we improve the title?

Starting with "China..." gives IMO an unnecessary political framing to this story. These were researchers at Chinese universities (almost all the same, Tsinghua), so it should say so. There are some cases where we say "America .." or "Russia.." but usually in the context of a body directly under gov't control such as military or federal policy. We never say "Europe develops a sailboat" or "India develops new diabetes medication", it would be "French startup develops" or "India's Biocon develops". Sure the government may have stronger hand in affairs than in other countries, but there's no reason to invoke a different framing because of a political difference.

"Tsinghua University Researchers Develop Photonic AI Chip" feels more appropriate.

Well, with the current chip disputes (no last generation lithography for china, China throws AMD and Intel out from telecommunications) China feels quite appropriate for the headline. I know out of my head that Tsinghua U. is in China, Beijing. But I doubt that many readers know this.

I doubt many readers would know it either, but everyone has to start learning somewhere, however I'm certain almost every visitor would be able to tell it's one of two mandarin-speaking places. Regardless of one's opinions on China, it is becoming more important and wealthy and it cannot hurt to learn the names of important institutions slowly.

Yes, Tsinghua is the Harvard of China. It would be weird if Stanford research came up with a breakthrough but the headline merely stated "US". Let's begin to treat Chinese universities as a non-monolith.

*MIT PKU is the Chinese Harvard

I used to know someone (about 8 years ago) who studied photonic chips at Stanford. She didn't seem to be confident about this direction. one challenge she mentioned was it's difficult to achieve non-linearity with photonic chips alone, like the activations. hence they had to read back signals and use electronics to add non-linearity, which in turn diminished the benefit of using photonic chips.

I wonder if things have changed a lot since then.

That's astonishing I thought light has the facility to (self) interact in wonderful non linear ways.

Photonic AI, is this how we get 'Raised by Wolves'.

I really want to see optical processing become mainstream as quickly as possible. Not just for AI, but in mainstream PCs as well.

Then we need to help train more physicists and material scientists. However, because of the advertising economic boom and high salaries of computer science, a vast majority of the next gen are studying cs instead.

> To support modern AGIs, we designed Taichi—large-scale photonic chiplets

“Modern AGIs” because we had classic AGIs? Last I checked AGI remains in the realm of science fiction and prank tweets. I hope this is a translation snafu as it reads like clickbait.

Plenty of people have been using the term AGI for the general purpose nature of what the SOTA LLMs can do. And, I think it's reasonable, because they are very general when you compare them to the previous generation of narrow AI models.

We don't have another good term. Unfortunately, "AGI" is also overloaded to mean not just general purpose but also alive, embodied, complete emulation of human, sentient, godlike, etc. And everyone has a slightly different definition of which humanlike or godlike characteristics of people or animals something should have to be qualified as AGI.

I think this is a failure of language or language use as much as anything.

But prior to having something like GPT-4 or Claude-3 with vision you had to train a different model or use a different library for every task. Now you can use the same model for many very different tasks.

It's clearly got the "general" part, even though it doesn't have _all_ of the cognitive abilities of humans or even animals.

Science is a pretty solid magazine.

We don't even have consensus on the existence of general intelligence, artificial or otherwise, so I think the linguistic specificity bar is set rather low here.

Term 'AGI' is used for what we had 'AI' before. Now 'ASI' is used instead of 'AGI'.

We will run out of 'A[A-Z]I' before there will be actual AI.

> We will run out of A[A-Z]I before...

No problem - we'll just ask AI to find a new solution!

Honestly if the word AGI experiences some term inflation that wouldn’t be the worst thing.

It would be a language genocide

faux pas from a time traveler

Reply on news.ycombinator.com