New Research in Quantum Fault-tolerance

Dr. Quantum Computing has spent the last couple of years working for OpenAI, and returns to announce:

Let me end by sticking my neck out. If hardware progress continues at the rate we’ve seen for the past year or two, then I find it hard to understand why we won’t have useful fault-tolerant QCs within the next decade. (And now to retreat my neck a bit: the “if” clause in that sentence is important and non-removable!)

I think they created one logical qubit that lasts for a millisecond or so. Or something like that, I did not read the details.

I am just passing this prediction along. Seems doubtful to me.

The Invention of Large Language Models

Most people think that LLMs were invented by Google in 2016, or by OpenAI several years later.

AI expert Andrej Karpathy wrote on his blog:

The Unreasonable Effectiveness of Recurrent Neural Networks

May 21, 2015

The concept of attention is the most interesting recent architectural innovation in neural networks.

He constructs some LLMs, and his results seem pitiful compared to what is done today, but I would say he has a proof of concept.

Google introduced the transformer in Attention Is All You Need, a 2017 paper. As you can see, attention was already a hot idea at the time.

I am not sure who should get credit for inventing LLM. The basic ideas of neural nets go back decades. They got a whole lot smart when gaming GPU chips became fast and widely available, and AI researchers figured out how to use them efficiently.

No Quantum Nonlocal Effects

This is an elementary fact about quantum mechanics.

Dr. Bee explains:

Most importantly, if you do something to one of the pair of entangled particles, 4:46 that does nothing to the other. Say you turn the spin of the particle on the right upside-down 4:53 even though you don’t know what it is. Then the spin of the other particle doesn’t change at all. 4:59 No, it doesn’t. You wouldn’t believe how often I see even physicists get this wrong. I just 5:05 the other day heard a talk from someone who works on quantum computing for heaven’s sake 5:10 who said that if you manipulate one of a pair of entangled particles then that will non-locally 5:16 affect the other. It will not. It’s just that if you make a measurement on one of the particles, 5:23 then that will tell you something about the other. Because they’re correlated. ...

The person who 5:50 makes a measurement on one end can’t tell that a measurement was even done on the other end. 5:56 So: Entanglement is real, we know that. Whether spooky action is real is still a matter of debate, 6:04 but you can’t send information faster than light with either.

She is correct. No one has ever found any nonlocal quantum effects. Just correlations.

It should not be complicated. The same happens classically.

What she says about "spooky action" is a little confusing, so here is her explanation.

2:36 Strange or not, Bohr said that when we measure a particle, 2:40 this superposition “collapses” and suddenly the particle is in only one place. It’s this collapse 2:47 that Einstein referred to as spooky action. Because it would indeed be faster than light. 2:53 The moment you find the particle in one place, you instantaneously know it can’t be elsewhere. 3:00 Einstein disagreed with Bohr. Einstein thought that quantum particles are really only in one 3:06 place and that the sudden update of the wave-function just means that you 3:10 have learned the particle isn’t elsewhere. And his main argument, here it comes, was that by 3:17 claiming the collapse is a physical process, Bohr was introducing a “spooky action at a distance”. 3:26 Even if this spooky action existed though, it couldn’t transfer information. Just because you 3:32 find out what’s going on elsewhere doesn’t mean you sent information there.

When you find a classical particle, you immediately know it cannot be elsewhere. If that is spooky, then classical mechanics is spooky. Regardless, no information or anything else goes faster than light.

Usually she branches into a plug for superdeterminism. Mercifully, she did not this time.

Five Ways to Think About Quantum Supremacy

Aventine reports:

When Google announced that it had achieved quantum supremacy in 2019, the headlines were thrilling.

The world of quantum computing had taken a remarkable step. Google, with its Sycamore quantum processor, had performed a calculation in 200 seconds that, the company claimed in the journal Nature, would take a supercomputer 10,000 years.

This feat, named quantum supremacy by John Preskill, a theoretical physicist, back in 2012, promised to usher in a new world of computing performance. ...

Only it didn’t play out the way Google hoped or expected. ...

And then, earlier this summer, researchers from Shanghai Artificial Intelligence Laboratory in China completed the same task in just 14.22 seconds, driving a final stake through the heart of the Google quantum supremacy claim.

It’s not the only warning sign for the industry. Venture capital investment in the sector has fallen off a cliff, from $2.2 billion globally in 2022 to about $1.2 billion in 2023.

It then reports the opinions of five experts, but none of them say that quantum supremacy has been achieved. The closest is Scott Aaronson who says:

Quantum supremacy can be achieved and then unachieved later. It’s a little bit of a moving target in that sense. But all expect that we’ll eventually get to a place where quantum computers are just routinely doing things that classical computers cannot replicate within thousands of years or millions of years, and at that point there’s no more arguing about it.

Achieved and then unachieved? This is a bit like a mathematician saying something was proved, and then disproved. If it was later disproved, then it was never really proved.

Gil Kalai argues that quantum supremacy is impossible, and explains further here.