Item — HN Reader

Looking at the paper, cherry picking 5 out of 21 devices is in itself not a deal breaker IMO, but it's certainly something they should have disclosed. I bet this happens all the time with these kinds of exotic devices that take almost a year to manufacture only for a single misplaced atom to ruin the whole measurement.

Average of positive and negative Vbias data and many other manipulations are hard to justify, this reeks of "desperate PhD needed to publish at all costs". Yet at the same time I wouldn't fully disqualify the findings, but make the conclusion a lot weaker "there might be something here".

All in all, it's in Microsoft's interests that the data is not cooked. They can only ride on vaporware for so long. Sooner or later the truth will come out; and if Microsoft is burning a lot of cash to lie to everyone, the only loser will be Microsoft.

3 comments

> cherry picking 5 out of 21 devices is in itself not a deal breaker IMO

Might as well draw a straight line through a cloud of data points that look like a dog

It's a physical device at the bleeding edge of capabilities. Defects are pretty much a guarantee, and getting a working sample is a numbers game. Is it really that strange to not get a 100% yield?

Having 5 working devices out of 21 is normal. The problem is that the other 16 weren't mentioned.

Well you also need to account for what kind of deviation are we talking about between the 21. If they selected the 5 because they were the best, but the others showed results that were within say 0-5% of the 5, then sure that is acceptable. But if we’re talking about flipping a coin 21 times, seeing heads 16 times and then choosing the 5 tails outcomes as the results, then I would say that’s pretty unacceptable.

I do not think this is the right metaphor. Having 5 devices work out of 21 is actually a better yield than what TSMC would get with a brand new process. This is not just normal, this is expected. It is all the other allegations that make this be a very questionable case.

Like I said, a single misplaced atom is enough to wreak havoc in the behaviour of these things. That's not the problem, everyone knows there's a large gap between phenomena observed, and making it consistently manufacturable with high yield.

It wasn't just that by itself. There was a list of several undisclosed data tweaks and manipulations. None were particularly fraudulent or anything, but once you have them all included in the paper, as the former author was complaining, it seems more likely that they just manipulated the theory and data as needed to make them match. There's a big difference between predicting something and demonstrating it in experiment, versus showing your theory can be made to fit some data you have been given when you can pick the right adjustments and subset of data.

> the only loser will be Microsoft.

Not really - that cash could have been allocated to more productive work and more honest people.