The answers can be recorded and reviewed. The other points are true, or is there a way to make outcomes deterministic, when compared to previous versions while allowing to add more knowledge in newer versions?

It's possible to make any model deterministic. Used to be just to save the seed, but I'm not sure it still is now that everything is distributed. Maybe a little more effort.

determinism isn’t really enough, we want “predictable”. Most of these AI wavefunctions are “chaotic” - tiny changes in state can cause wildly divergent outcomes

A part of my question that you didn't go into was, can new knowledge be added in a new version without making the answers with knowledge learned in previous versions non-deterministic?

that’s not really how training works.

changing the input (data) means you get a different output (model).

source data has nothing to do with model determinism.

as an end-user of AI products, your perspective might be that the models are non-deterministic, but really it’s just different models returning different results … because they are different models.

“end-user non-determinism” is only really solved by repeatedly using the same version of a trained model (like a normal software dependency), potentially needing a bunch of work to upgrade the (model) dependency version later on.

This requires an exact lock-down of things like the hardware and driver version, doesn't it? Is that sustainable?

It shouldn't. It didn't used to, at least.

But that won't survive an upgrade, will it?

It's also not possible to tap into the full amount of reduction because of the land use

The figure they measured, in terms of global acreage, works out to an increase in forestry across 8% of the Earth's land area. That's completely viable, in fact some multiple of that is as well. The overwhelming majority of land on this planet is doing pretty much nothing owing to our habit of packing ourselves into tiny densely packed cities which are supported by a fairly small amount of agriculture - 39% of all land there to be exact. [1]

So this is completely viable, and then some! Even more so if one accounts for the fact that increases in CO2 are creating a substantially greener planet making this even easier.

[1] - https://data.worldbank.org/indicator/AG.LND.AGRI.ZS

The nature of "being unsure" often inherently precludes being aware of that unsureness. Some companies are well known to exploit this basic fact of existence.

Let's not pretend that context doesn't matter, including the average technical proficiency of the intended audience, or the fact that HN's conventions are recognizable to people who have grown up with Usenet, BBS, forums, aggregators, etc. That is not the case here.

You don't need context there is terms and services that no one reads

This argument kind of defeats itself.

Even if you can not compute the result in a finite number of steps through the naive approach there could still be a better approach, let me call it shortcut, for determining the value that can be computed. E.g. the geometric series result is known (a/(1-r)) but not computable through evaluating the series itself. A problem is undecidable if you can prove that the shortcut can not exist, it is decidable if you know a shortcut exists (knowing the formula is not required), and potentially either decidable or undecidable if a proof in either direction is unknown.

While it technically correct to say this it also gets the wrong point across because it leaves out the fact that ordering changes create only a small difference. Other examples where arithmetic is not commutative, e.g. matrix multiplication , can create much larger differences.

> ordering changes create only a small difference.

That can’t be assumed.

You can easily fall into a situation like:

  total = large_float_value
  for _ in range(1_000_000_000):
    total += .01
  assert total == large_float_value

Without knowing the specific situation, it’s impossible to say whether that’s a tolerably small difference.

Obviously this is not how video compression and packets work but for the sake of the argument consider the following. The article speaks of a 300 fiber cable. A one bit per pixel square image with approx. 300 pixels is 17x17 in size. Not your typical video resolution.

Not your typical frame rate either.

Optics also have signal integrity issues. In practice OSNR and SNR limit optics. Cutting the fiber still breaks it. Small vibrations also affect the signal's phase.

Phase variations will not introduce any issues here, they most certainly are talking about intensity modulation. You can't really (easily) do coherent modulation using incoherent light sources like leds.

SNR is obviously an issue for any communication system, however fiber attenuation is orders of magnitude lower than coax.

The bigger issues in this case would be mode-dispersion, considering that they are going through "imaging" fibres, i.e. different spatial components of the light walking off to each other causing temporal spread of the pulses until they overlap and you can't distinguish 1's and 0's.

Mode dispersion is frequency dependent phase changes.

That's chromatic dispersion, mode dispersion is spatial "path" dependent phase changes. Vibration is actually somewhat more relevant because if it wasn't for that we could theoretically undo mode dispersion (we would need phase information though).

That said all of that is irrelevant to what the previous speaker said, vibration induced phase variation as an impairment. Thats just not an issue, vibrations are way too slow to impair optical comms signals.

How do the gravity wave optical paths solve the vibration issues? Couldn't TSMC do something similar?

aha! that is true with lasers that are coherent. But not with LEDs. We don't care about modes, polarization, or phase. Also, no worry about feedback into the lasers, so no isolators. LEDs are way easier!

No, not necessarily. If you can distinguish different amplitude levels you can do better. For example four amplitude modulation (4AM) carries two bits per symbol. There is also the option to use coherent optics, which can detect phase, and carry additional information in the phase.

Algorithms can see the difference between RGB 245, 245, 245 and RGB 246, 245, 245 (it's 1, 0, 0) but the eye can probably not, also depending on the monitor hardware. Thus the blurring effect might not be as strong as it looks like at first glance.

Maybe they also provide an AI hotline they have to sell to you?