It is a very fad driven field. Everyone brands everything. It isn't enough to give things boring titles like, stacked open linear dynamical system with selective observations and learned timestep.
There is nothing theoretical that stops anyone from trying that. It is just a matter of engineering and architecture. It is (in principle) possible to define "spiking transformers" and there is more than one way to do so, but this will not mean necessarily that you will immediately get good performance. Also current generation hardware doesn't as efficiently simulate SNN, so this is another constraint.
Trying to figure that out right now :), there is a tension between an algorithm like this and the biophysical reality of actual brains. My intuition is that whatever the brain does is not so much an approximation of any known algorithm, but rather an analog of such an algorithm. Now that we know how in a spiking neural network gradient computation looks like, we can ask what assumptions are violated in the brain, just like people have done before for the vanilla backpropagation algorithm. There is some recent concrete experimental evidence that points towards a solution.
That you can't backpropagate is a common misconception. There is recent work (I am one of the co-authors), that derives a precise analog of the backpropagation algorithm in spiking neural networks (https://arxiv.org/abs/2009.08378). It computes exact gradients and requires only communication at spike times during the backward pass. The reason this works is because "x is not differentiable" lacks a statement "with respect to y". It turns out that since gradient computation is local, a gradient is defined almost everywhere. The gradient computation is only ill-defined at places where a spike would get added or deleted. This is similar to how in a ReLU network an exact zero input is "non-differentiable".
Have looked at the T1 Code? It far from obvious how you would use that as a starting point for a new chip. You need a lot of expertise for custom memory blocks etc. and there aren’t any tests included if I remember correctly.
Julia is a language that was developed by academics, with not the best aesthetics in the implementation. I see a lot of value in starting with a clean, well engineered layered approach. If anything this is something I would expect Lattner to be able to deliver, whereas Julia while certainly powerful is tremendously messy.
They used to do nuclear physics research (and still do to some extend) there and had an experimental fast breeder reactor (capable of producing weapons grade plutonium) on campus. They also prepared for rapid development of nuclear weapons capabilities in the 60s. It was a site for a potential German nuclear weapons program, I think they would have been able to produce enough material in ~6 weeks. They frame it as "nuclear disarmament" now of course https://www.fz-juelich.de/en/news/archive/press-release/2022..., but effectively if there is any place where scientists in Germany have active knowledge how to develop nuclear weapons it would be there. There are several former military installations close by, including barracks for a guards and supply company.
You don't want something like that in a city centre.
search and rescue is also code for the military might find it useful. whenever a drone is supposed to deliver aid packages to firefighters for example.
The hardware they have in production is underpowered and wouldn't be able to render those high fidelity avatars. That is also why their version of the meta verse looks like it does. They want to be independent of the "legacy" platforms, so their device has to do all of the compute while also being affordable for wide adoption.
Ironically the very first publication on "Backpropagation" by a Finish master student (1972) already pointed out how to compute gradients in computational graphs. This if anything is the main thing which distinguishes Backpropagation as an algorithm from simply applying the chain rule.
It is the same in many areas I think. You get in if you know someone, or have a prior established rapport. At the very least that is a way more pleasant way of finding a new job.
I've tested it on a various areas of expertise in physics that I am familiar with and it often makes up content which sounds very plausible except that it gets pretty much always details wrong in very important ways. On the other hand I've found it very useful in providing reference articles.
Well in both cases you can ignore a lot of unknown physical details and still reach a valid conclusion. In the case of the climate this is done in terms of energy balance for example. The details are super complicated but you also have historical data that you can use to calibrate and validate your models on.
Measurements are adjoint to state preparation. Depending on what you measured you can prepare a special state at that time to make the system time reversal symmetric.
Volkswagen planned to hire 10000 new people for their "Car.Software" division in Wolfsburg, Ingolstadt, Berlin, as well as Seattle und Peking until 2025. They are also building up their hardware division with plans to do cutting edge node designs in 5nm and beyond. I know that they have extensive recruiting efforts in both areas.
Have you lived in Germany? I would only consider going to the US for substantially higher pay. Also there are 80 million Germans, many of whom are fairly well educated and would like to work for Mercedes in Stuttgart.
Both Mercedes and Volkswagen have absolutely massive software divisions, easily 10k+ people. They pay really well for German standards and have access to a large talent pool of people that would like to live in Germany.
While in principle you can evaluate certain path integrals by using MCMC, in practice you run into all kinds of issues like the sign problem, which even with things like complex Langevin dynamics can't be resolved to this day completely.
Even optical computing needs lots of standard electronics for storage etc. as you point out these consume most of the area, so we are talking about something that consumes 100x more area for a function which is not the main energy expensive thing in traditional computers anyways, memory and communication make up at least 1/3 of the energy budget, which doesn't go away by making the rest optical.
In fact reprogramming the optical non-linearities typically is much slower and energy intensive than retrieving and flipping some bits. Which makes non-static non time-multiplexed computation extremely slow compared to whatever the "best" case static scenario is.
Visible light and near ultra-violet light has a wavelength ~400-800 nm, current gen transistors have a pitch of ~40 nm. This gets worse because scaling is actually either quadratic (2d) or cubic (future 3d integration). So we are talking about 100x to 1000x worse spatial scaling disadvantage at the moment. The only redeeming quality of light is wavelength multiplexing, but that is only useful for a subset of applications, like optical communication and (maybe) convolutions (see below).
Moreover even in a hypothetical scenario where we somehow were able to find materials applicable to smaller wavelength, the deBroglie wavelength of an electron is ~1000x smaller, than that of a Photon at the same energy. So in terms of integration density electrons will always have a 10^6 - 10^9 (2d - 3d) theoretical advantage over photons, which means that investment in electron based computation will have a much more likely eventual payoff.
Take for example https://www.nature.com/articles/s41586-020-03070-1, they have a bunch of projections for what they hope to achieve over time. The most fantastical figure they give is 50 Peta MAC / s, but this doesn't take into account the PCM programming time.
If you take a look at the supplementary material https://static-content.springer.com/esm/art%3A10.1038%2Fs415... it becomes clear that they currently have a much lower TOPS/Watt figure than current generation ML ASIC like the TPU and this neglects all the expensive experimental optical equipment they would need to miniaturise. So even in their most favourable comparison they are 5x worse. Most of these papers unfortunately are full of hype and claims like that.
More power to them if it works, you can do cool things with it for sure. I am continuously amazed what you can do in quantum optics etc. but it isn't anywhere near to practical except for optical communication.
One of the things that I fully not expect to be successful is optical computing. There are just a lot of academic groups that are doing optics and they like to invent new reasons why whatever they are up to is relevant. For physics reasons the integration density of optical compute elements is abysmal and will remain so forever. Other technologies like spintronics at least have the chance to work sometime in the future. There were projects on wafer scale optical computing already in the 80-90s at MIT Lincoln labs, so this isn't exactly a new idea either. We have a new group at our institute doing "Neuromorphic Quantum Photonics", they publish in high-impact glossy journals, doesn't change that it is in my opinion mostly hype and bullshit.
It’s all hypothetical at this point, I reluctantly followed the guidance, more or less because I didn’t want to be inconvenienced too much. You are right I could have suffered worse, but then again I have absolutely no risk factors and statistically my chance of death or serious outcome was very close to zero regardless of vaccine status.
There was a clear indication from the data in Israel that nothing was working as it was advertised in Europe. Now high profile publications are coming out that confirm some of the worst fears, so who knows maybe at some point these people will have to take some accountability.