The short answer is that tracing is way, way easier to implement in a predictable and reliably performant way. This especially matters for distributed computation and automatic differentiation, two areas where JAX shines.
AST parsing via reflection means your ML compiler needs to re-implement all of Python, which is not a small language. This is a lot of work and hard to do well with abstractions that are not designed for those use-cases. (I believe Julia's whole language auto-diff systems struggle for essential the same reason.)
Glad to see that you can make ensemble forecasts of tropical cyclones! This absolutely essential for useful weather forecasts of uncertain events, and I am a little dissapointed by the frequent comparisons (not just you) of ML models to ECMWF's deterministic HRES model. HRES is more of a single realization of plausible weather, rather than an best estimate of "average" weather, so this is a bit of apples vs oranges.
One nit on your framing: NeuralGCM (https://www.nature.com/articles/s41586-024-07744-y), built by my team at Google, is currently at the top of the WeatherBench leaderboard and actually builds in lots of physics :).
We would love to metrics from your model in WeatherBench for comparison. When/if you have that, please do reach out.
Agree looking at ensembles is super essential in this context and this is what the end of our blogpost is meant to highlight. At the same time, a good control run is also a prerequisite for good ensembles.
Re NeuralGCM, indeed, our post should have said "*most* of these models". Definitely proves that combining ML and physics models can work really well. Thanks for your comments!
HN never disappoints, jeez. Thanks for chiming in with some expert context! I highly recommend any meteoronoobs like me to check out the pdf version of the linked paper, the diagrams are top notch — https://www.nature.com/articles/s41586-024-07744-y.pdf
Main takeaway, gives me some hope:
Our results provide strong evidence for the disputed hypothesis that learning to predict short-term weather is an effective way to tune parameterizations for climate. NeuralGCM models trained on 72-hour forecasts are capable of realistic multi-year simulation. When provided with historical SSTs, they capture essential atmospheric dynamics such as seasonal circulation, monsoons and tropical cyclones.
But I will admit, I clicked the link to answer a more cynical question: why is Google funding a presumably super-expensive team of engineers and meteorologists to work on this without a related product in sight? The answer is both fascinating and boring:
In recent years, computing has both expanded as a field and grown in its importance to society. Similarly, the research conducted at Google has broadened dramatically, becoming more important than ever to our mission. As such, our research philosophy has become more expansive than the hybrid approach to research we described in our CACM article six years ago and now incorporates a substantial amount of open-ended, long-term research driven more by scientific curiosity than current product needs.
From https://research.google/philosophy/. Talk about a cool job! I hope such programs rode the intimidation-layoff wave somewhat peacefully…
Google uses a lot of weather data in their products (search, Android, maps, assistant, probably others). If they license it (they previously used AccuWeather and Weather.com, IIRC), it presumably costs money. Now that they generate it in house, maybe it costs less money?
(Former Google employee, but I have no inside knowledge; this is just my speculation from public data.)
Owning your own data and serving systems can also make previously impossible features possible. When I was a Google intern in 2007 I attended a presentation by someone who had worked on Google's then-new in-house routing system for Google Maps (the system that generates directions between two locations). Before, they licensed a routing system from a third party, and it was expensive ($) and slow.
The in-house system was cheap enough to be almost free in comparison, and it produced results in tens of milliseconds instead of many hundreds or even thousands of milliseconds. That allowed Google to build the amazing-at-the-time "drag to change the route" feature that would live-update the route to pass through the point under your cursor. It ran a new routing query many times per second.
ERA5 covers 1940 to present. That's well before the satellite era (and the earlier data absolutely has more quality issues) but there's nothing from 170 years ago.
+1 one of most common mistakes for PhD students is picking a project based on research interests rather than the adviser. I believe it is relatively uncommon to speak with former students of an adviser, but this is something that everyone entering a PhD should do.
Finding a good mentor -- someone whose values you agree with and who sets you up for career success -- is far more important than working on any particular topic of interest. The world is full of interesting research topics, and very few PhDs work in the precise area of their PhD research for their entire career.
I work on weather prediction, both with traditional simulation methods and machine learning. I have not seen any examples of cellular automata used for useful predictions in this space.
As former quantum physicist, I find it little troubling to read "quantum theory has reached a dead end" in specific reference to the interpretation of quantum mechanics. Most quantum physicists could not care less about how quantum mechanics is interpreted when it makes highly accurate quantitative predictions, and there are still plenty of interesting open problems for quantum theory (e.g., related to the practical design of algorithms and hardware for quantum computers).
This article also misses what is likely the leading interpretation of quantum mechanics by actual quantum physicists, namely that the measurment problem is solved by decoherence (the quantitative theory of how classical states emerge from quantum states):
I read somewhere that quantum mechanics is the most tested of all scientific theories. And it has been shown to be right, every time.
Hawking espoused this idea he called “model dependent realism”. The idea is that every human understanding of reality is model-dependent, that is, it is not “reality” that we truly understand (we can’t) but rather in every case we have some model of reality that is useful in particular situations. For instance, we know that Newtonian physics are not “real” but they are perfectly accurate in certain situations. So they are not “wrong” when they are used in those situations, in fact, they are right.
The author of the article writes, “While Einstein won a Nobel Prize for proving that light is composed of particles that we call photons, Schrödinger’s equation characterizes light and indeed everything else as wave-like radiation. Can light and matter be both particle and wave? Or neither? We don’t know.”
In model dependent realism, we can ignore this apparent contradiction. In some situations the model of light as a particle is the most useful, and in others, the model where it is a wave is the most useful. We have to accept that it is not “really” either of these models, but that no matter what we do, any model we come up with for it will still just be a model.
>> Can light and matter be both particle and wave? Or neither? We don’t know.
But we know! The answer is neither.
Light and matter are weird things that is impossible to describe with usual language, but they can be described very precisely with math language. The problem is that the equations are too complicated and difficult to use.
They have been tested thoroughly, for example in particle accelerators but in experiments with very few things moving around. It's very difficult to use them when the experiment gets bigger.
In some cases, you can make some approximations and get almost the same result if instead of the full correct equations you use the wave equation. It's just an approximation. Light and matter are never waves, but in some case they can be approximated as waves.
In other cases, you can make some approximations and get almost the same result if instead of the full correct equations you use the particle equation. It's just an approximation. Light and matter are never particles, but in some case they can be approximated as particles.
And in other weird cases, bot approximations get very inaccurate predictions.
Apparently Hawking coined this term in 2010. Robert Anton Wilson coined a similar term, "Model Agnosticism", all the way back in 1977 in his book Cosmic Trigger:
"The Copenhagen Interpretation is sometimes called "model agnosticism" and holds that any grid we use to organize our experience of the world is a model of the world and should not be confused with the world itself. Alfred Korzybski, the semanticist, tried to popularize this outside physics with the slogan, "The map is not the territory." Alan Watts, a talented exegete of Oriental philosophy, restated it more vividly as "The menu is not the meal."
"All models are wrong, some are useful" describes it very well for me. It still amazes me, how late I really understood this and how many intelligent people not fully understand it.
he expanded on that quite a bit in quantum psychology (1990), giving an expanded treatise on the software of the mind and how it maps our interpretation of reality.
quite a good read if you liked his previous works.
We are human and prefer elegance, so enough will continue to try to unify models anyway. I don’t buy that we cannot understand nature but this model dependent realism is fine as a practical way of working until we do understand it.
I don’t think anyone is saying that, but in pondering this issue, I remembered how in Spanish there are two verbs for “to know”, “saber” and “conocer”. That latter verb is often explained in English as “to be familiar with”. The usage makes the point best: you can “conocer” a person but cannot “saber” them. That is, you can be acquainted with someone but you cannot truly “know” them, no matter how close you are to them.
Think about it: how well do you know yourself? You live in your own head and yet you are probably surprised by some of your own reactions, or dismayed by your actions, or fearful of certain emotions. If you do not fully understand yourself then what does understanding nature even mean? I cannot inhabit the mind of my wife, let alone inhabit a photon.
They aren't "utterly incompatible", they're largely compatible. For example, lasers work here on Earth, and between the Earth and its moon. Moreover, hydrogen maser clocks and cesium et al.'s hyperfine transitions are used in clocks which are sensitive to nearby mass concentrations, and altitude above the Earth.
There are whole textbooks written on the limit in which General Relativity and Quantum Mechanics work well together, with Birrell & Davies 1984 https://books.google.co.uk/books?id=SEnaUnrqzrUC being the most widely used by graduate students (and as a reference book for researchers).
Indeed, such textbooks go into where GR & QM make incompatible predictions, and almost all of those are in the limit of strong gravity, which in turn is almost certainly always deep within an event horizon, or isolated in the very very very early universe.
Semi-classical gravity (SCG) works well as an "effective field theory", and simply marries a classical curved spacetime (General-Relativity style) with a relativistic quantum field theory (standard-model-of-particle-physics style). In particular, with minor caveats, on the cusp of strong gravity SCG is successful enough in the astrophysical study of stellar remnants that it is reasonably believed to be good everywhere outside black hole horizons and after the very early universe. https://en.wikipedia.org/wiki/Semiclassical_gravity -- one of the caveats is noted there, namely given a sizeable mass (> kilograms) brought into a superposition of space, it is not clear at all what SCG predicts a cavendish apparatus or other gravimeter will point to. This is a possible incompatibility of SCG's two more-fundamental theories in the weak gravitational field, low-energy matter, and low-speeds-compared-to-c limit, and is a puzzle that hopefully will be informed by clear experimental data some day.
Since we can't get information back from inside a black hole horizon; can't see anything in the very very early universe (electromagnetism hadn't "frozen out" of the GUT yet for instance); direct detectors of very early universe gravitational radiation are implausibly hard engineering tasks; and a bowling ball sized mass will be extremely hard to keep in a coherent state for reasonably long periods of time; these are really academic problems rather than practical ones.
As a former quantum physicist who has just decided to go back into quantum computing, this was my take as well: Introductory quantum physics courses may still include wave-function collapse and all that nonsense, but I have not met many physicists who use this as a mental model.
To be a bit more specific as to how _decoherence_ solves this, one way to see it is that classicality (i.e. observables having specific values) is an emergent property in the limit of near-infinite degrees of freedom in the same way that e.g. thermodynamic properties (temperature etc.) are emergent properties of classical systems in the limit of near-infinite degrees of freedom.
Putting it on the edge, claiming that quantum theory is at a dead end is like claiming statistical physics is at a dead end.
One of my personal favorites for how to formalize this is the work on "pointer states" by Wojciech H. Zurek. There is a freely available Physics Today articls [0], and you can find surveys of further work e.g. in the introduction of [1].
The collapse just stands for the unknowable details of the interaction with the environment during measurement, a good quantum physics course will explain that and include experiments that make that clear. For instance the Stern-Gerlach experiment illustrates this well.
Decoherence is a consequence of interaction between particles in the coherent state, environment plays no role there, you can do all experiments in the vacuum and they will still work the same way, obviously.
> namely that the measurment problem is solved by decoherence
I think "solved" is too strong. The Wallace paper you reference, for example, does not claim that decoherence solves the measurement problem. His claim is only the more modest one that understanding decoherence helps to clarify what the measurement problem actually is.
I'd love if decoherence was now the dominant perspective, and I at least am largely convinced barring some huge revolution, but it would surprise me very much. In terms of my acquaintances (largely the department of my university) team "I don't care, does it matter?" takes top spot followed by equal parts collapse and decoherence. Oh, and some Bohmian people but I don't know them as well, as interesting as it would be if they were right.
Decoherence theory is really somewhat orthogonal to the measurement problem. Decoherence explains how the "collapse" happens gradually if you have a non-perfectly isolated system by seeping entanglement. But at some point you still need to activate the stipulation that you as an observer is drawn into the entanglement and at that point whatever's left of the wavefunction "collapses".
In essence the Copenhagen interpretation is still correct as a simplification that can still be OK in most cases. This is reflected by the fact that practising solid state physicists have successfully used this 20's style of QM for 100 years now.
I can't speak for the parent, but I have a physics degree from 30 years ago. Since then my career has diverged from academic physics to the point where it would take some effort to re-learn even my graduate level QM coursework, much less familiarize with current research topics. So I could see where the "former" status comes from.
I guess "quantum physicist" can mean either an academic credential or a job title. In my case, I have the former but no longer the later. I finished my PhD nine years ago and no longer work in the field.
My experience was that robo-investors are great until you need something special. Then they can become rather painful.
Exmaple: I got divorced last year. Betterment took weeks of time and many phones calls until they were able to figure out a way to divide our assets evenly, without a large difference in cost basis. Their automatic algorithm for dividing accounts just didn't know how to handle it.
If UBS figures out how to offer a higher level of service on top of robo-advising, that could be a real win.
Anyone who believes anything said by a company from that region with respect to investments in the western world was always in for a surprise. It’s best to assume they think we’re all fat, dumb, lazy, and easily duped, and assume they’ll act accordingly.
Exactly. I'm less charitable because of Foxconn and how that story developed. Big companies promises tons of money and small/mid sized economies will give them huge tax cuts in exchange. In reality they build a much smaller plant and still take the tax cuts.
I don't think you can generalize between these deals when:
1. The semiconductor companies on one side of the table are very different: Taiwanese company mostly employing in China versus a US company. National cultures around commitments and deals vary.
2. The politicians on the other side of the table are very different. Scott Walker and Donald Trump with the Wisconsin deal and Mike DeWine in Ohio.
AST parsing via reflection means your ML compiler needs to re-implement all of Python, which is not a small language. This is a lot of work and hard to do well with abstractions that are not designed for those use-cases. (I believe Julia's whole language auto-diff systems struggle for essential the same reason.)