isotypic

I don't follow this argument, and there would still be issues with the computation anyways.

1) Pretend I want something written, and I want to minimize emissions. I can ask my AI or a freelancer. The total CO2 emissions of the entire industrial sector has nearly no relation to the emissions increase by asking the freelancer or not. Ergo, I should not count it against the freelancer in my decision making.

2) In the above scenario, there is always a person involved - me. In general, an AI producing writing must be producing it for someone, else it truly is a complete waste of energy. Why do the emissions from a person passively existing count when they are doing the writing, but not when querying?

3) If you do think this should be counted anyways, we are then missing emissions for the AI as the paper neglects to account for the emissions of the entire semiconductor industry/technology sector supporting these AI tools; it only computes training and inference emissions. The production of the GPUs I run my AI on are certainly an unalieanable part of having an AI do some job.

The way this paper computes the emissions of a human seems very suspect.

> For instance, the emission footprint of a US resident is approximately 15 metric tons CO2e per year [22], which translates to roughly 1.7 kg CO2e per hour. Assuming that a person’s emissions while writing are consistent with their overall annual impact, we estimate that the carbon footprint for a US resident producing a page of text (250 words) is approximately 1400 g CO2e.

Averaging this makes no sense. I would imagine driving a car is going to cause more emissions than typing on a laptop. And if we are comparing "emissions from AI writing text" to "emissions from humans writing text" we cannot be mixing the the latter with a much more emissions causing activity and still have a fair comparison.

But that's besides the point, since it seems that the number being used by the authors isn't even personal emissions -- looking at the source [22], the 15 metric tons CO2e per year is labeled as "Per capita CO₂ emissions; Carbon dioxide (CO₂) emissions from fossil fuels and industry. Land-use change is not included."

This isn't personal emissions! This is emissions from the entire industrial sector of the USA divided by population. No wonder why AI is supposedly "100-1000x" more efficient. Counting this against the human makes no sense since these emissions are completely unrelated to the writing task the person is doing, its simply the fact they are a person living in the world.

What exactly do you mean by non-trivial/modern standards? While certainly the largest and most complicated theorems are currently out of reach of proof verifiers, there isn't a shortage of usage of them to prove important/modern theorems (of course certain fields are much more developed/more amenable to verification than others).

* https://xenaproject.wordpress.com/2024/01/20/lean-in-2024/ discusses a few recent usages in recent papers and the author's grant to formalize Fermat's last theorem.

* The liquid tensor experiment (https://www.quantamagazine.org/lean-computer-program-confirm...)

* Feit-Thompson was formalized in 2012.

I do largely agree that formal correctness within mathematics is not as important as it may seem, though this doesn't mean formal verification of a proof is completely orthogonal to understanding it - you can't formally verify something without really understanding the proof in the first place.

Handling two different writes to memory is not really a concern - existing speculative/out of order processors already solve this issue by completing (perform architectural side effects) instructions in-order. So even if two writes are made, one in each branch, by the time the write is meant to be completed, the prior branch is resolved and we know which write is actually meant to be made and the bad one can be discarded.

Doubling the execution units also isn't strictly needed - you can use the existing out-of-order core to send two sets of instructions through the same functional units. There will be more contention for the resources, possibly causing stalls, but you don't need to fully double everything.

Things similar to this idea are already done in processors - simultaneous multithreading, early branch resolution, conditional instructions, are all ideas that have similar implementation difficulties. So the reason this specific idea is not done is more in line with your last two points rather than the first two.

Has this WSI chip been taped out/verified? I must admit I am somewhat skeptical of TBs of SRAM, even at wafer scale integration. What would the power efficiency/cooling look like?

Section 3 Building a ARM7 CPU is in my view the best example of this - building a pipelined processor is exactly the sort of thing that sounds impressive, but in truth teaches you exceedingly little for the effort it requires. * Architecture wise, implementing a core in verilog versus simulating it in a higher level language are nearly the same from a learning perspective, there really aren't any special. insights from hardware when just starting out. And doing this in verilog just makes the entire process more obtuse because its hardware. * Microarchitecture wise, a simple pipelined processor teaches you basically nothing about how a modern processor (i.e. superscalar, out-of-order) functions from the perspective of a person higher on the stack. If you were to be using it to springboard to writing an OoO superscalar processor, it would be far better to be given a well-designed multistage core and the assignment be to add pipelining - this avoids the student implementating the multistage core of the processor, which is really mostly straightforward implementation of simple but large state machines, and instead focuses purely on the pitfalls of pipelining while also giving a model of good verilog/processor writing practices.

Another thing which is sortof weird is worrying about the type of RAM - that's getting just way into the weeds of processor design and teaches you nothing except stuff like "how to interface with dual port versus single port RAM blocks". And given the time constraints it makes more sense to just simulate memory anyways.

But for many the drawbacks of ranked choice systems are far more preferable to those of FPTP. Additionally, Arrow's theorem only states that the spoiler effect cannot be completely eliminated by ranked choice - it says nothing about how often such an event actually occurs, and in practice spoiler candidates will occur less frequently with ranked choice systems compared to FPTP.

Additionally, rated choice voting systems are not subject to Arrow's theorem.

> and the model math is using is evidently not very good.

There isn't really one "model of math". Plenty of people study alternative foundations of math or various axioms you can attach to ZF, etc. and its great that they do - choice, and lots of other set theoretic/foundational stuff, is very weird. But the reason mathematicians have largely settled on ZFC as a default is that it (and maybe nowadays + an inaccessible cardinal) also reduces some pathology and makes math more convenient.

Let's list some things which are related to the axiom of choice: * The cartesian product of nonempty sets is nonempty (equivalent) * The reals can be partitioned into more parts than there are real numbers (consistent with the negation of choice) * All vector spaces have a basis (equivalent) * All commutative rings have a maximal ideal (equivalent) * All fields are contained in an algebraic closure (implied by choice)

The first two, at least to me, are similar to Banach-Tarski in that they are things I would like to be true and false, respectively, and are not if we do not accept choice. The point here is that the weirdness of Banach-Tarski is as much related to the axiom of choice as it is to generally the fact that (uncountable) infinities are just very weird, and while introducing choice does introduce pathologies like Banach-Tarski, it also reduces some.

The last three illustrate a more practical perspective. Suppose a universal decree, that math is no longer allowed to be done in ZFC, you could only use ZF, was imposed on mathematicians. This doesn't really change anything - all that mathematicians will do is, if they were perfectly fine with choice before, simply replace instances of "vector space" with "vector space with a basis", or "commutative ring" with "commutative ring with a maximal ideal", because the types of mathematical objects they care about are the ones with these desirable properties - they only use the axiom of choice because it ensures the general objects they work with also have those desirable properties. And mathematicians (often even the same ones as before, just wearing a different hat) who _do_ care about choice will continue studying those weird instances of the general object just as before.

As a final point for the convenience of axiom of choice, there are lots of instances where a proof will use choice purely as a matter of convenience. Maybe with some technical set theory or a smarter argument its possible to completely eliminate the requirement of choice, or use a weaker, less-objectionable version of choice. Other times, while a general vector space having a basis requires choice, the vector spaces in your particular application have a basis regardless of your thoughts on the axiom of choice. But of course its much easier to simply assert "vector spaces have a basis" and introduce a (faux) dependence on the axiom of choice. For a concrete example of this last point, see https://mathoverflow.net/a/35772. While about inaccessible cardinals/"universes" rather than the axiom of choice, the principle I want to illustrate is the same. The proof of Fermat's last theorem, if you trace citations back, eventually depends on "universes". But the actual invocations of those theorems are applied to objects where that generality is not required.

(Why not simply only have done the non-general scenario? Because "holds in general with additional axiom, holds in specificity without" is more knowledge than just "holds in specificity")

Sortof - sioyek will parse the text and attempt to create a table of contents (this does not modify the base pdf), but the accuracy/usefulness varies. For papers I have found it works reasonably well, just sometimes misses a few sections or includes a few junk entries. For textbooks, especially ones with heavy math typesetting, there are too many junk/missing entries a lot of the time in my experience.

For papers or for textbooks I only need a small subset of, I find the autogenerated table of contents + bookmarks is typically sufficient. For other cases I like to use pdf.tocgen (https://github.com/Krasjet/pdf.tocgen) to semi-manually generate a correct table of contents.

Should add (2021) to title.

I disagree with some of the later points the author brings up - the sentiment of "SEO is ruining searching for things" is certainly agreeable, but I find the connection to academia a bit more specious and extreme.

> Today it would be effectively impossible for a physicist to personally know all the other top physicists. In the U.S., the number of physics PhDs awarded per year grew from under 20 in 1900, to nearly 2,000 by 2018. Imagine trying to get all the leading theorists together for a group photo today.

What is the point here? It is a fact that knowledge expands and grows. If you collect all of the "top physicists", by some magical metric, you get physicists from vastly different subfields who probably wouldn't understand each others research. This is a good thing! We understand so much more about physics that you have entire sub-sub-fields warranting the same attention as quantum mechanics did at its beginnings.

> A doctorate just means much less, now, than it did. The fraction of physicists who can generate genuine new knowledge and insight is fairly small. Most of today’s physicists would not measure up to the standards that the field held for itself when that photo was taken in 1927.

This is... frustrating. Of course if you compare to the mythologized greats of physics you won't be as good. That doesn't mean what you are working on isn't new, or interesting, or important. The author seems to want every PhD to discover quantum mechanics. But that's already been done. Also, I would pushback on the author having the ability to judge whether a physicist is able to generate new knowledge.

> The dilution of the top experts doesn’t only mean that it’s harder for you and me and the New York Times to tell the first-rate researchers apart from the sea of mediocrities.

Work from physicists from the early to mid 1900s is of course going to seen to be on average better than the average work in physics today - we aren't going to still be looking at the stuff that was below average/didn't work out! There is an inherent survivorship bias in what ideas survive 100 years, but that doesn't mean that all ideas from that time were of equal importance. Also, if you are looking at the probably greatest period of advancement in physics, as the author seems to be, given their prior focus on physics, of course modern day stuff is going to pale in comparison. But of course if you compare to the 99th percentile everything else seems lesser.

> It’s also harder for those on the cutting edge to find each other.

This is just untrue - the people working in these areas know which journals/conferences are higher quality and which are worse, and are in direct communication with their colleagues. As an outsider attempting to peer in, yes, you won't have this knowledge. But the conclusion from that isn't that the experts also struggle with this, the conclusion is the hidden knowledge should be made more open.

> When an important new theory is published, instead of being circulated and torn apart by the two dozen people whose opinions matter, it’s sent off to some nameless apparatchik for the hollow bureaucratic ritual of peer review, so that the reviewer can demand bogus citations for himself and his friends.

I don't see where the author is going on this - the "circulated and torn apart" step still happens, though perhaps less violently. Academics do not see a peer reviewed article and assume "this is a good article, I will accept it at face value", they read over it, they might discuss it with colleagues, the authors presents at conferences, etc. Peer review is not the end of the process, its just a gatekeep to prevent the lowest quality from making it through. Yes, there are problems with the process, but certainly not enough to warrant the disdain the author shows. Furthermore, preprints are a common feature of some fields. If your theory is truly groundbreaking, post it to the arXiv, and it will eventually make its way around. Perelman's work is only on the arXiv, for instance.

I hope my points make some degree of sense.

A statement by the University of Michigan has been made: https://publicaffairs.vpcomm.umich.edu/key-issues/licensing-...

The full quote is "Clearly formulates new ideas for advancing equity and inclusion at Berkeley and within their field, through their research, teaching, and/or service." The last two are naturally more relevant for fields where advancing DEI through the research itself isn't as feasible. The examples given in the rubric are all of teaching/service categories, as well.

The version in the paper only is ever multiplying by 1 or -1, so you do not need a full multiplier circuit. The weights are also stored as signed integers, not floats, so no complicated floating point circuity. I am not sure what current state of the art is, but considering the cost of multipliers/floating point circuity I would be surprised if it changed to those if signed integers work.

All branch predictors need some way of storing their state and selection logic, and the way a perceptron branch predictor stores its data is just a big table indexed by some hash of the program counter of the branch, which is pretty standard for branch predictors. Also, all branch predictors have a sort of "backpropogation" in that pipelined processors produce the actual result of the branch (possibly many) cycles later, so this also is not as much of a factor. Since the training is a function of the weights you do not need to store extra data beyond the threshold, but that is already being computed as the prediction anyways.