rsfern

8 ms·

91.
by rsfern 2y ago
I’m not GP, but I don’t think their position is necessarily in tension with leveraging computation. Not all FLOPs are equal, and furthermore FLOPs != Watts. In fact a much more efficient architecture might be that much more effective at lev
92.
by rsfern 2y ago
It’s not that off topic IMO, NIST (along with all the other federal agencies) is finalizing restructuring/reorganization plans this week in response to one of last month’s executive orders. It’s not clear when the extent of the cuts wi
93.
by rsfern 2y ago
I think you’re right that it’s better to avoid these issues in the first place, but your perspective on furloughs instead of layoffs seems a bit fatalistic. I think a company that did rolling furloughs (and importantly was transparent about
94.
by rsfern 2y ago
Yes - this isn’t some new or controversial phenomenon, DoD has funded all kinds of academic research for a long time through organizations like the Office of Naval Research, Air Force Office of Scientific Research, etc. The article says one
95.
by rsfern 2y ago
Octonions are cool! The one application I’m familiar with is in crystallography - you can represent the interface between two crystals with a unit octonion https://doi.org/10.1016/j.actamat.2018.12.034 (Open access pdf
96.
by rsfern 2y ago
sure - there are absolutely lousy PMs as well as great PMs, but this kind of situation happens in all kinds of organizations and I think you are overgeneralizing. There are lots of reasons that excellent scientists might not want to activel
97.
by rsfern 2y ago
I don’t know what the distribution of advanced degrees is among NSF program managers, but I strongly reject the implication that career program managers are somehow a lower tier or less well suited for the job. I’ve personally served a few
98.
by rsfern 2y ago
> If fusion is commercially viable, it doesn't require taxpayer (i.e. mine and yours) money to be developed. This is the reason that NSF exists - many exciting and potentially promising future technologies are not commercially viabl
99.
by rsfern 2y ago
The SpaceX/DOGE/FAA situation seems to be a bit more porous than your are making it out to be, per Wired. I think it’s unclear how that will play out https://www.wired.com/story/faa-doge-elon-musk-space-x&#x2F
100.
by rsfern 2y ago
“Middle managers” is a pretty gross mischaracterization of NSF program managers. They are more like investment portfolio managers, responsible for coordinating and leading the external reviews of research proposals, evaluating the proposal
101.
by rsfern 2y ago
Same story with people at USDA working on the H5N1 response https://www.reuters.com/world/us/usda-works-rehire-bird-flu-...
102.
by rsfern 2y ago
For now.
103.
by rsfern 2y ago
So what did the 40k for your research cover? It sounds like you are probably discounting the cost of your own salary as a graduate student supported by the grant, which is usually the bulk of the direct cost for NSF grants like this and pro
104.
by rsfern 2y ago
Consider a complex LLM pipeline with multiple steps. each LLM evaluation has an associated prompt to shape the style/context of the response. Conventionally these prompts are treated like hyperparameters that have to be manually adju
105.
by rsfern 2y ago
Lots of politically charged comments on this discussion. If you disagree with NSF’s merit and broader impact review criteria, that’s fine and it’s important to have a healthy conversation about it, but I think it’s important to take a step
106.
by rsfern 2y ago
Yes - for example when modeling solid materials it’s common to use a plane wave basis set for the electronic wave function instead of atomic orbitals
107.
by rsfern 2y ago
FYI, the author used to be a programming languages theory researcher before changing fields to biomedicine. So I think you’re reading into the field a bit You last point definitely resonates though
108.
by rsfern 2y ago
Here’s the preprint - https://chemrxiv.org/engage/chemrxiv/article-details/652414e... Your impression is accurate, they have a bunch of experimental synthesis results and then some modeling to try to understa
109.
by rsfern 2y ago
Usually standard reference materials are just regular materials that have been measured really precisely and carefully. So you’re paying for the certificate of analysis and the tracking of lot numbers and such for wherever the material was
110.
by rsfern 2y ago
Review by committee sounds like not very much fun, but reading the description it could actually be kind of nice to work through one’s thoughts about a paper though web chat with (anonymous) friends. One of the big slogs with the current so
111.
by rsfern 2y ago
For this particular issue, it took a while but the NIST password policy does follow 800-63 now. They changed it a while back.
112.
by rsfern 2y ago
I think you’re just looking for the surprise factor in the wrong place. The notability is all about Sandia’s public release criteria, which are pretty much orthogonal to whether or not the information is publicly known. I don’t think the au
113.
by rsfern 2y ago
In the original KAN paper, they do two things to address this: first they have some sparsity-inducing regularization, and second they have a symbolification step so that you can ideally find a compact symbolic model after learning a sparse
114.
by rsfern 2y ago
They say in the text that it’s the average value of the data points they fit to. I think whether to count it as a parameter depends on whether you consider standardization to be part of the model or not
115.
by rsfern 2y ago
Doesn’t everyone? But seriously, my actual question is whether an ideal resistor network can compute nonlinear functions since the individual resistors are linear in their input, ignoring possible nonlinear effects like temperature depende
116.
by rsfern 2y ago
This is really cool, but I was confused by the framing as a resistor network since I think that should be linear (to first order? I’m not an EE) What they have is a transistor network, and they constrain all the transistors to the ohmic reg
117.
by rsfern 2y ago
Follow-up on this feedback - it would be nice if the UI supported tapping individual blocks sequentially like the game strands does. Currently (on safari on iOS at least) if you go to highlight e.g. 6+4 one at a time the UI reports “invalid
118.
by rsfern 2y ago
I don’t see a PyTorch import, and the transformers import is just for the tokenizer which I don’t really consider a nontrivial part of mamba So it’s just numpy and einops, which is pretty cool. I guess you could probably rewrite all the ei
119.
by rsfern 2y ago
The first link on the list is “the annotated transformer” which goes into a fair amount of detail on attention, it’s an annotated version of the paper that introduced scaled dot product attention
120.
by rsfern 2y ago
I like the asinh scale, I’ve come across it before as “signed pseudolog scale” - your name for it is nice and descriptive, do you have a writeup about it somewhere? https://win-vector.com/2012/03/01/modeling-t

More ›