lemursage

This is so weird, seeing this. Two years ago, I got a customs notice from FedEx asking to fill in my details. That was just a plain email from __some guy__ at FedEx with a PDF file attached. I wasn't expecting any package.

I wrote to their chatbot (of course, no human assistance) and after some time of "prompt engineering," or what one might call coercing, it finally directed me to a human consultant who confirmed it was indeed not a scam, and that it was indeed their messaging.

I opened the PDF, and it was pre-filled with someone else's data, with blank rectangles placed over fields in a bad attempt at redacting them (you could just move those rectangles around to reveal the underlying data).

The package later turned out to be a surprise from collaborators abroad. Years later, I still feel that scam aftertaste whenever I see the FedEx logo.

Reference on LLM architectures

I'm "almost" a lifelong user of a cochlear implant. I got my first one when I was 9. Before I got it, I was communicating through lip reading and speaking, I never knew sign language. Lip reading I still use relatively often -- when I'm at a crowded restaurant, or at an unbearably noisy party, and there's many interlocutors at the table, I persistently stare at their lips. They take me for a great listener, when in fact, I can't hear shit, and I'm desperately switching back and forth between people's mouths to catch what they're saying. I'm out of shape and this takes so much of my brain power to understand people that I often cannot contribute my thoughts.

Though my cochlear isn't perfect, I would never think of not getting it. In fact, I'd probably be rather angry at my parents for not helping me get one as soon as it was possible. During my childhood and up until late college, I've only ever met one person who was so severely hard of hearing and was about my age, and that was where I have been getting my speech lessons before I got my first cochlear implant.

I really like the book's subchapter on colours, wish it was even more elaborated on. Colours are one of the subtle things I so often find difficult to get right.

As to using matplotlib in published research: when I started out as an undergrad, everybody in the research team used OriginLab for plotting -- my impression of it then was pretty good. At some point, I started using matplotlib + Latex + science plots and it caught on, mostly because there's no need to shift all the data around to a separate programme. Scienceplots package does heavy lifting with fonts and styling for specific journals, so it's just a matter of designing the right plot geometry and information density [1].

[1] an obligatory Tufte citation.

In larger companies, and, specifically, bigger projects, systems tend to have multiple ML components, and those are usually a mix of large NN models and more classical (ML) algorithms, so you end up tweaking multiple parts at once. In my case optimising for such systems is ~90% of the work. For instance, can I make the model lighter or go faster and keep the performance? Or, can I make it go faster? Loss change, pruning, quantisation, dataset optimisation etc. -- most of the time I'm testing out those options & tweaking parameters. There is of course the deployment part, but this one is usually a quickie if your team has specific processes/pipelines for this. There's a checklist of what you must do while deploying, along with cost targets.

In my case, there are established processes and designated teams for cleaning & collecting data, but you still do a part of it yourself to provide guidelines. So, even though data is always a perpetual problem, I can shed off most of that boring stuff.

Ah, and of course you're not a real engineer if you don't spend at least 1-2% of your time explaining to other people (surprisingly often to a technical staff, but not ML-oriented) why doing X is a really bad idea. Or, just explaining how ML systems work with ill-fitted metaphors.

For a numerical "physicist" (yes, the quotation marks are indispensable), Sympy was somewhat of a godsend to me. Great for prototyping even more advanced models before optimising them later on in C++.

I haven't used Mathematica much, but I have a feeling that it's still more symbolically powerful (or requires less wrangling) than SymPy? I'd appreciate if somebody with more experience in Mathematica could lay it out flat for me if that's the case.

I wonder if this is in response to Sci-Hub/Arxiv proliferation or what else. Anyhow, I find it beyond rich -- the authors usually do all the typesetting, spell/grammar checks, formatting etc. themselves and now they get to pay for that (or their respective funding institutions). I presume that is what would land under a "processing fee". Beyond atrocious -- some journals have their custom LaTex templates only to discard formatting they put in them at the pre-publish stage. Also, reviewers are probably still not getting paid.

I'm not sure if that's an accurate sentiment, but I feel like there's an academic research winter with grants and funding are being slowly squeezed out of academia and moved to external consortia with loose academic affiliations (at least this is what I _think_ I see happening in Europe). If that's true, this trend of moving the paywall from readership to authorship is just sad.

As an engineer my concern about "google killers" is that I can't see an easy way to scale and control/optimize them in business settings. Apart from factual misstatements happening in the ChatGPT, what about source attribution? How is the relevance of a source determined? How is the flow of information through the network preserved (sourceA => sourceB => sourceC)? With Google we also don't know exactly but I can image some version of PageRank as tuneable. Finally, how to add new pages to index and measure potential "forgetting" that could happen?

Unless somebody could clarify those for me, this is what currently petrifies me -- some uncontrolled black box presenting its clandestine view of the web with no way to follow the breadcrumbs.

Here's the catch -- I don't have a dog! Utter ramblings, I say!

I actually tested a bunch of those on GH Copilot -- sure enough, it does autocomplete certain prompts. No wonder, some are probably known quotes. Anyway, when I'm feeling down in my VSCode environment, I enter some random prompts in search of a pun. Most of the time it's just random rambling but hey:

# (prompt: my dog ...) seems to be interested in the plot.

# I wonder what the dog is thinking about.

# Probably about the food.

# I should feed him.

under some code where I was plotting a forecast.

Maybe many PhDs attend conferences just because they have to have some paper/poster presentation rather than they are genuinely interested in making new connections? That's what I'm guilty of at least.