eyegor

Haproxy is great in the end but it is pretty awful to work with. Once you've toiled in the mines of acl commands, going back to nginx/caddy is a breath of fresh air. Unless recent versions have completely changed the game, using haproxy when you don't have to is a massive time waster.

The m1/m2 chips are generally stupid effecient compared to Intel chips (or even amd/arm/etc)... Are you sure the power draw is comparable? Apple is quite well known for kneecapping hardware with terrible thermal solutions and I don't think there are any breakthroughs in the modern chassis.

I couldn't find good data on the older mbpros, but the m1 max mbpro used 1/3 the power vs an 11th gen Intel laptop to get almost identical scores in cinebench r23.

https://www.anandtech.com/show/17024/apple-m1-max-performanc...

This is not generally true. Amazon is plagued by drop shippers now and as a result most of their solar junk is many years old warehouse stock. Harbor freight tends to be better.

Ex: https://youtu.be/W-fgI6Kyy2I

So what is the value of the "proof of universality"? I'm a bit confused as to what the value pitch is when compared to xxhash3, for example, which runs ~twice as fast and has a large user base (it's used when hashing in dotnet now).

From a dev perspective though, it seems very nice. Short and clean code with decent comments. Although I'm struggling to understand what `polymur_hash_poly611` is doing.

I mean he is in the top ~100 most cited researchers of all time, across all fields, and he's still alive. Guy isn't a great public speaker but his name is on quite a few foundational papers in his field.

https://scholar.google.com/citations?user=0dDO3SAAAAAJ

I'm fairly confident a coding specific model should be a lot smaller - 3b should be plenty if not 1b or less. As it stands, there are quite a few 7-13b model sizes that can predict natural language quite well. Code seems at its surface a much simpler language, strict grammars, etc so I wouldn't think it needs to be anywhere near as large as the nlp models. Right now people are retraining nlp models to work with code, but I think the best code helper models in the future will be trained primarily on code and maybe fine tuned on some language. I'm thinking less of a chat bot api and more of a giant leap in "intellisense" services.

Yew is pretty well loved, but I haven't had a chance to do much beyond toy projects with it.

https://yew.rs/

https://github.com/yewstack/yew

You can create true "early return" behavior by writing a lazy computation builder [0]. It feels esoteric but taking the time to craft a domain specific computation builder can drastically simplify the rest of your code base. I don't know if there's a more modern (more ergonomic) way to do this since this has been around since ~2012 or so.

Alternatively you can take the more functional approach and rewrite your function logic to use `if x then (value) else` throughout instead of `return (value)` or `yield (value)`. Early return isn't really a thing in f# but if-else is.

[0] https://fsharpforfunandprofit.com/posts/computation-expressi...

What kind of madman tries to deploy pytorch? If it's a product, convert the finished model to onnx. Otherwise it's probably for a dev and you can tell them to have fun because installing all the gpu libs sucks no matter what. In theory mambaforge should make most of these scenarios less painful (mostly portable faster version of anaconda), but you'll still need the right cuda runtime, rocm, etc.

I vaguely remember nuitka or pyinstaller being able to handle this for me, but I used vms or qemu to build them "natively" for each target platform. Never found a way to solve cross compiling.

Unless you have a way to manually control the focus, I doubt it. Tracking a tiny target like that above a magnet would confuse the auto focus on just about any camera. With manual focus and very steady hands, maybe?

The best results I've had with small things are using the super zoom telephoto lenses like on samsung s22. I haven't seen a smartphone macro lens that takes decent photos of things yet (unless the subject is perfectly still). Perhaps the latest gen has found a way but in smartphone land you're mostly relying on software post processing since they capture so little light.

Their docs seem perpetually out of date, but they recently released support for 3.10. I haven't been able to try it recently because our projects use 3.10 features but in the past it was easily a 10-100x speedup as long as all the project's libraries worked.

https://downloads.python.org/pypy/

I usually make a venv in ~/.venv and then activate it at the top of any python project. Makes it much easier to deal with dependencies when they're all in one place.

Mojo is vaporware for now, they still don't support classes. There are many other issues, but good luck finding a meaningful size python codebase with zero classes.

Playing rhythm games, I can easily feel ~3ms of input lag. However it's pretty easy to adapt even to 100-200ms of lag, so I'd say most people are in the audiophile placebo territory. I can hit a 5-8ms window reliably even with a ton of lag if I spend 20 mins getting used to how it feels (running a game at 10fps under wine on a 10 year old chromebook with bad wireless gear). Network lag is far more significant than any input lag due to input processing or display pipelines.

Fyi, blake3 was released in 2019 and should probably be used over blake2 unless you have some strong reason not to. It's basically a reimplementation of blake2 with performance tweaks.

https://github.com/BLAKE3-team/BLAKE3

If you need a non cryptographic hash, the xxhash family is crazy fast.

https://github.com/Cyan4973/xxHash

The second to last section "some potential issues with the paper" discusses the top-2 finding.

Tree style tabs is a vertical tab extension for ff and the best ui around if you use a lot of tabs. I can't even use chromium vertical tabs anymore because they feel terrible in comparison. You can also fully style tree style tabs via css.

https://addons.mozilla.org/en-US/firefox/addon/tree-style-ta...

For myself and several friends who started with poor senses of smell, covid boosted our sense of smell for a brief time after having it. My smell was easily 5x as sensitive for a month after having covid the first time. Now it's back to baseline.

As someone with a weak sense of smell, I don't care much for most food. I'm a fan of spicy foods but little else. Perhaps this is because it stimulates other senses. If I'm left alone without being reminded, I'll go 1-2 days without eating on a regular basis.

That's a bit rough as a price tag for an esp32s3 board... They've been available with lcds included for $20 for quite a while at this point. Bare they're around 10-15$.

Wasn't optane a pretty noticeable leap in voltage stability? They could've eventually pivoted to high density packing if they hadn't given up.

Do you happen to be familiar with audio classification? There's been a ton of research on text classification and prediction but not many good papers I've seen for general audio classification. I'm talking more feature extraction, not speech recognition. There are a lot of speech recognition papers. So far I've been stuck on fft - image processing pipeline but I haven't gotten great results in real world tests, only on nice teat datasets.

Personally I don't have much experience working beyond mlp/rnn/lstm/cnn models.

What about: add more dropout or noise layers and train an ensemble of models. Submit the best one. Is this considered dirty?

I'm not familiar with most of these models in detail, but training time is generally less interesting than inference time to me. I don't care if it takes a month to train on $10k of gpu rentals if it can be deployed and run on a raspberry pi. I should definitely look into fasttext though.

You may want to consider adding a note to the top. Seems like a lot of people are lazily skimming/reading the headline and see it as "gzip paper full of beans, gzip approach sucks" when really I see this as "gzip approach not better than dnn models but mostly competes and much cheaper to run". The paper is still solid.

You should try running visual studio 2022 and teams. On a medium sized project vs2022 will happily gobble 10+ gb of ram, and teams will take 1-2 more without supporting any meaningful scrollback. Even with ~500 tabs I can't hit 10gb ram in ff (thanks to tree style tabs to make it feasible to have so many tabs open).

Living in Ohio (no inspection "rust belt" state), I saw a minivan's axle pop loose ahead of me on the highway, which sent it into the concrete barrier pretty much instantly. Perhaps you don't have rust issues where you live, but requiring some minimum mechanical function to register a car should be everywhere in my opinion. Generally inspections are either mechanical/safety or emissions. Emissions are based on the car model in every state I've lived in, so as long as you don't modify your car to blow smoke it'll pass. Mechanical inspections are to make sure there's no major rust, the brakes work, the lights work, tires aren't bald, etc.

Since this is an engine extension, I wonder if it would have any effect when combined with others. For example, timescaledb [0] acts on underlying tables. I wonder if this would have some effect if you did something like

    create table xyz(...) using orioledb;
    select create_hypertable(xyz, ts);

[0] https://github.com/timescale/timescaledb

How do you define "core business"? Any laptop os would crumple purely by running out of resource handles on the connections, even without doing work. I've run into file descriptor / handle size limits on small clusters before, I can't imagine trying to deal with 1m+ concurrent tcp connections.