edude03

Yeah that's a ux papercut - pre flakes nixpkgs was always the nixpkgs on your machine. There is docs but if you're not expecting that to happen you wouldn't think to look up the docs.

But it covers five? Languages so if all equal it’s just 20 hours per language.

I'm not talking about a bit perfect reproduction though, just being able to understand dependencies. Take for example a simple Dockerfile like

``` FROM python:latest ADD . RUN pip install foo ```

If I run this today, and I run this a year from now, I'm going to different versions of `python` and `foo` and there is no way (with just the Dockerfile) to know which version of `foo` and `python` were intended.

Nix on the other hand, forces me to use a git sha[^0] of my dependency; there is no concept of a mutable input. So to your point it's hard to 'upgrade' from version a -> b in a controlled fashion if you don't know what `a` even was.

[0]: or the sha256 of the depedency which yes, I understand that's not easy for humans to use.

Nix has been my daily driver for about 10 years, all 10 on servers, and about 3/4 on a laptop.

The thing that hits close to home for me, is the inability to use software that doesn't support nixes opinions on how to do version management (for example a post of mine from years ago[^0]), software that likes mutable state for its configuration (Gnome for example) and yeah, trying new things that aren't packaged for nix means writing a nix derivation.

That said, I feel like nix does more good than harm for me so the paper cuts are bearable.

[0]: https://community.roonlabs.com/t/unable-to-get-roon-to-start...

Containers are only reproducible at run time not at build time. Once you build a container and pull it down by its sha256, you'll get the same environment each time. However if your Dockerfile does any IO (curl, apt get, pip install etc) you're quite likely to get different images on different machines.

As a heavy container user myself - I've been using containers since I needed to build my own 3.x kernel to test them - docker doesn't solve the reproducibility problem nix solves - IE, I can make a Dockerfile that does `RUN curl foo.com/install.sh` and who knows if that'll work ever again. Nix on the other hand doesn't allow you to do IO during builds[^0] only describe the effect of doing the IO.

[0]: Though apparently darwin (mac) doesn't support sandboxing by default, so you can bypass that but anyway

And of course - often apps ask for permissions they don't need at opprotune time ie "enable development mode to get 100 more gems" or "enable notifications to use this app"

I'm hoping I haven't become an out of touch tech bro, but I believe if you're able to flash libreboot on your software development machine you're able to afford $50.

_should_ you have to spend more money to support bloated software? No of course not, especially since many users of such software _aren't_ tech bros, but as someone in tech, shelling out a little more money seems like a much more pragmatic solution vs having your computer be slow and/or waiting for the industry to change.

Honestly I worry more about hiring an oppositional / defiant / insubordinate “asshole” more than the AI use itself because what does it say about an applicant if their first instinct when challenged is to say “your process is invalid, I’m going to cheat instead”?

Point taken, although I feel like $2k a year would be really cheap if AI delivered on its hype.

This is exactly the problem with optimizing metrics though. It seems like they’re getting more money from less customers but it puts them at might higher risk that those big customers will churn and tank the company which also increases as the price goes up as it will eventually hit a point where they can find something cheaper.

Maybe I'm reading it with rose coloured glasses - but I feel like the only thing kubernetes "did wrong" is allowing them to host multiple control planes. Yes, you need 3+ CP instances for HA, but the expectation is you'd have 3 CP instances for X (say 10) workers for Y (say 100) apps. Their implied ratio was insane in comparison.

Since you can't run the Fargate control plane that indirectly solved that problem for them

Like most tech stories this had pretty much nothing to do the tool itself but with the people/organization. The entire article can be summarized with this one quote

> In short, organizational decisions and an overly cautious approach to resource isolation led to an unsustainable number of clusters.

And while I emphasize with how they could end up in this situation, it feels like a lot of words were spent blaming the tool choice vs being a cautionary tail about for example planning and communication.

I used this product when I lived in a building where I could get a 500mb and a 100mbit connection but not a gigabit connection. I had the server side setup on digital ocean and it "worked" in so far as I could pull 600mbit/s but was impractical since 1) I would get the latency of the worse (and in my case more variable) connection and 2) since it was essentially a VPN to a cloud provider many sites blocked me as a suspected bot.

Nerves and hot code reloading got me into erlang after I watched a demo of patching code on a flying drone ~8 years ago.

While I can't imagine hot reloading is super practicle in production, it does highlight that erlang/beam/otp has great primitives for building reliable production systems.

Daily nix user across Mac and Linux, though I use Mac for actual development. No problems here moving between the two with my dev env defined on GitHub [0]

[0]: https://github.com/edude03/dotfiles

What kind of issues are you having?

Transformers were made for machine translation - someone had the insight that when going from one language to another the context mattered such that the tokens that came before would bias which ones came after. It just so happened that transformers we more performant on other tasks, and at the time you could demonstrate the improvement on a small scale.

Yes although how close to verbatim is debatable. For example there are questions that you’d ask that other people have asked many times before that you’d like the exact answer for (e.g. when does daylight saving time end?)

What does it mean to predict the next token correctly though? Arguably (non instruction tuned) models already regurgitate their training data such that it'd complete "Mary had a" with "little lamb" 100% of the time.

On the other hand if you mean, give you the correct answer to your question 100% of the time, then I agree, though then what about things that are only in your mind (guess the number I'm thinking type problems)?

Envoy is legendary in (dev)ops circles, but I don't understand what it lends to the AI space. I feel like building a separate backend service that runs behind envoy would make more sense but that's just me.

Nice to see Jeff finally got some success running GPUs on the pi. Though, as the conclusion kind of points to, it's not very practical today. To be practical I guess we'd need to see faster bus speeds, newer GPU support AND probably, a more practical way put the setup together (I'd love the see a Ghost Canyon[0] style setup) which would make building a high performance ARM system at home practical.

[0]: https://www.pcmag.com/reviews/intel-nuc-9-extreme-kit-ghost-...

I think the point was this isn’t the norm though. If you know you need to be pinned to a core you CAN configure kubernetes to do so but it’s not the default and therefore you are unknowingly leaving performance on the floor

was surprised by this as well

I’m a bell customer in Canada and it used to be the case that the ISP provided modem had a CPU too slow to run PPPoE at a gigabit despite the ISP selling plans up to 1.5gb/s (it could only do 600mb/s or something but don’t quote me). That model has a sfp ont and so you could swap it into something else with no hacking but now you can only get the model with the ont built it. The new model is better hardware wise but just as bad software wise so it feels like a step back in practice.

I think selling users SFP ONTs is probably the right balance of ISP control vs allowing customer freedom

_generally speaking_ most data science is "we tried this and it improved this metric" the how and why is often poorly understood, and even when it is, it's often later once the system has been working for awhile.

But also I'm not convinced that getting someone to pay more is the same as screwing them over. For example - countries where fines are based on your income, or needs based programs that charge you less when you make less are mostly accepted as social good

IIRC the actual model is private and only downloads once you activate the device so presumably if the company was shutdown you couldn’t use the device anymore unless someone shelled out for training

I mean let’s be frank here. The sort of people who are enthusiastic about this also aren’t going to be stopped by the law especially with the low risk of getting caught.

I’ve been saying this since Tesla released autopilot because I personally think teslas implementation is the best but there isn’t an easy way to compare it to something like radar cruise with lane keeping other than saying they’re both level 2.

That kernel panic sounds like a hardware fault just from first blush. Sounds like it's turning on some device but the device isn't responding after being turned on