davisp

> On correctness: LLMs may make mistakes. But these videos are code sitting in our repository, and we can maintain them. Every report and review becomes a fix in the source, and everyone who watches after that gets the corrected lecture. Every recorded lecture on the internet is stuck with its mistakes. Ours will continuously improve.

Models change. Models are unpredictable. The assumption that we can constrain them between prompt bumpers has not been proven and seems unlikely to be provable given how these model works.

Beyond that, how do you know if the reports and reviews are even correct? Are you just gonna rely on the kindness of others to provide that value? Of all things that LLM’s have killed, I’m pretty sure the open sharing of knowledge and understanding are the first on the chopping block. I don’t care whether or not you’re pro or anti LLM, publishing knowledge for reputation is now a dead end for anyone wanting to make a living.

Wrapping others’ knowledge in a black box that may or may not accurately represent that knowledge is basically enshitification on steroids. The reason we hold folks like Lovelace, Sagan, or Feynman in high regard is because they weren’t wrong when they shared their understanding and knowledge.

But the best of luck to everyone involved! I’m sure everything will work out!

> If I build a Markov chain based on a statistical analysis of word sequences in Hamlet, and then use it to produce a new sentence that isn't found in the text of that work, I have not created a derivative work of Hamlet under any applicable sense of that term.

Uh, that is exactly what a derivative work is. You literally specify that Hamlet is an input to your work. I believe you're conflating derivative with transformative. You're certainly creating a transformative derivation of Hamlet, but you are by definition creating a derivative work by training a Markov chain on the text of Hamlet.

The obvious follow up here is whether an LLM is creating transformative derivations or not. A lot of folks argue that yes, an LLM spitting out statistically sampled code that matches existing code is not transformative and is (or might be) infringing the terms of the license it was released under. Others argue that there's not an exact copy of the original source in the LLM's weights so by definition it must be a transformative work. I think it's a pretty obvious "somewhere in the middle" that is gonna make a bunch of lawyers a whole lot of money.

Personally, I don't care one way or the other. I'm one of the folks that thinks software shouldn't be copyright-able in the first place.

I’m gonna give this a very charitable read by saying that while I find the ways that the treatment of burn victims was advanced by abhorrent means, we as a society have still benefited from those means.

> So once the model is created there isn't a good reason to encumber it by how it was created.

I am trying to be very specific here. I assume no untoward motivations from the parent commenter. I am not intending to cast aspersions. Whoever wrote this, I feel no ill will for you and this is not meant as a personal slight.

And I will be very clear, this statement as written could probably be defended because of the “by how it was created” clause.

However, “So once the model is created there isn’t a good reason to encumber it” is so… fucking I don’t even know, because what the actual fuck?

I apologize for the profanity, I really do. But, really? Are you fucking kidding me?

These models should not exist. Ever. By any means. Do not pass Go. Go directly to jail.

I understand the engineering brain enough to contemplate abstract concepts with detachment. That’s all I think happened here. But holy fuck, please pause and consider things a bit.

I'm no rustc expert, but from what little I know it seems like disabling panics for a crate would be an obvious first step. You make a great point though. Turning that into a compiler assertion of "this function will never panic" would also be useful.

What caught my eye in the article was the desire to have something that doesn't panic with a release profile, while allowing for panics in dev profiles. Based on other comments I think the general "allow use of std, but don't panic" seems like something that could be useful purely on the "Wait, why doesn't that exist?" reactions.

Does anyone know if there's an obvious reason that adding a `no_panic` crate attribute wouldn't be feasible? It certainly seems like an "obvious" thing to add so I'm hesitant to take the obvious nerd snipe bait.

Absolutely correct! Had the bird strike not occurred, there wouldn’t have been a crash. Had things with the go around been handled properly, there would have been no crash.

Etc etc. The fact that a wall was 50m out of compliance or whatever it ends up being will be a footnote at best in the review of this crash.

That’ll just depend on whatever code was deployed to the cluster. For the clusters I used to operate, the answer would be absolutely all nodes talk to all nodes all the time.

I personally never operated anything above roughly 250 nodes, but that limit was mostly due to following the OP’s advice about paying attention to the configuration of each node in the cluster. In my case, fewer nodes with fancier and larger raid arrays ended up being a better scaling strategy.

> If you actually have 1000 nodes worth of work, the heartbeats are not at all a big deal.

I think you’re missing the fact that the heart beats will be combined with existing packets. Hence the quoted bit. If you’ve got 1000 nodes, they should be doing something with that network such that an extra 50 bytes (or so) every 30s would not be an issue.

Most likely those are just the states where they already have a tax presence. For whatever reason they happen to currently employ folks in those states so adding employees is easy. Adding new states means getting lawyers and CPA type folks involved which is a hurdle to hiring in larger organizations.

A wild tangent but reading “heritable metabolizing” really hit me on the “are viruses alive” question.

I’ve been around enough biotech to have considered the differences between plasmids and viruses versus archaea, bacteria, and eukaryotes. I’ve always considered “heritable change” as the base definition of “life”. As in, “life” is progeny resemble their parent(s)? Or “heritable change”.

“Heritable metabolizing” quite nicely captures that difference between the levels of single molecule “life” and singular/multicellular “life”.

Apologies for the random aside, it was just one of those random “I have a vague idea of why mitochondria are important, but I don’t see them as fundamental” parts of my “What is life?” definition being refined.

Sometimes science doesn’t have to be precise to demonstrate a result.

Consider trying to measure feedback from a microphone and speaker. You don’t have to be an expert to know that there’s a quick change in system behavior when the microphone gets too close to the speaker.

This matches my experience. I spent a decade operating Erlang clusters and using hot code upgrades is a superpower for debugging a whole class of hard to track bugs. Although, without the tracking for cluster state it can be its own footgun when a hotpatch gets unpatched during a code deploy.

As for relups, I once tried starting a project to make them easier but eventually decided that the number of bazookas pointed at each and every toe made them basically a non-starter for anything that isn’t trivial. And if its trivial it was already covered by the nl (network load, send a local module to all nodes in the cluster and hot load it) style tooling.