x3haloed

I’m sorry you’ve been inconvenienced by a murder… Or is that the country we’re living in now? “Murder happens, just don’t bother me.”

I believe this is the least known and most important use-case for LLMs today -- understanding and inferring the meaning in language. We've all seen cases where Google can occasionally surface the right content for indirect descriptions of what you're looking for. The most famous example I can think of is searching for "eyebrows kid" returns images of Will Poulter. Google's knowledge engine is getting pretty good at this kind of thing, but LLMs are way better.

Language models are exceedingly good at understanding the meaning of your language without the use of specific keywords. Here's an example from a recent search I did.

> metals can be flexed or bend, and it will regain some of its prior shape

in Google returns either "ductile" or "shape-memory alloy," which are both incorrect.

> What is the property of a material where it prefers to stay in its current form? This is often found in metals, where it can be flexed or bent, and it will regain some of its prior shape?

in ChatGPT correctly explains that

> The property you are referring to is called elasticity. Elasticity is the ability of a material to return to its original shape after being deformed (stretched, compressed, bent, etc.) when the external forces are removed.

We all know that LLMs can hallucinate, and they are therefore not a reliable source of truth or knowledge. I'm not necessarily trying to say that LLMs are more accurate than something like Google's knowledge engine. The value in LLMs is that they can infer your meaning to some degree of accuracy (just like asking a human) so that you can productively continue your own research in more depth.

I wonder what the play is here. Any ideas?

The key must be somewhere in the statement "we’ve been increasingly asking ourselves how we should work with computers. Not on or using computers, but truly with computers."

I wish I had some experience using Multi so I could picture this better. But is there a chance this is for a sandboxed execution environment that would allow models to interact with software alongside a human counterpart?

Startups always have an exit strategy. Makes sense.

Blame sales and marketing

I think wrappers are fine if they add meaningful value. It does seem that mostly you’re getting charged for a little prompt engineering and a UI skin.

I’d personally like to see a good wrapper for autonomous LLM coding. Nothing we have yet quite fits the bill, IMO.

Exactly. I'm not sure if this is brand new or not, but this is definitely on the frontier.

I was literally just thinking about this a few days ago... that we need a multi-modal language model with speech training built-in.

As soon as this thing rolls out, we'll be talking to language models like we talk to each other. Previously it was like dictating a letter and waiting for the responding letter to be read to you. Communication is possible, but not really in the way that we do it with humans.

This is MUCH more human-like, with the ability to interrupt each other and glean context clues from the full richness of the audio.

The model's ability to sing is really fascinating. It's ability to change the sound of its voice -- its pacing, its pitch, its tonality. I don't know how they're controlling all that via GPT-4o tokens, but this is much more interesting stuff than what we had before.

I honestly don't fully understand the implications here.

Yeah, I'm annoyed that OpenAI has deprecated its text completion models and API. I think there's a ton of value to be had from constrained generation like what's available with the Guidance library.

I don't understand. What's an example of something that could be achieved using a system like this?

I’m shocked they were even able to do this. From what I’ve been reading the last couple years, it seems that the US and UK like to just destabilize countries with natural resources and send in the corporations.

I don’t think we’ll make meaningful progress on giving people their land back until corporations are massively reined in.

I agree with a lot of the general advice here. I definitely agree with you on some stuff, some stuff maybe you could flex on a little bit.

Lots of small-minded idiots here. But this is a giant country. There are different people in different places. Try moving as close to Silicon Valley as you can. You will, without a doubt, find a whole bunch of people with your exact same mindset.

Americans are not lazy. If you want to work hard, you’re in the right place. We’re statistically one of the hardest working counties. This is where all of the antidepressants come in though. We’re overworked, and we don’t have a sense of purpose outside of work.

Which leads me to… you’re doing a lot right now. A ton. It’s pretty difficult and rare to be able to keep up with how you are living, especially if your responsibilities grow. Be careful not to over strain yourself and make sure to keep a sense of self and purpose outside of your hustles. I personally had a meltdown in my early twenties that was partially brought on by these factors.

Finally, like others said, try to find a way to find common ground with the muggles around you. The faster realize that you’re not better than them, the faster you will find relief from the stress.

> I don't think society needs public, or at least public by default, social media. This is even more true now in the age of LLMs and mass data harvesting. Depends on your definition of “social media.” Is a message board / forum considered social media? Because while they can be problematic, they’re still extremely useful. It would be interesting to think about decentralized Reddit and why lemmy is going nowhere.

I’ve also been thinking about this problem for a while. The push model and account portability are definitely the most important dimensions of this issue.

Can’t we just solve this problem with IPFS? And something like DHT tables for protocol-level awareness of state changes?

Your client app would just be responsible for pushing and pinning content to IPFS, scanning state tables for interesting updates, and then sending you push notifications.

Really simple to take your IPFS key with you can switch to a new app.

So what if it's ugly? Is this a serialization format or paintbrush?

Also, I think TOML stdlib gets to the heart of what I've been arguing. People who like syntactic whitespace like YAML. Everyone else (like C devs) don't. It makes quite a bit of sense to me that C devs don't want to touch YAML with a 10-ft pole.

Yeah, if I had my druthers, brace-less shorthand syntax like that would never be allowed. I never use it in my own code.

Beautify takes care of all questions about what an indent means. C-syntax devs never assume an indent has syntactic meaning. We use it has a helpful hint. We innately understand that it's the braces that matter, and it's really easy to Beautify a file if the original author was sloppy.

C-syntax devs read differently. We're not reading a book. We're reading code.

And we generally strongly prefer correctness. Braces avoid all unintended bugs related to where the instructions are located on the screen. Pretty is nice. Structure and correctness are better.

The legibility argument.

What if I told you that C-style syntax can be formatted to use indentation exactly as you prefer?

Amen. My current theory: data science. Data science has become much more prevalent in the last decade, and I suspect data scientists prefer English-like syntax because they're not engineers.

Engineers use language like mechanical components. It should be precise, neat, and functional. We're designing machines. I have a feeling that data scientists don't want any of that. They want to use language as a way to describe data transformations in a format that resembles a log rather than functional instructions. Much like SQL.

I think I have a solution to keep the peace, if we can all universally respect one rule:

Never use YAML outside of the Python ecosystem.

That way, people who love esoteric scripting formats that prioritize legibility over correctness, durability, and maintainability can keep all their tab characters, loose typing, and cryptic syntax. And the rest of us never have to! Those of us who prefer C-style syntax can keep our sanity.

Edit: OK, I think I've finally got my finger on the crux of the issue here. I think I can't explain this whole thing without being spicy. For C-syntax devs like myself, syntactic whitespace is pure madness. Whitespace is not information or an instruction -- it's formatting. Good formatting is helpful and useful, and good C-syntax devs care about legible formatting. In Python (and YAML), the formatting is instructional information. This has the benefit of making all functional code legible. But why the blazes does code have to be legible to be functional? Imagine working with a YAML coworker. You send them a long message. They reply with, "What? This is nonsense." You suddenly realize your mistake. Without separating your paragraphs with an empty line, you've broken the meaning of what you sent. Adding the empty lines back in and re-sending the chat message, your co-worker can now read what you wrote. Without syntactically-accurate formatting, the information you sent was meaningless.

> Not sure I see the point in hating on one specific configuration language. If it was that bad, nobody would use it. And if you still think it's bad anyways, you can always improve on it. But very few actually want to put on the enormous amount of work needed to improve YAML or create a new language.

The point is that this format is like a virus. It doesn't need to be improved. It needs to go away. We already have a lot of great serialization formats that have all the use-cases covered. Whenever someone chooses to serialize configuration into YAML, they are unleashing hell upon all of the humans that have to use it.

Let me put it this way. If I was designing some software to sell to you that requires a config file, why would I skip over TOML (INI), JSON, or XML to pick YAML -- a brand new format that has a lot of funny rules that take training to understand? I would be subjecting you to confusion for no good reason. If I'm designing the software for a developer, I'd use JSON or XML. If I'm designing for a non-dev, I'd use TOML. If I'm intentionally trying to cause suffering, I'd choose YAML.

Edit: I wrote this comment in a way that makes it sound objective. Programmers like to think our arguments are all objective. Truthfully, I think the hate train is really about that fact that YAML seems to be very divisive -- some people really like it, and some people feel very frustrated with the format (like me). When those of us who don't like the format encounter some system, tool, or package that relies on YAML serialization, we lose our sh--. We feel that we're being subjected to a difficult and silly new format for reasons that seem entirely arbitrary.

Here's the thing. We've basically been using TOML since the INI file format from MS-DOS. It works, and it's useable, and we've all seen it. TOML is just an evolution of the INI format to fix some of its shortcomings. YAML blasts onto the scene like "hey, what if we just rewrite the JSON format to make it legible?" which brings with it a million edge cases of problems.

YAML is the new kid on the block. And despite having several great formats for different use cases already, some kind of mass hysteria caused big players to adopt YAML. I suspect everyone who adopted YAML is either a Python dev or they were drawn in by the legibility of the format. It's easy to read. While that's true, did anyone stop to think about what it's like to actually use the format?

YAML is bloody terrible, and I've hated it since the first time I was forced to use it.

JSON or TOML are always preferrable, based on the use-case. If you want humans to be able to change the file, use TOML. If only machines and debug programmers need to see the file, use JSON.

YAML is in this horrible in-between where it's extremely unintuitive at first glance and it's really easy to break. The only positive aspect to YAML is that it's legible. But legible is not the same as understandable. TOML is just as legible, but it's also easy to understand and change. JSON is hard to read and hard to understand, but the structure is very durable.

I'm going to take a guess here and assume that there's a strong correlation between people who like Python and people who like YAML.

NoMachine works great for me :) It’s not perfect, but it’s darn good, and I can consistently use it across all my Mac/Win/Lin devices.

I’m glad someone is working on this! We need it.

Same.

Ugh. This still isn't quite what I want. I can't stand Docker. It's so damn heavy.

All I want is something like the following:

- Sandboxed applications

- With an agent to control which ones run on a machine

- A GUI to easily observe and manage deployments

- Infra-as-code using a sensible file format (anything but YAML)

I imagine this working best with VM-based runtimes like .NET and WASM. The ability to control resources consumption isn't there yet, but I don't see why you couldn't have a runtime that gives fine-grained controls over sandboxing and resource consumption.

This idea came about from observing discourse on replacing conventional hypervisors with WASM/WASI.

Forget Docker. Forget OCI. Forget Kubernetes/SWARM. We just need a simple system for orchestrating apps that are already VMs.

It seems interesting, but I’m having a hard time imagining what an appealing use case would be.

I’ve gotten to a point where Windows is now my least favorite OS, but the fact of the matter is that modern Windows has the widest gaming compatibility of all operating systems.

Proton/Steam seem like the only ones who could possibly challenge that at this point.

I think the point is that creative experimentation has really been pushed to the far edges. You would never see something so daringly unusual on broadcast television these days. What’s left of broadcast television, anyway.

Broad media consumption all just formulaic and uninspired.

The 60s and 70s were just full of weird, exploratory, boundary-bending shit, which is what makes them a time of such progress.

Yep. Most human activity is a blight on animals.

OK, after I read this, I can't stop thinking about how much I love this idea. I've been searching and searching thinking that there must be a toolset in alpha out there somewhere that I can play with.

The only think I can find right now is Krustlet, which is not really what I'm looking for. I want something that I can run on my own hardware. Have you (or anyone else 'round these parts) hears of a WASM replacement for Docker in development?

We're looking for a WASM runtime that can handle orchestration of multiple assemblies on the same machine, I suppose: https://www.docker.com/wp-content/uploads/2021/10/Docker-Web...

EDIT: I found what we're looking for! Docker Engine is a container runtime. What we need is an alternate container runtime that will work with WASM. K8s-compatible options are described in the following article: https://kubernetes.io/docs/setup/production-environment/cont...

It turns out that both CRI-O and containerd can run WASM assemblies via the WasmEdge runtime in crun:

CRI-O: https://wasmedge.org/book/en/kubernetes/cri/crio.html

containerd: https://wasmedge.org/book/en/kubernetes/cri/containerd.html

And you can create WASM containers without docker hub images by controlling crun directly: https://wasmedge.org/book/en/kubernetes/container/crun.html