ar-nelson

Forgejo was supposed to, but that feature has been in limbo forever. For better or worse, ATProto is easier to implement, because most of the important details of ActivityPub are undocumented compatibility details of implementations like Mastodon (the spec alone isn't very useful for a working, compatible implementation).

AI would make it easier to implement and test complex protocols like this, but the community using Forgejo is mostly against it, which means it'll likely stay in limbo for a long time.

The smallest quant is still over 100GB, so I don't know how that fits in 64GB RAM. It would have to be partly streamed from disk.

I thought that sounded good, so I looked it up. I can find Dells with 1920x1080 screens for ~$560, but the 4K ones, even refurbished, are closer to $2,000. But maybe I just wasn't looking in the right place?

> It does not assume that progress will be in LLMs

If that's the case then there's not as much reason to assume that this progress will occur now, and not years from now; LLMs are the only major recent development that gives the AI 2027 scenario a reason to exist.

> You have have 2 AIs, then 4, then 8.... then millions

The most powerful AI we have now is strictly hardware-dependent, which is why only a few big corporations have it. Scaling it up or cloning it is bottlenecked by building more data centers.

Now it's certainly possible that there will be a development soon that makes LLMs significantly more efficient and frees up all of that compute for more copies of them. But there's no evidence that even state-of-the-art LLMs will be any help in finding this development; that kind of novel research is just not something they're any good at. They're good at doing well-understood things quickly and in large volume, with small variations based on user input.

> But the thought experiment doesn't seem indefensible.

The part that seems indefensible is the unexamined assumptions about LLMs' ability (or AI's ability more broadly) to jump to optimal human ability in fields like software or research, using better algorithms and data alone.

Take https://ai-2027.com/research/takeoff-forecast as an example: it's the side page of AI 2027 that attempts to deal with these types of objections. It spends hundreds of paragraphs on what the impact of AI reaching a "superhuman coder" level will be on AI research, and on the difference between the effectiveness of an organizations average and best researchers, and the impact of an AI closing that gap and having the same research effectiveness as the best humans.

But what goes completely unexamined and unjustified is the idea that AI will be capable of reaching "superhuman coder" level, or developing peak-human-level "research taste", at all, at any point, with any amount of compute or data. It's simply assumed that it will get there because the exponential curve of the recent AI boom will keep going up.

Skills like "research taste" can't be learned at a high level from books and the internet, even if, like ChatGPT, you've read the entire Internet and can see all the connections within it. They require experience, trial and error. Probably the same amount that a human expert would require, but even that assumes we can make an AI that can learn from experience as efficiently as a human, and we're not there yet.

I find Yudowsky-style rationalists morbidly fascinating in the same way as Scientologists and other cults. Probably because they seem to genuinely believe they're living in a sci-fi story. I read a lot of their stuff, probably too much, even though I find it mostly ridiculous.

The biggest nonsense axiom I see in the AI-cult rationalist world is recursive self-improvement. It's the classic reason superintelligence takeoff happens in sci-fi: once AI reaches some threshold of intelligence, it's supposed to figure out how to edit its own mind, do that better and faster than humans, and exponentially leap into superintelligence. The entire "AI 2027" scenario is built on this assumption; it assumes that soon LLMs will gain the capability of assisting humans on AI research, and AI capabilities will explode from there.

But AI being capable of researching or improving itself is not obvious; there's so many assumptions built into it!

- What if "increasing intelligence", which is a very vague goal, has diminishing returns, making recursive self-improvement incredibly slow?

- Speaking of which, LLMs already seem to have hit a wall of diminishing returns; it seems unlikely they'll be able to assist cutting-edge AI research with anything other than boilerplate coding speed improvements.

- What if there are several paths to different kinds of intelligence with their own local maxima, in which the AI can easily get stuck after optimizing itself into the wrong type of intelligence?

- Once AI realizes it can edit itself to be more intelligent, it can also edit its own goals. Why wouldn't it wirehead itself? (short-circuit its reward pathway so it always feels like it's accomplished its goal)

Knowing Yudowsky I'm sure there's a long blog post somewhere where all of these are addressed with several million rambling words of theory, but I don't think any amount of doing philosophy in a vacuum without concrete evidence could convince me that fast-takeoff superintelligence is possible.

It's surprisingly funny for AI, but there's just so much of it... It has no sense of pacing. It repeats the same jokes for too long, without including bits of normalcy in between as a breather. Still, it's a lot better than I would have expected from something written 100% by AI, and I'm very curious what the prompt involved.

For everyone who's having a hard time parsing what Octelium does, I found this page to be the clearest explanation: https://octelium.com/docs/octelium/latest/overview/how-octel...

It's clearer because, instead of starting with a massive list of everything you could do with Octelium (which is indeed confusing), it starts by explaining the core primitives Octelium is built on, and builds up from there.

And it actually looks pretty cool and useful! From what I can tell, the core funtionality is:

- A VPN-like gateway that understands higher-level protocols, like HTTP or PostgreSQL, and can make fine-grained security decisions using the content of those protocols

- A cluster configuration layer on top of Kubernetes

And these two things combine to make, basically, a personal cloud. So, like any of the big cloud platforms, it does a million things and it's hard to figure out which ones you need at first. But it seems like the kind of system that could be used for a homelab, a small company that wants to keep cloud costs down, or a custom PaaS selling cloud functionality. Neat!

I really like Mithril.js (https://mithril.js.org/), which is, IMO, as close as it gets to web IMGUI. It looks a lot like React, but rendering happens manually, either on each event or with a manual m.redraw() call.

I find it interesting that this kind of "animal intelligence" is still so far away, while LLMs have become so good at "human intelligence" (language) that they can reliably pass the Turing Test.

I think that the LLMs we have today aren't so much artificial brains as they are artificial brain organs, like the speech center or vision center of a brain. We'd get closer to AGI if we could incorporate them with the rest of a brain, but we still have no idea how to even begin building, say, a motor cortex.

Obligatory shameless plug whenever Zod is posted: if you want similar, but much more minimal schema validation at runtime, with a JSON representation, try Spartan Schema: https://github.com/ar-nelson/spartan-schema

A thought I had while reading this: what about putting a flexible membrane above the wheels (or belt) with the dots? This would require the user to press down to feel the dots, but it would remove the issue of fingers or hair getting caught in the wheels.

There is at least going to be a Paper Mario: The Thousand Year Door remake coming out sometime in 2024. I'm hoping it's a sign of more traditional Mario RPGs to come, but who knows. The TTYD remake is something that most of the fanbase thought would never happen, so I guess anything is possible.

Disagree. Mastodon isn't so much a Twitter alternative as a Tumblr alternative with Twitter-like UX. And its value proposition is much the same as Tumblr: curate a collection of interesting discussions and memes from across the network, and in the process find people who like the same kinds of things and build connections with them.

It's not about "number go up" in the same way Twitter or Instagram is because there's no algorithm to give it a feedback loop.

Both Twitter/Instagram and Tumblr/Mastodon are about getting community/attention from strangers by posting what you want and letting them come to you (as opposed to Reddit or forums where you join existing discussions). But the Twitter/Instagram model relies on an algorithm so it only benefits those who are already famous or who invest in gaming the system, whereas Tumblr/Mastodon make it easy for everyone to find their community of a few dozen mutuals without being buried by the algorithm.

I've had the opposite experience. On Mastodon I get a small, but consistent, amount of engagement with my posts, and I have over 400 followers. Twitter was like shouting into the void.

FreeBSD jails don't have a one-command way to install a preconfigured jail for a specific service. And they don't have a way (that I know of) to write a config file describing a bunch of jails and services, then run that file and install all of those services on a new machine.

Docker is less elegant than jails in a lot of ways, but not having to figure out how to install each service on your own is a huge advantage.

I tried using Ansible to set all of my services up, and it really didn't work. Probably because I was using a different machine as the Ansible controller, and there is no Ansible plugin to connect to jails on a remote machine. I tried writing my own, and it just barely worked, but with a lot of bugs I could never fully track down (because Ansible's internals aren't documented at all). Eventually it just wasn't worth it to keep trying to use FreeBSD.

I really wanted to like FreeBSD. ZFS is cool, jails are cool, pkg has such a wide variety of packages that it's comparable to Arch.

In a world before Docker and btrfs, this would be enough to make installing FreeBSD worthwhile. I tried it anyway, for a home server, and it was mostly painful.

Jails are tricky and require third-party (?) tools and scripts to manage them effectively. And they don't give you much that Docker doesn't already. And, although it's not as bad as Arch, FreeBSD does update quickly and if you aren't frequently updating your server you can be left behind; it's not noticeable until you try to update and a bunch of packages break, or your jails are no longer running the same major version as their host and that causes subtle bugs.

I recently replaced my FreeBSD home server with a basic Debian machine running a docker-compose cluster, and it's so much simpler. I wanted to like FreeBSD. At its core, it's probably a better-designed system. But it requires too much additional knowledge and maintenance for something that Linux and Docker do just as well.

My guess is it's a Magic: The Gathering reference.

Using the Jitsi website has worked well enough for me for years, and the calls are P2P as far as I'm aware. What's the advantage of hosting your own? Just preventing jitsi.org from seeing call metadata?

I'm just imagining a type system based on this. I've had ideas about a Datalog-based type system before, but it was never really practical; this could be the breakthrough that would make it practical!

Yes, but it's rare to see linked libraries that use this as their API, even though it would greatly simplify FFI.

I've had a lot of ideas for cross-language libraries that would need a C API, and this issue always comes up. The idea I had several years ago---but never implemented, because most of the projects I'd use this for are in limbo because I never seem to finish anything---is an API with only one function, which takes JSON and returns JSON, possibly via JSON-RPC. Basically a library that pretends it's a remote service. Slow, yes, but not as slow as some alternatives, and it makes FFI setup with other languages easy.

Would you consider supporting Deno? (The npm module can probably be imported as-is in newer Deno versions, but a dedicated module is always nice and likely wouldn't be much work)

I get the sense that you think I'm saying more here than I actually am. I'm not proposing that this is the only kind of news that should exist, only that it would be nice if it existed for the kind of people that want to read (only) this kind of news.

And I'm generally in agreement that most attempts to quantify 'truth' in media are hopelessly dependent on personal bias---but this mostly shows up in category #2 in my list. In things where you'd never know the difference if it were true or not, because it would never affect you either way.

The reason I thought a news service like this would work better as a government service than a private entity is because a government news service's commitment to the principles I listed could be defined by enforceable laws. "Likely to materially affect people" is something that you could reasonably argue about in a courtroom, just as much as other fuzzily-defined legal concepts like libel or false advertising.

I'm imagining a news agency whose legal responsibilities were defined in such a way that it could be sued if one of the following happened:

1. It reports something that no reasonable person would believe meets the criteria.

2. Readers experience some kind of material harm that could have been avoided if they had read news reported in another outlet but not this one. And this harm is not the result of the reader being in some very small minority of readers (say, <1%), because after a certain point this will always be true for things that affect a very small number of people.

I don't see how an unmoderated, anarchic space like 4chan is close to what I described; the entire point of what I had in mind is a very specific kind of moderation: stories are only published if (a) they're reasonably likely to materially affect some significant portion of readers sometime in the next 6 months, and (b) there's something they can do about it or in response to it.

For example, if the readership of this news service was entirely US-based, then it would only publish a single article on the Ukraine war---when it started---and then might only ever mention it again if it has a direct practical effect on US residents, like travel restrictions.

The way I see it, news serves three purposes:

---

1. To let you know about local or world events that could affect you and those close to you.

2. To let you know about world events that affect others far away, in order to judge the effectiveness of political decisions and the necessity of future political decisions.

3. As a form of entertainment derived from the ongoing story of world history (or celebrity gossip, or whatever else).

---

#1 is the reason that news feels important enough that tuning it out completely feels irresponsible, but it's a very small component of most news.

#2 is perhaps necessary for democracy to work, but it's so easy to manipulate, and the incentive to manipulate it is so high, that it's questionable whether this type of news has ever existed without being more manipulation than fact---and this has been true since long before the internet.

#3 is the real reason most people (myself included) read news, even when they convince themselves it's #1 or #2. And it becomes unhealthy because, as long as you're convinced you need to care deeply about what you're reading because it's actually #1 or #2, it will inspire constant anxiety.

I would be interested to see a type of (perhaps government-funded) news service whose sole purpose is to publish only news that fits into category #1: if it is not reasonably likely to affect the average reader in an actionable way in the next 6 months, then it can't be published in this outlet.

This could be implemented as an import assertion[1]:

import foo from 'http://example.com/foo' assert { sha256sum: '...' }

This feature would even be useful in the web platform in general. Would also be nice to see an assertion for importing plain text files (not just JSON files) and assertions for per-module Deno permissions.

[1]: https://github.com/tc39/proposal-import-assertions

> Should it have a "repository" of whitelisted domains?

Yes, actually.

Deno's security model allows you to whitelist file paths and network domains with --allow-read, --allow-write, and --allow-net. But this doesn't apply to static imports. I don't see why it needs to be this way.

There should at least be a flag or deno.json config option that would require you to explicitly approve every new import domain, which would prevent typo attacks and make long chains of transitive imports obvious. deno.land could be allowed by default.

iirc I wanted to port it to Guile originally but there was one major bug in Guile R7RS support that was stopping me. I think it was something about the way imports were named. I submitted an issue but it languished for at least half a year, I think because the dev in charge of that part of the code was busy with something else and not actively working on Guile. I haven't looked at the issue in a couple years, maybe it's fixed now.

And thanks!

(edit: I looked it up, they merged my fix for the bug in May 2021, maybe supporting Guile would be feasible now? https://debbugs.gnu.org/cgi/bugreport.cgi?bug=40252)

From my experience working on Schemepunk, Gerbil has a lot of interesting functionality but it's also very immature. Frequently modules will not quite match their documentation, or a simple mistake will cause a segfault. I wouldn't feel comfortable using it in production for something internet-facing.