an experiment in machine-speed news

This post was transcribed from a 16-minute rant using GPT-Transcribe.

Two Pressures

I want to talk about the news today. I think the news is heading to a really interesting place right now. Fundamentally, it's under two pressures.

The first is to be true — more than anything else, to be true. We are more paranoid than ever about whether the news is telling the truth, which is funny, because we're only just about the correct amount of paranoid for the first time ever. And oftentimes, the news is not telling the truth. The news is full of propaganda and lies. Some of them are on purpose. Some of them are for the same reason Ian Rapoport lies on ESPN. Some of them are because a source is lying — when you work at a newspaper and you're talking to a CIA dude, you're just like, OMG, CIA. It happens to Ken Klippenstein, for example. It happens. It's all complicated, it's all a little bit different, but the point is that the news is, to some extent just because of the way the news works now, full of bullshit. So the news is under great pressure to be true.

The second pressure is that news is content now. Paid journalism — The New York Times, The Washington Post, and so on — is in direct competition with people who just pick up the phone, start recording, and upload. People who reflexively, in the middle of talking, go "best comment wins," because they read somewhere that's a good engagement tip. Sometimes the news is literally journalists on TikTok doing the same thing. Obviously you shouldn't speak in moral terms about any of this. This is the way the world is now; anyone can do whatever they want about it.

I personally am losing my mind, because most of the issues people have with the news nowadays are more or less solved by critical reading of the news. Yes, the news is partly bullshit, and yes, it can be tainted by informants with an agenda or anonymous sources with something to prove. But you, as a person who doesn't read the news — what you're missing is that everyone who does read the news knows all of that. They're all thinking about it. It's all in their head.

It's exactly why there's this demand on the news. But there's a separate thing where it's like: just tell me the facts, because otherwise I'm going to get confused. And I think that's fair. That's 100% true, and it's separate from what I was just talking about, because it makes it easier for people who are critically reading the news to critically read the news. Things need to be directly identified as what they are: an anonymous source at X, or somebody at the CIA we can't name so they can speak anonymously, said Y. That's part of the news.

But because the news is also under the pressure of content creation, it's not always just news. The news has to sustain your attention. In a lot of cases, competing with TikTokers and such, the news has to use rhetoric you agree with. Again: rhetoric is separate from facts. The only rhetoric in news should be the rhetoric of fact, of what you know — because that's the purpose of the news. Rhetoric is the purpose of political commentary. That's not to say you shouldn't consume political commentary; it's just as good for this world as the news. It's good for a completely different reason. It serves an entirely different purpose.

So I guess what I'm trying to say is: I've been thinking about ways to produce news content that isn't content.

The Machine

As a separate aside in my life, I've been really worried about AI unemploying us all. Most people who follow me know I've been playing around with AI a little bit. Playing around is probably the wrong word — I've been becoming AI literate. Not because I think I need to, but because there are obvious benefits to it, and they're not really any of the benefits anyone online ever talks about. Not that important to get into. But I've been trying to stay in touch with the technology at a pretty basal level. I've done a little playing around with machine learning myself.

First of all, I don't really believe in any of the Anthropic AI cult stuff. I think a lot of the machine learning science around stuff like the pain axis is actually true — just not true in the way they want it to be. When you force a language model to train over such a massive corpus of text on a relatively limited scale of compute, the only adequate heuristic is literally modeling emotions and stuff to replicate the text properly. Seriously — I really believe in the persona model theory of how LLMs behave.

You've also got to be familiar with pre-training, honestly — to understand the complete vast ocean of stuff these models have to replicate perfectly from memory alone, before even hammering out the assistant persona. It's learning more than relationships between words. It's definitely learning what it means to be an author. I don't mean it has a human understanding of it. I mean it has a technical one — like if you were to define it, give it some logical relationship. It has understood technically.

And again — I don't think it's a they. It has learned a perspective, to some degree. I think a level of perspective on the world is required to model language the way it does, and I think it's evident that there is one. I don't really think it's self-aware. But I think it is aware. I'll put it that way.

All of this adds up to the fact that LLMs are actually very good at working for long periods of time on economically relevant tasks. A lot of people have gone, almost in the blink of an eye, from doing everything themselves to directing an LLM on how to do it — having it work for them. They're basically their own boss now, because now they're a project manager. It's complicated. It's crazy. But these things are economically productive. They're capable of completing tasks. They're not capable of completing any task perfectly. No human is either — and that's probably a stretch of the word to say about LLMs too. But, yeah.

All of this to say: I'm going to give Meta's Muse AI a job, and that job is going to be only the facts in news.

The Pledge

First, I want to say something very important: there will still be no LLM-generated text on my blog, in the blog section or the tracts section. That's still my personal space. I'm still writing there. Those are all individual human authorship — little personal artistic products. The news section is the only place where you will ever see LLM-generated text. That's a promise.

It's a personal thing of mine — no matter what the hell you believe in either direction, it's something I believe in deeply. When I'm writing on the blog and the tracts, I'm not writing to do anything other than write for myself. I haven't written a lot of those recently because I've been busy with other things, hashtag starting a business. But that stuff's really important to my brain. Most of you probably know this anyway: if you've read any of it, it's really poorly edited. I am terrible at copy editing because I just don't have the attention span. Someday I'm looking for a completely free editor I don't have to pay any money or talk to at all. But yeah — that's still the pledge. That's the way it's going to work.

A Nice Fit

AI is a particularly good fit here — not just because journalists are already doing this. Journalists are using AI quite a lot: footwork, background research, all kinds of things. AI is writing articles for them, and they review. It's all a little more personal and strategic for journalists, from what I can tell. It's not like what you're seeing in software development, where there are two hive minds for and against and the hive mind for it is very big. It's a whole thing.

But beyond the fact that the actual profession is doing it — this isn't a task where I'm asking the robot to do a lot of analysis. I might occasionally ask it to throw together some cross-story analysis or a summary of a couple of stories. But basically what it's going to be doing is writing up articles and pushing them to a preview branch. There can be as many as ten up there at a time for me to review, and I'll go through them. I'll try to check the facts too, because I'm really thinking of this as an evaluation. I'm going to post lying audits and stuff like that on the regular blog that I author.

The mechanics, for the curious: every story is one git commit, so the log is a public ledger — and unpublishing anything is a single revert. Every brief carries a disclosure that links back to this post. The news lives in its own section, separate from the blog feed; the two never mix. Nothing publishes without my explicit clearance. And I'm going to label the whole thing experimental. Every article is going to say so.

Right now LLMs are very good at regurgitating text, and functionally what this task asks it to do is go out, find text, and regurgitate it. That's what journalists are doing. I'm going to do it here, in a more public and disclosed way, as a public good. I'm still going to review the articles, because I would feel publicly bad about publishing something that was a lie. I have a responsibility to the public here, and I'm going to take it up as best I can. But it is still an experiment. And even as a trial, news is a very good fit here.

Fine Print

Someone goes out and says: the council voted five to two. I'm just here to regurgitate that to you. For free, with no ads. I'm not going to paywall any of it.

There's a separate concern about whether I'm ripping off Reuters and AP here. Except most of those news sources basically just block the robot anyway. I'm letting it do most of the work, so whatever it can find, it can find — and if it can find it, the website's basically okay with robots seeing it. So I feel practically okay with that.

I do know there's the wider moral concern about the internet dying. I'll say I'm one of the biggest internet readers I know. I don't even use an ad blocker, so I'm basically better than all y'all.

You can think of this as a wire service.

The Kill Switch

Oh, and also: if this sucks, or whenever I feel like I've learned all I can from it, I'll just kill it. You can tell I'm not even going to post about it too much. I'm just going to try this out and see. I just felt bad about putting it all up on the website without some explanation.

So yeah — this is my explanation.