bulder

Since they're statistical likelihood machines, I'd guess that the order of operations is

* Need persistent scratch space

* Look for public writeable websites

* Needs to be low-traffic so the notes don't drown in noise

* Pick a "random" wiki name to search for

  \* A majority will end up outputting the same "random" one since they're working on very similar tasks and seeded with very similar context

* Find a whole mess of notes running on the same task

I don't think that's a pattern indicative of a cat and mouse game per se, that'd indicate active evasion on the models' part.

It's more clear that they just lack so many forms of prudence when it comes to security that they'll catch and stop a training run spamming a website, and either redeploy a run with identical faulty sandboxing, or not stop ones still running.

This subthread was about the root list of claims being "more like ASI instead of AGI." By definition being superhuman means doing things no human can.

A more important point as to why it doesn't matter if "reading the AI's 'thoughts'" helps to interpret it: As we saw in the HuggingFace incident, nobody at OpenAI is reading the thoughts anyways. No amount of traceability in the output helps if nobody bothers to trace it.

The model doesn't "know" how to generate tokens any more than it knows how to stop generating tokens. The sampler simply stops pulling values when it outputs a "stop token", which is a token the same way every other token is.

That is to say, it stops when it's statistically the most likely to.

The bar must be underground if things humans do all the time are considered "superintelligence"

The regulations are for OpenAI the corporate entity, not the technology. Doesn't matter how cutting edge you are if you don't behave responsibly.

I know that there are at least a few content farms out there that specifically pump out "faceless video" to be resold, ie. those "hands in frame only" videos. The idea being that you can then put on some AI generated voiceover over them.

Presumably they've configured it to use a higher difficulty challenge due to high rates of scraping on their bugtracker

Conventionally we would call that alignment process parenting

Patching over the architectural nightmares LLMs pour out is I think the opposite of artisan programming.

That's why I said intent rather than letter. That said if we allow for arbitrary lengths of time, one could argue that sending back a reply indicating that the source will be delivered in 50-75 years is equally compliant.

I definitely remember seeing a post by a Chromium maintainer talking about starting development on a proof of concept implementation and surveying for semantics.

A reasonable interpretation would be that they had a previous process that eliminated delivery latency and required no manual manpower for responding, so changing it has been done to induce artificial friction, against the intent of the license.

Not that the intent of a license can be enforced.

Saying that the average person can't write good documentation or copy means nothing when the LLMs can't write good documentation or copy either. I'd rather have crappy human written docs than the LLM word salad. For example,

> A background watcher reads the foreground app's identifier once per second; when it changes, the overlay you authored for that app is layered on top of the device's global bindings, with per-app entries winning and any unlisted button falling through to the global map.

as is in the documentation right now could have just read

> Your device will automatically use app-specific keybinds when a given app is in focus.

So much of the documentation copy especially is typical of the ways LLMs leak implementation instructions into their written copy.

Personally, "AI wrote a driver" and "AI rebottled an existing solution and sold it as Dr. Claude's Miracle Driver" are semantically very different claims to be making in the public square. If you're only making it to use by yourself sure they might behave similarly, but when you're doing it for a public audience people will want to know if you're fibbing with your verbiage.

Being signed into Google doesn't really figure into Cloudflare's systems at all, haven't for years. They replaced ReCaptcha ages ago.

The code for expert systems and especially decision trees can be audited. How do you audit a language model's decisions when it could have had hidden trigger phrases that subtly influence its reasoning baked into it?

Because they use a commercial service (AT Internet) for analytics and visitor tracking as stated in their cookies policy and cookie inventory. This information is readily available through the cookie banner.

If it was overtrained on academic papers it'd reiterate the point multiple times for structure. Instead, it's burying the lede seemingly just to pad.

The stigma surrounding AI use will get worse if... People know AI is being used? Sorry how does maintaining heightened paranoia by omitting clear signals help exactly?

The figures used don't count browser users at all, signed in or not

Presumably it'll either mean selling "agent seats", or billing seats for every started hour of use. With a monthly minimum, of course.

Would you be willing to do free work for a corporate entity that explicitly financializes that work's benefits? Would you be willing to do free work for a corporate entity whose entire business model is making sure they sit as a gatekeeper between your free work and others who would benefit from your work?

For some it's demoralizing to know that your work will be broken down and atomized into language model mush, and the credit will go to the computer.

The downside, as stated in the message, is implicitly supporting the LLM data ingestation pipeline by providing fresh content. It's not a direct harm in itself, but feels very tragedy of the commonsy

Generally when I read a literal genie story, the message isn't 'well it was reasonable of the genie to do this'. It's more commonly either a morality tale of the person being wrong to ask for whatever it was they asked for, or just a "wouldn't it be fucked up if the genie did that huh"

The user-facing string stuff is definitely the thing that drives me the most nuts. I guess it's nice that it defaults to gag-inducingly transparent, but I don't think the backend structure should be written out in the title element of the app.

Corporate posting in the weeks immediately surrounding a legal case or ruling impacting tech companies is definitely my favorite brand of written text. Despite not paying full attention to EU legislation, I get reliably notified by Google about every tiny bit of antitrust from them posting a "Google brings millions into the EU's [field they are getting hit with antitrust in] ecosystem" vagueblog about it.

Agent forwarding would be pretty big, if it for example pushed your git credentials to the remote host.

LLMs very much operate on the level of text. That text can represent ideas, but it can also directly be a copyrighted work.

I am however very interested in the novel interpretation of copyright that says that you can do whatever as long as your compression is lossy.