malwrar

It’s kind of funny reading this blog announcement as a corporate product announcement, because I’ve been actually playing with nreal’s air 2 ultra glasses. They are absolutely perfect for open XR work, I love them. You get the cameras, buttons, accelerometer, all over the cable without fuss, and they show up in Linux as just a regular DRM device. Literally by default most window managers just see them as another display. It was such a nice experience getting a 3d engine hooked up to them and adding 6DoF tracking working with visua slam (two fisheye cameras gives you instant depth, a luxury compared to the usual monocular focus). Honestly these glasses are good enough that I use them with a pinephone as my primary computing device these days. And this was on the side executed by an agentic harness while I focusing on other things! I guess I’m also now competing with Google, and so could you too.

Predicting it now that people will vibecode open source clones built from having bots play the closed game for training data.

Countdown to Zero Day probably! Super good book, technical without being unfriendly to casuals or inaccurate for pros.

I think generative content will get good enough that people simply don’t trust what they read online. Humans might still generate content, but I predict over time they will be crowded out and worse blatantly copied to the point of discouragement. AI-powered miscreants will outpace enforcement efforts.

Personally I’m betting that we will see a major comeback in local radio and print media, as global network effects regress local with collective trust.

> But what got those engineers in the same room in the first place?

The interest in a shared problem and the passion to solve it?

> A patent is the best thing we’ve collectively come up with

I think that’s a bold claim in the face of the numerous contributions from e.g. universities, open source communities, single-person projects, etc that have pushed the world forward. We simply do not know a world free of the oppression of patents, the institution sustains itself.

> If you don’t like the game, get better at it.

I don’t think people will be playing this game much longer in the face of generative AI, unless ofc the ruling classes succeed in locking down access to that tech, gobally.

How can price discovery work if everyone’s price is different? What downward pressure would even exist on surveillance-derived prices besides resale black markets? Societally, do we really want to further monetize the expansion of the mass surveillance state?

> In July, Meta announced that it would push an update to detect when an LED has been “physically tampered with or destroyed” and stop people from being able to film

Would love to know how they plan to do this, sounds like a patch that is bypassed with a $0.01 resistor.

Hard to find better solutions when we have no agency to enact them. The “arrangement” was one-sided from the start.

Any tips on measuring the impact of not having swap? I never use it on my machines for similar reasons.

I feel a similar fate will befall engineers too. Your predictions anre quite interesting from that perspective.

Markets defined entirely by law have distorted our collective understanding of what can actually be built with the knowledge our species has accumulated thus far. How will traditional shields that have protected capital accumulation in tech to survive in a world where governments now realize control of technology is a national priority? Especially as we see its impact on modern warfare, and that such conflict looks like it’s only escalating over time.

Mathematicians appear to me (as an outsider) to exist in a field without such distortions, and I think offer engineers a preview of what’s to come. I certainly have completely ceased sharing original ideas online at this point.

What about when the open models get to the point that you don’t need these fancy GPUs? It seems inevitable this will happen over time, we see cracks right now within even just GPT w/ expert streaming and flash attention. ASICs are coming soon too.

Is the fact that Nvidia is also massively financing the closed-models companies, which stand to lose when open model project(s) take off, a conflict of interest here? I can’t tell if this is nvidia hedging or planning to embrace & extinguish.

Don’t get me wrong, governments putting these things up for one purpose and then using them for another (my read on your use of “malicious”) is a major concern I have.

It is broader than that though, and is the distinction I sought to capture with “anyone with access” in my original post. They are connected to the internet, and to a cloud no less! Anyone could be the NSA now, from anywhere in the world. I bet we’ll see these used to collect marketing data too.

Perhaps a limiting principle can be that we simply do not use ALPRs or other surveillance tech, or at least that we don’t connect them to the internet and require a warrant to access and use their data? Personally, I’d just prefer cops enforce traffic law the old fashioned way.

Thanks for this term, I’ll try using it!

Interesting perspective.

As others have noted, I believe the brightline is less “general purpose computer…” and more “…that is remotely accessible & modifiable effectively anywhere in the world”. I worry not just about OEMs and cops, but also criminals, foreign intelligence agencies, etc. Honestly potentially anyone at this point, now that folks have figured out you can use LLMs to do analysis.

This is a sound foundation for concern by itself, don’t you think?

I don’t doubt it was intended by the manufacturer, but do you think the average customer knows this? Every normal person I’ve ever told this to, that their Ring cameras can also be looked through by the police, has looked at me like I’m crazy. I’m sure some people would buy a Flock camera for their porch, but I’d bet many others wouldn’t if they truly knew what they were signing up for.

The Super Bowl commercial that Ring put out about finding lost dogs spooked people for a reason!

It frustrates me when people call them license plate readers. I guess you need to call them something, but they are general-purpose internet connected cameras. They will do whatever their firmware tells them to do, and could be reprogrammed at any time by anyone with access. No one expected doorbell cameras to join a mass surveillance network, but later the manufacturers added that feature. Why do we treat these cameras like they can do only one thing?

We are only ever one computer science paper or github repo away from all of their moat fading to dust anyways. The question everyone financial should be asking is “what happens when folks don’t _need_ to pay these people anymore?”

The sizes for the matrices are decided up front (“hyperparameters”, more or less guesstimated) and the values that are learned (“parameters”, for completeness) and remain static after training. The layering just refers to the core “transformer block” being chained inside the model that contains most of the weights. Those themselves are pretty simple, usually just an MLP followed by an attention function, with variations on formulation & “multi-heading” in most of the new models. Behind the terms and math and inner architectural choices, it’s pretty much still the same GPT pattern of ‘“lie” = decode(sample(lm_head(T_1(T_2(…T_n(encode(“the cake is a”))))))’, with T_i(…) being the ith layer.

You know, I haven’t kept up with MoE and etc where there’s a bit of selection going on, so I should probably be more humble. I think new work has only added different “paths down the same hill” though (no recurrence, just select different matrices), but could be wrong there. I don’t think I’m wrong on my general intuition, just want to be epistemically honest!

Yes.

Each token prediction is one big function call. Then you just recursively generate more tokens until run out of context or the model predicts a next token indicating end of sequence. Technically the model outputs a matrix where the last row is a probability distribution, but I’m counting sampling from it as part of the chain. Hundreds of billions of dollars has gone into just making the function fatter and gradually changing pieces here and there.

They log all DNS requests made to their public resolver in a searchable internal database, at least when I worked there a decade or so ago. I wonder if they seed their crawler with it?

One of my personal faves on that site: fauux.neocities.org

> We really need an open protocol to build on.

I’d bet making a slack-compatible client or bridge isn’t hard, we all just instinctively know whoever develops it is going to get sued or taken down.

It feels like we quietly gave up on adversarial interoperability awhile ago, and act like we need a whole separate “open walled garden” when what we actually need are legal protections that prevent companies from suing/banning people who call their APIs. Slack, Facebook, etc, are walled gardens only because they can ban/sue people who compete with their client experience.

I figure that will probably never happen in the US (maybe if someone rich starts it), but eventually someone outside of it will make such an adversarial integration and host it from some region that doesn’t care about US laws. Then, when they get away with it, we’ll all praise them as a genius and wonder how Slack could exist at all. The US has many international agreements keeping this illusion alive, but my guess is that even formerly stable markets like Europe could spawn such work if they decide to stop caring about ~1990s-2010s era contempt-of-business-model US laws.

Perhaps “censorship” & manufacturing consent?

I think both political extremes have their own angles: liberals might be concerned that conservative censors will censor kids from learning about LGBT people and minorities, conservatives will be concerned that liberals will force too much LGBT and minority content onto kids. Or whatever issue, they want to control what your kids read!

This will almost certainly be used to censor adults too, the only reason we aren’t doing that is because it hasn’t been possible to consistently identify people before. Considering who is pushing for this, they’re absolutely going to tie this into advertising, and if they know who you are so do all of the spooky upper echelons who could implement a true censorship regime.

“The only way they can do this is by controlling what you read, shouldn’t that be the parent’s choice?”

My proposal is that we increase the effectiveness of child controls on devices, so kids can’t get access to the adult world on them. Parents in other threads however tell me that this won’t work, they’ll just use the internet on their friends devices, and what we need is to completely exclude children from the adult internet by confirming the identity of every single person who wants to use it.

Now I’m trying to convince people that identifying adults won’t work either, because kids will still get around it. All the while, we adults live with the negative externality of no longer having anonymous communication. Woe betide us all should this come to pass, but I guess desperate people are how bad things like this tend to happen.

I heard a billionaire is funding this, he’s probably afraid of the one famous instance of someone like him getting killed by one of those.

My point is that the entire check is bypassed easily and instantly, and in the meantime the government gets data that someone _will_ figure out how to make personally identifying for adults, or will argue for changes to make it so. Alcohol age limits are a simple physical check for a vice that everyone accepts those who want it can get at. I’d rather demand that device manufacturers give parents effective controls before we try solving this problem by identifying internet users wholesale.

Headline news: children infiltrate the universal adult one time password scheme for porn, parents panic! Turns out the 18 year olds started selling access to their younger friends, who resold it to their younger friends.

These things passing the Turing Test makes anthropomorphizing their behavior awkward, but don’t forget it’s just an analogy to communicate an experience. If you convey a certain written voice to these models in your input, you get a somewhat consistent end effect. I think that’s all that is being communicated.

It was precisely that for me! Another commenter captures it well; “the bitter lesson” indeed.