Notices by Mark T. Tomczak (mark@mastodon.fixermark.com)

@lauren part of me wants to observe that the real question is whether AI transcription makes fewer errors than human transcription.

... But I think the question of category of error is also very important. If the algorithm is extrapolating full-on made-up statements then I think that might be a kind of flaw that we can't really accept for this use case.

@lauren Yeah, I agree. There's definitely a serious conversation to be had about the possibility that the failure mode when this machine fails is worse enough that one shouldn't try an apples to apples comparison between machine and human transcription accuracy.

Cards on the table / where I'm coming from in all of this: I don't exactly advertise it but I'm working on autonomous vehicles these days. So the safety story is omnipresent on my mind, and there has definitely been one horror story I'm aware of where the vehicle failed to realize there was a pedestrian under it and kept going.

... Unfortunately, I hear that story and I weigh it against the stories I'm aware of where a human driver, absolutely panicked about facing consequences (or ignorance they had struck someone, we know that humans don't have as much situational awareness about what's going on outside their car as it feels like we do when we're driving one), made the worst possible decisions for the pedestrian they had injured. So I feel a lot more comfortable doing apples to apples comparisons of autonomous versus human operated vehicle safety.

I might be guilty of assuming things about a field I don't know when I don't cut medical transcription a similar benefit of the doubt to the benefit I cut my own field.

@lauren I've been in about a half dozen since they aren't deployed in my city yet.

Generally felt pretty good. I don't feel any less safe in one than I feel being driven by an Uber driver at this point.

Coincidentally, my last ride home from the airport I was driven by an Uber driver and a Tesla and he switched it to FSD from entering the highway to my door. I did not realize that it could do full on highway to off highway grid level road yet. Driver is fully responsible for what the vehicle is doing and has a HUD indicating what the vehicle sees, they are supposed to be matching that continuously with what's going on out the window. This does give me some UI questions... I know that we have research from the '90s about having to take your eyes off the road, so I suspect that that HUD would be better served if it were projected on the glass.

@lauren

@lauren

And then what happens?

Generally, the cops show up since we're taking about devices with camera coronae and always on cell connections. If someone starts a fight with an AV, there's ample legal evidence of what happened.

You seem to be envisioning a "riots in the street" scenario that I find unlikely (not impossible, but unlikely).

Trapped on the side of the road

.... I don't generally worry about suddenly dying on the side of the road when my car stops. If I did I wouldn't drive for fear of flat tires and battery failure; AAA has taken two hours to find me in the cold before when my regular human-driven Hyundai failed to continue being a car. Do we have reason to believe the always-online camera-studded robotaxi has more risk of that failure mode?

@rantingsteve

@mirabilos @dalias

Do you know how computers work?

Gosh I hope so, or this computer science degree and following career has been an elaborate fabrication. ;)

I think you might be flipping the logic arrow on this problem. You seem to be implying that it's created by a deterministic algorithm and therefore it can't be transformative. Unfortunately, deterministic algorithms can be applied to vast, vast quantities of copyrightable material. Proven copyrightable material, like this isn't even up for debate. That's the universality of computing.

So clearly, you can't reduce the question to "It's just a deterministic algorithm." That's too broad a constraint.

The questions that the law addresses are the ultimate purpose of the work, whether the changes enacted upon it make it an entirely different kind of thing. And I think the law argues, and I would agree with it, that using the content of the New York Times to train a large language model (along with content from other sources) is self-evidently transformative. I can't pick up a New York Times and shake it and have a novel New York Times shaped article just fall out of it. Clearly, feeding the content through a training algorithm is doing something significant here.

Plus, at the end of the day the purpose of copyright in the United States is the promotion of science and the useful arts. That's the lens through which all of this is viewed. And the courts broadly seem to be coming down on the side of LLM-AI being exactly the kind of science that copyright and patent are intended to promote.

@dalias @mirabilos @davidculley This is one of the claims that the New York times made that I find most interesting.

They are claiming that output from AI generated on their content replaces their product?

I don't see how that can be the case. Their product is reporting on current events. There is no way for an LLM to do that by the nature of what an LLM is.

I'm sure if I can carve out the time to read their complaint in detail it will be more clear. To a cursory reading in the Reason article, that claim makes no sense, but the case passed basic scrutiny so I have to assume that summary compresses the complaint too far.

@mirabilos @dalias @davidculley So I might be wrong, I am not in fact a copyright lawyer... I don't think that determinism is the thing that determines whether something is transformative.

I wrote a blog post on this a while ago. There's an algorithm I can plug into blender that will generate an infinite number of asteroids by iterating on a relatively simple mathematical description.

While I might in fact be wrong, I am fairly certain that the output of such a generator is transformative enough to be copyrighted. If it's not, there's an awful lot of art out there that probably cannot withstand copyright scrutiny, since its creation is nothing more than a sequence of deterministic transformers applied to some root seed...

https://blog.fixermark.com/posts/2022/on-ai-art/

@dalias I mean, Independent of whether we donate to them they are correct, both in terms of the law and in terms of what the law is for.

We can dislike it (though I would recommend we sit hard with precisely why we dislike it and consider all implications... Most ways to take power from AI companies via copyright put it right in the hands of the Disney corporation). But their understanding is correct as far as I understand.

Thinking about the era of technology where the easiest way to display images was to take all the air out of a tube and point a death ray at your face.

It's fine though. There was a thin screen between you and the death ray that caught all the death and turned it into pretty colors.

@sparrows @mhoye This is another extremely good point. Sometimes people use the system in ways other than we anticipate because they have people in their lives that are inspecting the ways it's normally used.

@mhoye "But why is she keeping something mission critical to her health in her unpublished tiktok reels" fuck you is why. We don't actually get to decide how users choose to use the system, we get to provide it, provide soft incentives to use it the way we intended, and keep it running when they get creative.

Now, is she going to be able to show the doctor what she needs to show them or is it going to be on you that she can't?

The UK is trying to ban nudes (store, transmit, display) on smartphones owned or operated by kids under 18.

Is young people seeing nudes bad? Separate problem, and could be its own thread, but for the sake of argument let's say "yes."

Is the problem tractable? I think the answer is actually "probably" for a partial solution. It'll be expensive (Apple and Google's money and everybody's electricity and, of course, privacy), but I'm playing Evil Oracle in my head and I'm not seeing a defeat strategy for "cliient-side hash check routed to server for detailed confirmation and revocation of account on offense."

... But blocking seeing them at all? Intractably expensive in energy and time for anything approaching a 100% case when bright and motivated young people are in the loop.

@lauren I believe a bloom filter (or trained recognizer) on the video layer will generally be hard to subvert; Lord knows they'll try. That gives a rough cut of "probably porn," and the whine screen can then be passed to an online detector for verification and then follow up (notify parents, ban account, whatever). It's impractical to use that to prevent the image from being seen though, unless we want the kid phone experience to be total ass.

It would also be very expensive battery wide, depending on how frequently it's scanning the whole screen. And "your phone can screenshot what you are up to at any time and send it to the government" feels like it should be pretty viscerally horrifying, even to Brits.

(I don't think this is a good problem to try to solve; my noodling is on whether it's actually unsolvable and under what constraints it becomes so).

@haicen @soatok

@thomasfuchs This is true, even if their incentive to do so is market-driven (they're a hardware company; data mining their users' PII beyond things like "tracking app store purchases" just doesn't have benefit outweighing risk).

@sj I've always considered it a bit of a litmus test. If we're really getting the kind of freedom we wanted out of all of this open tech, there should probably be a couple of furries or anime mascots in it.

It's a little shibboleth I know, but... Seems to work? If you find yourself in an area completely devoid of anime mascots and furries, be ye warned... Ye may have wandered deep into corporate computing. ;)

@sundogplanets Even if the engineering is feasible, it's a plumb-stupid idea.

"You mean I can push my compute into a place where my only cooling solution is black-body radiation, and I get to deal with the computational errors introduced by unshielded cosmic rays... But if do both of those things, I get... a fraction of the compute power of an equivalent-sized terrestrial system at 100,000 times the up-front cost? Sign me the fuck up!"

Even Scott Manley, notorious (and delightful, this is not criticism) Liker Of Technically Complicated Engineering Things In Space, could only see one justification for this: it's only the cheapest solution if the kind of datacenter you want to build has been banned on Earth. They're hedging against the kind of mass uprising that, I hate to say, makes it risky to assume you can get something as vulnerable as a rocketship off the ground if people hate your product that much.

(And I think he was being generous. I can think of other reasons one might want to do this, but expounding on them would have gotten me banned on Facebook and Twitter and I choose to be more polite on Mastodon. 😉 )

@EuphoriaLavender @thomasfuchs On a lark, I tried throwing a locally-running Qwen at Common LISP using the CLSQL library.

It had no idea the API for the library and did not give me runnable code. But what was fascinating was it did give me syntactically-valid LISP (just trying to call nonexistent functions), and the shape of it matched the shape of the CLSQL API---function names were wrong, but arguments and even interrelationships like "make a connection and then use it to execute SQL" were mostly right.

... which suggests to me that at a fundamental level, the structure of SQL code is just a common pattern, so common that it could be extrapolated across language and library boundaries. And that means making code that talks to a backend via SQL should be automatable.

@thomasfuchs I basically concur. I should have saved the link to it, but someone did a blog post awhile back that was basically "LLMs work well on your code because your code is shit." I have observed that, notably, they struggle with common LISP (although that may also be a consequence of the training dataset).

But, I would extrapolate to observing that most code is shit because it doesn't actually pay to write deeply concise code. There has always been a tradeoff between "getting it done today" and "getting it done perfectly," and the people who want the machine to do the thing want today. In fact, if you don't know your problem domain perfectly, I'd argue that trying to make your code optimally concise is counterproductive.

For those reasons, we can expect LLMs to be a time-saver to the extent that they can execute on "Take this fuzzy pattern and apply it to the codebase" and I expect they will end up a permanent tool in the toolbox (though not in their current form; a whole datacenter to do a 'soft-grep' is overkill, my prediction is that the open source projects will succeed in condensing the tool down into "works 90% of the time on the most popular languages and fits on one or two graphics cards").