All that extra is clear as day compared to the mystery of how neural network training decides to divide and balance the weights in even small neutral networks.
We can, at best, approach a good set of weights, even in tiny neural networks.
Imagine if we found a way to calculate the exact optimal weights for a given loss function. I mean, there is an exact optimal solution, it exists, but we can't find it exactly, even for a neural network with just 50 parameters.
There is no point in that because the loss function itself is already an approximation. No one knows what is the exact loss function for any given non-trivial real-world task.
I mean, things humans defined can be pretty clear. Like your electricity rate. Natural systems less so. Not pretending no complexity in human made things, but at least some models can be fully specified.
I mean, the subtlety of the neural network weights that emerge from training are not fully comprehended by anyone, man or machine.
Every individual calculation is understood, and every step of training is understood, but the exact nature of those weights that divide the responsibility of responding to subtle changes of input in intelligent ways is beyond me.
What are the implications of AI output not being copyrightable?
Can I just grab a copy of my company's vibe coded app and start selling it myself? What stops me? Not copyright law.
I suppose what really stops me is that power wins, and the company has more money and power than me, and so regardless of whatever the law says, my life will be fucked up good if I did copy their codebase.
We need a way to interleave our own notes into the source code.
It's like Bible study, everyone studies the same Bible, but everyone wants to have their own highlights and notes.
Some might say to use comments, but my Monday morning chain-of-thought doesn't belong in a comment.
I want a way to do literate coding. I want to interleave writings, and drawings, and editable code snippets that point to actual code in the code base -- and I want to see when those code snippets are stale, or have been edited by other developers (or agents).
I want a way to build up some artifact of my understanding that I control, that exists primarily to aid my understanding, and isn't constantly shifting as a thousand LLMs pass through the actual codebase changing everything.
The code used to be the canvas of the programmer. Now programmers have no canvas. They need a canvas. They can't understand things without a canvas to hold their thoughts.
The irony is, this would be a pretty trivial one shot project for a modern LLM to create that interfaces with the source control repository of your choice. It's basically an advanced note-taking app, which LLMs have been trained to death on. I bet you could get it to write it as a VS Code extension as well (or emacs macro, or... etc.)
What is an "algorithm"? Showing the 20 most recent posts is an algorithm, but I doubt that's what they're targeting with the law.
I'm not just trying to find a technical gotcha here. I think it's worth identifying what exactly about an "algorithm" causes it to cross the line and become harmful.
When does an algorithm become harmful?
When it is designed and optimized to maximize the amount of time it engages the user?
What about an algorithm that was optimized the present the user with content they would find interesting? That doesn't sound so bad on the surface.
What about an algorithm that maximizes the amount of ads people see? Or the amount of money that the company makes?
Most Australians dislike social media. We are underrepresented on it, and it shovels the US’s issues down our throat. The current algorithms mainly send us US political propaganda and things that are supposed to make us hate nearby Asian powers like China.
> Social media is not capable of shoveling anything down anyones throat.
This statement is manifestly untrue. The malign power of social media is precisely the reason that legislation has been introduced in Australia and elsewhere, and that Meta et al are battling and losing "addiction" lawsuits worldwide.
> What about an algorithm that was optimized the present the user with content they would find interesting? That doesn't sound so bad on the surface.
If only we had 20+ years of scientific and societal evidence to show that regardless of what you think is "on the surface", the effects are catastrophic to mental health, democracy, the concept of truth, and everything else.
I'm sure the final law will outline the specific details. But the intent of the law is not algorithms in a computer science definition, but modern recommendation engines.
Showing a chronological feed of accounts you have followed in this case would be non algorithmic. This is how all social media worked before the modern hellscape we are in.
To you and me it's a very wide category, and there may be hundreds or thousands or even millions involved in me simply typing this, it being saved somewhere and then served back to you.
For lay-people though, it's become a term which means "The thing which measures your preferences and usage patterns then decides what to show you". So for lay-people to want to turn off the algorithm, they tend to mean going back to "just what your friends posted, in order"
> When does an algorithm become harmful?
I think this is a research subject. AFAICT it's pretty certain now that social media 'algorithms' have been optimised for addictiveness, which is likely to overlap somewhat with what users find interesting, but also cover things they find mildly titillating and/or outrageous.
The harms come in a bunch of forms. The never-ending scrolling of bite-sized videos seems massively destructive to attention span, and people seem to get addicted to it and may neglect other parts of their life. The platforms also seem really keen to send men and boys off to pretty hateful 'manosphere' content which is apparently making life hell for women (including teachers of teenage boys). Then we have all the body-image and anxiety issues caused in teenagers constantly exposed to influencers showing off their (faked) wealth, beauty etc.
Switching off the "shoving this all in your face" part of social media seems to me to be a win. When it was actually a way to keep in touch with friends and family it was a different game.
> The thing which measures your preferences and usage patterns then decides what to show you
That (plus a clause that exempts algorithms with user controls with full informedness and ability to confidently and reliably predict the outcome = old good filters and such) is the thing that should be in the legislation’s definitions section.
But then in the actual law it’s worse than “algorithm”, they’re actually banning broad over generalizations and design patterns instead, stuff like this:
>> recommender features – content selected and displayed based on information associated with the user’s account;
(Obvious legit use cases are regional content and user preferences.)
>> endless-feed features – continuous feeds of end-user content;
(Pure design pattern, pagination can be addictive at any span size.)
Dark patterns are called that for a reason. Having to click "next page" is a point of friction that slows and stops consumption. Pageless increases engagement in an addictive way - it's why companies do it.
I disagree that any of this is "overly broad" - it's all long overdue
I respectfully disagree. I’ve clicked way too many “next page”s (or refreshed pages, or whatever) to be aware that while the friction is real, it’s not the actually addictive component.
Endless scrolling aids, but isn’t what makes a system addictive. Continuous relevance gamble “maybe the next one is good, or the one after” is the actual driver, and “I stop at this page” is a weak stopper condition. Not negligible (as proven by testing) but not the fundamental piece.
Actually, I found an experiment that directly contradicts your statement. So my disagreement is IMHO valid and your appeal to "mountains of telemetry" is not. To me it seems that you've just parroted the chorus without trying to understand the problem by decomposing it into pieces and trying to attribute the potential causes. Which just happens to be something that I spent some time doing (total amateur hour and N=1 but sort of worked for my purposes).
Etsy ran A/B tests with and without endless scrolling. Same content source. Infinite scroll haven't produced any increased engagement: surprisingly, it decreased it - people searched less. Potential reason: it was a store, not a social media feed, the feed wasn't addictive in its nature (decreased relevancy, lower chances of finding something good buried down further and further) and the whole alternative pagination design didn't help at all. This is my interpretation, but c'mon - if you're appealing to the telemetry: here's the telemetry, how'd you explain it?
In the neighboring comment, I was mistaken in that it would still increase length of browsing a bit, turns out it was off-putting.
With addictive content, endless scroll makes it more addictive and increases consumption.
I'm not sure why you think this is a slam dunk. Fruity flavors increase the adoption of vaping because it tastes better, but that doesn't work when you apply it to broccoli. Doesn't mean we shouldn't regulate flavored vapes.
Yes, exactly, with addictive content it boosts the engagement even further. We all agree about that. And - my point here I’m trying to convey - with nonaddictive content it doesn’t do that. Which is why I’m saying “we need to classify better”. I don’t understand why you’re against a more nuanced approach, when the goal is to allow non-harmful cases. I’m very much favor of regulation, just a more careful one.
Vaping analogy is not perfect, as it doesn’t hold true in some aspects, but... Nobody makes vapes that go unpleasantly sour for a while after a fourth puff. But if they do we’d better not ban them from adding flavors, when the flavoring here does the right thing.
We’d still want to regulate vaping on other aspects (such as harm to health, which - unlike speech - doesn’t get protections), but that’s another, different story.
(Sigh) And what exactly does it show? What makes you think you interpret the statistics correctly?
It shows a correlation that suggests - at best - that certain design patterns decrease friction and increases duration of engagement, making system more captive. It does NOT show that those design patterns are captive (= are a problem) by themselves.
Have those mountains of telemetry ever included an experiment where the addictive component was intentionally not present or maybe even designed to be anti-addictive (next element becomes increasingly less likely to be interesting/relevant past some threshold, frustrating the viewer) but the endless scrolling remains, so we test on its presence alone? Ideally, with an engagement-optimized feed as a separate control.
I don't know of such experiments, so I cannot counter your point with a definitive proof. But your interpretation isn't a proof either - there are multiple factors and no control to tell them apart. Those mountains of telemetry didn't have a goal of isolating factors to determine their actual effect, they just had to optimize engagement, find the best combo that works.
If someone ran something like that - I'd love to hear about it, and what they show. I'm utterly confident (as someone who had dealt with various addictions) that it'll show endless scrolling would have mild effects at best (maybe just a bit longer engagement compared to classic pagination as telemetry suggests, but I expect it to pale in comparison to an actual addictive/engagement-increasing feed source logic).
Your claims that these specific features provide no harm are belied by our knowledge that they do.
If "recommender" algorithms were the same as "just show me what my friends posted in chronological order", then there would not be such resistance to implementing the latter.
If endless scrolling were the same as pagination, then apps optimized for addictiveness would show a roughly equal distribution of both. But they do not.
Recommender algorithms as defined aren’t just the recommender algorithms you’re thinking about. It’s merely this idea that precise definition matter, nothing more. I’m not arguing that it was a bad idea to restrict some crappy patterns, or anything like that. Just that lawmakers defined that poorly.
And endless scrolling is pagination with a page size of one and a gesture (scroll down or swipe) for the next page. And trust me, “next page” can be helluva addictive. Endless pagination is an aesthetic preference. It may make it more slicker, frictionless, more visually appealing to get addicted, sure, but it doesn’t change a thing about the core element of the addiction. Same endless feed can be easily made counter-adductive by degrading relevance past some threshold, which clearly tells me the issue is elsewhere.
About the infinite scroll: I just found not just a hypothetical counterexample, but an actual experiment that suggests that infinite scroll is not addictive on its own. Not the ideal experiment, but it proves the key part of my point. So now I have not just logical inference, but also a real-world empirical confirmation.
Excuse me… what? You must've misunderstood what I argued about. Yes, they’re not the same!
I'm going to reiterate for the third time, to break it down even further in a hope it clicks.
We have:
1. Predatory/addictive algorithms (A_bad). We don’t like those.
2. Old good user preferences-based algorithms (A_good). Those are fine.
3. Probably some other algorithms but let’s go with just two alternatives for simplicity’s sake, as extra (A_more) won’t affect the argument here.
Definition from the law (A_law) covers both (A_bad)+(A_good). As I already said twice, it’s overbroad.
You said “if (A_law) wouldn’t be bad there wouldn’t be an issue”. That is obviously true, because (A_law) contains (A_bad). What you said is true, and it doesn’t contradict what I said - that (A_law) containing (A_good) in addition to (A_bad) isn’t great and lawmakers should’ve done better.
Just so we’re clear, I’m not arguing that A_law = A_good (that’s not true, they’re not the same! but neither A_law = A_bad either, it's the fact that it's A+bad+A_good[+A_more] is the issue), or that A_bad isn’t bad, or that A_bad isn’t covered by the law, or anything like that. Your knowledge that A_bad is bad, is my knowledge too.
As for the second point, I provided a counterexample of how endless scroll can be a non-issue. If endless scroll can be a problem (no argument against that from me - yes, there are clear examples we all have seen) and endless scroll can be non-problematic too (my counterexample design), that, says endless scroll is not what’s problematic per se. That’s just simple propositional logic.
Endless scrolling can be a proxy/substitute for the actual issue if nothing better could be named, but I think a better target can be identified and pointed out.
In other words, all I’m unhappy about is that laws are poorly worded and loosely defined so they can potentially cause unnecessary collateral damage. All I’m saying is that lawmakers are missing the point and just vibe-legislate in the vicinity. They’re banning the thing they (and you, and I) see, but the way they do it kinda like how Plato defined what a man is. And I’m just plucking a chicken here.
If you see a logical flaw - please do point it out for me. I don’t see any.
It’s a good question, and it’s not well defined. I am hoping that it allows us to customise feeds more. If I have problems with body image, don’t show me topless photos. Same goes with alcohol and gambling, which some platforms do have. The answer must not be either be subjected to content you don’t want to see or don’t use social media at all.
I think the line for algorithms is anything machine learning or neural network based, not simple ordering.
Side note: When iterating on male avatars, ChatGPT incrementally increases muscular features without being prompted to do so. There are biases in these machine-learned algorithms.
White-hat and grey-hat hackers need to be able to perform penetration testing without permission. Nobody is able to build secure systems. The best we can hope for is that the good guys find the vulnerabilities first and report them responsibly.
This would be a huge inconvenience for companies and government organizations, so it probably won't happen. We will chose to sacrifice national security for the convenience of companies--what else is new?
Companies will say "it is our system, we are responsible for our own system", then, after a breach, they will say "our bad, we are not responsible". Same old story; half the nation's personal information is leaked twice a month and nobody cares.
Since I'm getting some positive feedback, I'll go even further and say that there should be security bounties established by law:
If a certified red-team of security researches breaches a company's system and discloses appropriately, the law should require the company to pay a security bounty.
The bounty doesn't have to be crippling to the company, but it should be large enough that the security researchers will be paid well and can live on collecting security bounties. We want an entire industry of good guys testing the security of everything.
There can be some regulation to. Like, it's not okay to run a massive DDoS to test systems. We want the red-teams doing constructive things, not just breaking everything. It should be legal for the red-teams to be annoying, but not purposely destructive.
While I agree in principle that sounds like incentives leading to a new economy similar to that of copyright and patent trolls. Just scan for known vulnerabilities in [commonly used software] used by [mom and pop shop] and demand a payout. The result is a dead web consisting only of big-corp-hosted sites as no one else can ever keep up with the infinite array of vulnerabilities.
In general, preventing consumers from having fully featured AIs (especially local ones) will be worth about 10 trillion dollars to companies. That's my prediction.
Amazon is one example. If my basement AI drives my browser and never clicks an ad, Amazon loses big.
Or, the ad industry in general -- imagine the basement AI filtering out all ads at the analog level so nobody in my house ever sees an ad, and no software update can change that.
Or other optimizations. Right now companies enjoy a power imbalance because they have people who ruthlessly optimize 24/7 to save a buck. I can't do the same, because I have to work, and also prefer to spend my personal time with family -- a foolish thing to do in a capitalist society, at least from a capital perspective. Imagine I have an AI wheeling and dealing on my behalf 24/7 too. Especially in an age of online shopping and product deliveries, my AI will absolutely order from your competitor to save a penny, my AI will absolutely spend 400 hours disputing a medical bill I don't think is right, barely an inconvenience to me.
One of the optimistic aspects of what's happening is that it doesn't seem like the leading model capabilities are outpacing the more open ones by that much- that just as a general principal there are pretty ok models relatively close behind the best ones.
This could mean the long term viability of my open claw bot to shop for me- that openai / google etc won't control AI, and I'll still just be able to use a model via API to do what's actually in my own self-interest.
The ads (in general) will be redirected to where you will see it, regardless of whether or not Amazon allows bots. You don't think you're getting off that easily, do you?
Once AI is good enough the AI will control what I see, plus, I will just spend a lot less time looking at screens altogether.
It's also trivial for Amazon competitors to pop up in this potential AI future. Maybe you only sell door knobs, but all you need to do is set up a mostly text-based API designed for AI consumption and you're in business. If you offer the door knob for 10 cents less, why wouldn't the AI agents buy from you?
Right now making an Amazon competitor will fail mainly because humans are lazy and can't be bother to check other sites. AI might change that.
"Look at how great our product is!"
reply