Honestly I should just get that slide tattooed to my forehead next to a QR code to Weizenbaum’s book. It’d save me a lot of talking!

I agree with you so strongly that I went ahead and updated my comment. The problem is general and out of control. Orwell said it best: “Journalism is printing something that someone does not want printed. Everything else is public relations.”

These articles frustrate the shit out of me. They accept both the company’s own framing and its selectively-released data at face value. If you get to pick your own framing and selectively release the data that suits you, you can justify anything.

I am once again begging journalists to be more critical of tech companies.

But as this happens, it’s crucial to keep the denominator in mind. Since 2020, Waymo has reported roughly 60 crashes serious enough to trigger an airbag or cause an injury. But those crashes occurred over more than 50 million miles of driverless operations. If you randomly selected 50 million miles of human driving—that’s roughly 70 lifetimes behind the wheel—you would likely see far more serious crashes than Waymo has experienced to date.

[…] Waymo knows exactly how many times its vehicles have crashed. What’s tricky is figuring out the appropriate human baseline, since human drivers don’t necessarily report every crash. Waymo has tried to address this by estimating human crash rates in its two biggest markets—Phoenix and San Francisco. Waymo’s analysis focused on the 44 million miles Waymo had driven in these cities through December, ignoring its smaller operations in Los Angeles and Austin.

This is the wrong comparison. These are taxis, which means they’re driving taxi miles. They should be compared to taxis, not normal people who drive almost exclusively during their commutes (which is probably the most dangerous time to drive since it’s precisely when they’re all driving).

We also need to know how often Waymo intervenes in the supposedly autonomous operations. The latest we have from this, which was leaked a while back, is that Cruise (different company) cars are actually less autonomous than taxis, and require >1 employee per car.

edit: The leaked data on human interventions was from Cruise, not Waymo. I’m open to self-driving cars being safer than humans, but I don’t believe a fucking word from tech companies until there’s been an independent audit with full access to their facilities and data. So long as we rely on Waymo’s own publishing without knowing how the sausage is made, they can spin their data however they want.

edit2: Updated to say that ournalists should be more critical in general, not just about tech companies.

No need to apologize for length with me basically ever!

I was thinking how you did it in the second paragraph, but even more stripped down. The algorithm has N content buckets to choose from, then, once it chooses, the success is how much of the video the user watched. Users have the choice to only keep watching or log off for simplicity. For small N, I think that @kersplomp@programming.dev is right on that it’s the multi-armed bandit problem if we assume that user preferences are static. If we introduce the complexity that users prefer familiar things, which I think is pretty fair, so users are more likely to keep watching from a bucket if it’s a familiar bucket, I assume that exploration gets heavily disincentivized and exhibits some pretty weird behavior, while exploitation becomes much more favorable. What I like about this is that, with only a small deviation from a classic problem, it would help explain what you also explain, which is getting stuck in corners.

Once you allow user choice beyond consume/log off, I think your way of thinking about it, as a turn based game, is exactly right, and your point about bin refinement is great and I hadn’t thought of that.

Thanks!

I feel enlightened now that you called out the self-reinforcing nature of the algorithms. It makes sense that an RL agent solving the bandits problem would create its own bubbles out of laziness.

You’re totally right that it’s like a multi-armed bandit problem, but maybe with so many possibilities that searching is prohibitively expensive, since the space of options to search is much bigger than the rate that humans can consume content. In other ways, though, there’s a dissimilarity because the agent’s reward depends on its past choices (people watch more of what they’re recommended). It would be really interesting to know if anyone has modeled a multi-armed bandit problem with this kind of self-dependency. I bet that, in that case, the exploration behavior is pretty chaotic. @abucci@buc.ci this seems like something you might just know off the top of your head!

Maybe we can take advantage of that laziness to incept critical thinking back into social media, or at least have it eat itself.

If you have any ideas for how to turn social media against itself, I’d love to hear them. I worked on this post unusually long for a lot of reasons, but one of them was trying to think of a counter strategy. I came up with nothing though!

taps the sign