Gad Zoox

In which we consider whether AI is killing dogs and cats

Amazon’s Zoox started offering driverless rides in San Francisco last week. The New York Times covers this in an article purporting to give a balanced look at self-driving taxis. But the list of downsides is so silly it seems calculated to win over skeptics. If it just straightforwardly talked about how amazing Waymo and (I presume) Zoox are then anyone not already on board would dismiss it as propaganda.

Instead it first focuses on how much momentum Waymo and Zoox have and how Waymo-pilled San Francisco is. Key Cialdini persuasion principle there.

The upsides are covered reasonably. Drastically less death sure is an upside. Capitalism and consumer choice, sure. Eliminating the risk of your Uber driver assaulting you probably shouldn’t count as an upside, since the risk is already so low, but the article admits that.

Then the article gets to the so-called downsides. Self-driving taxis don’t break the law, which is terribly frustrating for passengers and other drivers in too much of a hurry to obey the speed limit. And they gallingly prioritize safety in other ways, like not double parking to pick you up like human drivers do. You might have to walk around the corner — maybe with luggage! — to get to your car. What else? Some of them are more expensive. For now. And not counting Zoox, which is free while they get people hooked and will supposedly be priced competitively with Lyft and Uber when they do charge money. Worst of all, self-driving cars are only wildly safer than humans, not perfect. A Waymo even killed a cat once.

What’s this about murdering kittens?

It was just that one time, baby, I swear it. The article links to another New York Times article about the cat, which, after spending 1200 words talking about how much everyone hates Waymo for this, eventually admits that pets are killed horrifyingly often by human drivers and that in this case the cat was under the Waymo when it pulled away so presumably there was nothing the Waymo could’ve done to prevent it.

I lied about it being just that one time, though. Over Waymo’s 100+ million miles, it has happened... roughly twice? The only other publicized incident was a dog that ran out from behind a parked car in a way that seems to have made the collision unavoidable. In any case, my robot friends (GPT-5.1-Thinking, Claude 4.5 Opus, and Gemini 3 Pro, which, by the way, have gotten shockingly competent lately) all agree that Waymo seems to be orders of magnitude safer in terms of pet deaths and even with the most pessimistic possible assumptions is still many times safer.

And Tesla?

Speaking of things that have only happened that one time, the crowd-sourced data on Tesla’s latest Full Self-Driving (FSD) software, version 14, has now logged at least one so-called critical disengagement out of a few tens of thousands of miles or so tested. At worst, a critical disengagement could be a case where the car would’ve crashed had the human in the driver’s seat not intervened. More likely it was milder than that but we’re extremely in the dark about the state of Tesla’s self-driving abilities. It seems very impressive in YouTube videos (and I’ve even found some YouTubing Tesla owners I trust to cover this as objectively as they can). But being safe to read a book while driving, or not being in the driver’s seat at all, means a lot more 9s of reliability, as engineers say. No individual Tesla owner can distinguish between 99.9% safe and 99.999% safe in their own lifetime, and the former is not enough. It would only be a matter of time before a merely 99.9% safe autonomous vehicle killed you.

With a naive extrapolation from the crowdsourced data and assuming you drive as much as a typical American (plus pessimistic assumptions about what a critical disengagement is) it would take your Tesla about 5 months to kill you if you didn’t supervise it.

I think in reality it’s better than that, especially factoring in the Tesla robotaxis’ ~250k cumulative miles, with plausibly as few as zero would’ve-killed-someone disengagements. So I believe Tesla’s latest FSD is at least 99.9% safe, whatever we mean by that, but that Tesla just isn’t adding 9s as fast as other companies. I covered a big reason why in last week’s AGI Friday.

I know a lot less about how Tesla’s humanoid robots are going (not close to commercial deployment in any case) but maybe this is Elon Musk’s plan to leapfrog the competition on self-driving?

Random Roundup

• Suppose I tell you that outcome X is what’s most likely, and then I turn around and say X is probably not going to happen. Did I contradict myself? Technically, no. If outcome X is rolling a 7 with two standard dice then X is indeed what’s most likely, occurring with probability 17%. That’s higher than the probability for any other roll. But it means you can be 83% sure you won’t roll a 7. “Most likely to happen” and “probably going to happen” sound similar but are different. Why am I mentioning this? Because it’s what happened with AI 2027. I do think this was a strategic/rhetorical mistake on the part of the authors. On the other hand, I’m a huge fan of technical correctness. In any case, the AI 2027 authors have said that AI research has progressed a little slower than they expected so far and their new most likely year for hitting AGI is 2028 and their median estimate, which ranged from 2028-2035 among the authors when they published, has also moved later. Breathing room? I’m on the record saying from the start that these timelines seemed too aggressive and I still think so. I just can’t say that with much confidence. And of course “probably sometime in the 2030s” is still terrifying soon.

• To put that “slower than expected” in perspective, it means “only slightly super-exponential”:

  

  That’s assuming you’re on board with “length of programming task the AI can complete on its own” as a meaningful measure of progress towards AGI. Or towards superhuman coding, which does require additional assumptions to get from there to AGI.

• I enjoyed the documentary “The Thinking Game”, about Demis Hassabis and DeepMind, and it’s now free on YouTube. I wish it didn’t underplay the risks of AGI so much though.