Offensive advantage and the vanity of ethics

I have recently shifted my “person I am obsessed with listening to”: my new guy is George Hotz, who is an eccentric innovator who built a cell phone that can drive your car. His best conversations come from Lex Fridman’s podcasts (in 2019 and 2020).

Hotz’s ideas bring into question the efficacy of any ethical strategy to address ‘scary’ innovations. For instance, based on his experience playing “Capture the Flag” in hacking challenges, he noted that he never plays defense: a defender must cover all vulnerabilities, and loses if he fails once. An attacker only needs to find one vulnerability to win. Basically, in CTF, attacking is anti-fragile, and defense is fragile.

Hotz’s work centers around reinforcement learning systems, which learn from AI errors in automated driving to iterate toward a model that mimics ‘good’ drivers. Along the way, he has been bombarded with questions about ethics and safety, and I was startled by the frankness of his answer: there is no way to guarantee safety, and Comma.ai still depends on human drivers to intervene to protect themselves. Hotz basically dismisses any system that claims to take an approach to “Level 5 automation” that is not learning-based and iterative, as driving in any condition, on any road, is an ‘infinite’ problem. Infinite problems have natural vulnerabilities to errors and are usually closer to impossible where finite problems often have effective and world-changing solutions. Here are some of his ideas, and some of mine that spawned from his:

The Seldon fallacy: In short, 1) It is possible to model complex, chaotic systems with simplified, non-chaotic models; 2) Combining chaotic elements makes the whole more predictable. See my other post for more details!

Finite solutions to infinite problems: In Hotz’s words regarding how autonomous vehicles take in their environments, “If your perception system can be written as a spec, you have a problem”. When faced with any potential obstacle in the world, a set of plans–no matter how extensive–will never be exhaustive.

Trolling the trolley problem: Every ethicist looks at autonomous vehicles and almost immediately sees a rarity–a chance for an actual direct application of a philosophical riddle! What if a car has to choose between running into several people or alter path to hit only one? I love Hotz’s answer: we give the driver the choice. It is hard to solve the trolley problem, but not hard to notice it, so the software alerts the driver whenever one may occur–just like any other disengagement. To me, this takes the hot air out of the question, since it shows that, as with many ethical worries about robots, the problem is not unique to autonomous AIs, but inherent in driving–and if you really are concerned, you can choose yourself which people to run over.

Vehicle-to-vehicle insanity: While some autonomous vehicle innovators promise “V2V” connections, through which all cars ‘tell’ each other where they are and where they are going and thus gain tremendously from shared information. Hotz cautions (OK, he straight up said ‘this is insane’) that any V2V system depends, for the safety of each vehicle and rider, on 1) no communication errors and 2) no liars. V2V is just a gigantic target waiting for a black hat, and by connecting the vehicles, the potential damage inflicted is magnified thousands-fold. That is not to say the cars should not connect to the internet (e.g. having Google Maps to inform on static obstacles is useful), just that safety of passengers should never depend on a single system evading any errors or malfeasance.

Permissioned innovation is a contradiction in terms: As Hotz says, the only way forward in autonomous driving is incremental innovation. Trial and error. Now, there are less ethically worrisome ways to err–such as requiring a human driver who can correct the system. However, there is no future for innovations that must emerge fully formed before they are tried out. And, unfortunately, ethicists–whose only skin in the game is getting their voice heard over the other loud protesters–have an incentive to promote the precautionary principle, loudly chastise any innovator who causes any harm (like Uber’s first-pedestrian-killed), and demand that ethical frameworks precede new ideas. I would argue back that ‘permissionless innovation‘ leads to more inventions and long-term benefits, but others have done so quite persuasively. So I will just say, even the idea of ethics-before-inventions contradicts itself. If the ethicist could make such a framework effectively, the framework would include the invention itself–making the ethicist the inventor! Since instead, what we get is ethicists hypothesizing as to what the invention will be, and then restricting those hypotheses, we end up with two potential outcomes: one, the ethicist hypothesizes correctly, bringing the invention within the realm of regulatory control, and thus kills it. Two, the ethicist has a blind spot, and someone invents something in it.

“The Attention”: I shamelessly stole this one from video games. Gamers are very focused on optimal strategies, and rather than just focusing on cost-benefit analysis, gamers have another axis of consideration: “the attention.” Whoever forces their opponent to focus on responding to their own actions ‘has the attention,’ which is the gamer equivalent of the weather gauge. The lesson? Advantage is not just about outscoring your opponent, it is about occupying his mind. While he is occupied with lower-level micromanaging, you can build winning macro-strategies. How does this apply to innovation? See “permissioned innovation” above–and imagine if all ethicists were busy fighting internally, or reacting to a topic that was not related to your invention…

The Maginot ideology: All military historians shake their heads in disappointment at the Maginot Line, which Hitler easily circumvented. To me, the Maginot planners suffered from two fallacies: one, they prepared for the war of the past, solving a problem that was no longer extant. Second, they defended all known paths, and thus forgot that, on defense, you fail if you fail once, and that attackers tend to exploit vulnerabilities, not prepared positions. As Hotz puts it, it is far easier to invent a new weapon–say, a new ICBM that splits into 100 tiny AI-controlled warheads–than to defend against it, such as by inventing a tracking-and-elimination “Star Wars” defense system that can shoot down all 100 warheads. If you are the defender, don’t even try to shoot down nukes.

The Pharsalus counter: What, then, can a defender do? Hotz says he never plays defense in CTF–but what if that is your job? The answer is never easy, but should include some level of shifting the vulnerability to uncertainty onto the attacker (as with “the Attention”). As I outlined in my previous overview of Paradoxical genius, one way to do so is to intentionally limit your own options, but double down on the one strategy that remains. Thomas Schelling won the “Sveriges Riksbank Prize in Economic Sciences in Memory of Alfred Nobel” for outlining this idea in The Strategy of Conflict, but more importantly, Julius Caesar himself pioneered it by deliberately backing his troops into a corner. As remembered in HBO’s Rome, at the seminal engagement of Pharsalus, Caesar said: “Our men must fight or die. Pompey’s men have other options.” However, he also made another underappreciated innovation, the idea of ‘floating’ reserves. He held back several cohorts of his best men to be deployed wherever vulnerabilities cropped up–thus enabling him to be reactive, and forcing his opponent to react to his counter. Lastly, Caesar knew that Pompey’s ace-in-the-hole, his cavalry, was made up of vain higher-class nobles, so he told his troops, instead of inflicting maximum damage indiscriminately, to focus on stabbing their faces and thus disfigure them. Indeed, Pompey’s cavalry did not flee from death, but did from facial scars. To summarize, the Pharsalus counter is: 1) create a commitment asymmetry, 2) keep reserves to fill vulnerabilities, and 3) deface your opponents.

Offensive privacy and the leprechaun flag: Another way to shift the vulnerability is to give false signals meant to deceive black hats. In Hotz’s parable, imagine that you capture a leprechaun. You know his gold is buried in a field, and you force the leprechaun to plant a flag where he buried it. However, when you show up to the field, you find it planted with thousands of flags over its whole surface. The leprechaun gave you a nugget of information–but it became meaningless in the storm of falsehood. This is a way that privacy may need to evolve in the realm of security; we will never stop all quests for information, but planting false (leprechaun) flags could deter black hats regardless of their information retrieval abilities.

The best ethics is innovation: When asked what his goal in life is, Hotz says ‘winning.’ What does winning mean? It means constantly improving one’s skills and information, while also seeking to find a purpose that changes the world in a way you are willing to dedicate yourself to. I think the important part of this that Hotz does not say “create a good ethical framework, then innovate.” Instead, he is effectively saying do the opposite–learn and innovate to build abilities, and figure out how to apply them later. The insight underlying this is that the ethics are irrelevant until the innovation is there, and once the innovation is there, the ethics are actually easier to nail down. Rather than discussing ‘will AIs drive cars morally,’ he is building the AIs and anticipating that new tech will mean new solutions to the ethical questions, not just the practical considerations. So, in summary, if you care about innovation, focus on building skills and knowledge bases. If you care about ethics, innovate.

4 thoughts on “Offensive advantage and the vanity of ethics

  1. Thought provoking article. QI do agree that ethicists tend to be overly cautious and lack skin in the game. They often tend to be arrogant relative to the value of their contributions.

    But seriously, is there no limiting principle to innovation, some things we just shouldn’t do? Your use of the term “black hats” suggests you might agree. In a sense, aren’t some black hats likely to be innovators too? I’m not drawing any definite line today. Just something for your consideration.

    • Great point! I am not saying that there should be no limiting principle on innovation– just that 1) they should not be from outside ethicists and 2) they should not be made ‘firm’ in advance or based on the precautionary principle. Start inventing, and figure out as you go whether there are ethical issues, rather than sitting down to work out ethical issues and ‘assuming the can opener’ (https://en.wikipedia.org/wiki/Assume_a_can_opener). Regarding ‘black hats’… I hardly think they are likely to listen to ethicists anyway, so in their case, I think that we can frame the question as “how to reduce vulnerabilities” rather than “how best to advise evil people.” That is a much harder, but more practical, question than how ethicists should influence ‘white hats’, and maybe it will inspire a Part II to this post. Thank you!

  2. The trolley problem remains: what if the driver either cannot or does not make a decision? Assuming it is appropriate to program the vehicle to make a decision that the driver isn’t making, the vehicle is essentially deciding a local version of the trolley problem. But even if we decide that it is not appropriate that the vehicle make a decision, that is in itself a trolley problem decision.

    Surely we should keep control with the driver as much as possible, but we should be prepared for that not to happen. So not all hot air is out just yet.

Please keep it civil

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s