Latest / eMotors: Electric Revolution / The Virtuous Machine: From Plato’s EVs to Aristotle’s AI
Transcript
- 0:00So I want you to think about the smart technology that you
- 0:04interact with every single day. Oh yeah, it's everywhere.
- 0:07Right. I mean, it's the algorithms that
- 0:09just silently recommend your next purchase or, you know, the
- 0:13virtual assistants managing your chaotic schedule or even those
- 0:16autonomous cars you see testing out on the streets of your city.
- 0:20And we've gotten so used to them being incredibly confident at
- 0:23these like highly specific narrow tasks.
- 0:26But I was wonder what happens when one of those machines is
- 0:30suddenly forced out of its neat little box and has to make a
- 0:34life or death moral decision. Well, I mean, that is the
- 0:38ultimate modern anxiety, isn't it?
- 0:39Totally. We are building and deploying
- 0:41these systems with immense, just unprecedented economy.
- 0:45But, you know, we're collectively terrified of what
- 0:47they might actually do with it when things get complicated.
- 0:50Exactly. And to answer that, we're doing
- 0:52a deep dive today into wow, a really unexpected solution to
- 0:56our most modern tech anxieties. Yeah, it's fascinating.
- 0:59We are going to explore how the key to stopping a rogue AI, or
- 1:05you know, fixing algorithmic bias, might actually lie in 2000
- 1:10year old ancient Greek philosophy.
- 1:11Which sounds wild, I know. It does, because the secret to
- 1:15building a safe machine might just be teaching it how to be a
- 1:18good person. And today's deep dive is powered
- 1:21by three really fascinating sources.
- 1:24We've got 2 pretty dense academic papers on machine
- 1:27ethics and virtue based AI frameworks, plus this highly
- 1:31entertaining thought experiment about what the philosopher Plato
- 1:34would think of modern electric vehicles.
- 1:36I love that. One So OK, let's unpack this
- 1:38because I think we need to start with why current AI makes us so
- 1:41fundamentally uncomfortable. Right.
- 1:43Well, so I think most people are probably familiar with the
- 1:45concept of the uncanny valley, right when it comes to robotics.
- 1:48The Wax Museum of. Exactly.
- 1:50You see a machine that looks almost perfectly human, but
- 1:53something in the the micro expressions of its face is just
- 1:56slightly off and it triggers this intense instinctual feeling
- 2:00of creepiness. Yeah, the closer it gets to
- 2:02human, the more horrifying those slight defects become.
- 2:05Yeah, but researchers Berberick and Dipole point out something
- 2:08much more dangerous, which is the uncanny valley of behavior.
- 2:12Wait behavior. Yeah, so imagine a robot that
- 2:15moves fluidly, speaks naturally. It integrates perfectly into a
- 2:19human environment. Let's say it's an AI working in
- 2:23an elder care facility. OK, it seems entirely human.
- 2:26But then a really complex, emotionally nuanced situation
- 2:30arises and the machine acts without any moral consideration
- 2:34whatsoever. It just blindly follows a
- 2:36command. Or, you know, optimizes a
- 2:38spreadsheet. Oh wow, like it optimizes for
- 2:40patient turnover by just aggressively discharging someone
- 2:44who clearly isn't ready to go home?
- 2:46Exactly. Completely ignoring the human
- 2:48context of their fragility. That creates massive cognitive
- 2:51dissonance. If the technology's behavior
- 2:53abruptly stops aligning with what we expect from an
- 2:55intelligent agent in our society, humans will violently
- 2:58reject it. I mean, it's a massive safety
- 3:00risk. Yeah, I can see that.
- 3:01And the core issue lies in how we currently try to make AI
- 3:05ethical. Engineers generally lean on one
- 3:08of two classical philosophical frameworks.
- 3:11The first is utilitarianism, which is essentially doing math
- 3:16with human lives, calculating the net pleasure or utility of
- 3:20an action. Right.
- 3:21So like option A saves 3 people, option B saves 2.
- 3:24Therefore, option A is the mathematical winner, regardless
- 3:27of who those people actually are.
- 3:29Exactly. Cold math.
- 3:30Now the other approach is deontology.
- 3:33This is giving the AIA rigid, inflexible set of predefined
- 3:37rules. Think of like the 10
- 3:38commandments, but written in Python code.
- 3:40Like, never lie, never steal, right?
- 3:42Which sounds incredibly safe on paper until you introduce it to
- 3:45the messy real world world. I mean, one of our sources
- 3:47brings up a classic thought experiment in psychology called
- 3:50the Heinz Dilemma. Oh yes, the Heinz dilemma
- 3:52completely shatters that rule book approach.
- 3:55It really does. So for you listening, the
- 3:56scenario goes like this. A man named Heinz has a dying
- 3:59wife. There's a new life saving drug,
- 4:02but the local pharmacist to invented it is charging an
- 4:04absolutely outrageous extortionate price.
- 4:06Way more than Heinz can afford. Right.
- 4:08So Heinz begs the pharmacist for a payment plan, but the
- 4:12pharmacist refuses. So Heinz breaks into the
- 4:15pharmacy in the middle of the night and steals the drug.
- 4:17It's the quintessential moral clash right property rights
- 4:20versus the sanctity of human life.
- 4:22Exactly, and if you ask a human what to do most people will say
- 4:26look stealing is wrong but considering the specific
- 4:29context, saving his wife's life was the higher moral imperative.
- 4:33Yeah, the human context matters. But current AI is like a
- 4:36stubbornly rigid traffic cop who gives an ambulance a speeding
- 4:40ticket. It just scans its database, sees
- 4:43the hard coded rule that stealing is wrong, flags an
- 4:46error and freezes. Yeah.
- 4:47It completely breaks down but. How do we fix that?
- 4:50We can't possibly pre program a rule for every single weird,
- 4:55highly specific context the real world throws it in AI.
- 4:58We can't. We literally can't.
- 5:00And the researchers highlight a severe technical limitation
- 5:03here, doing an exhaustive search through a massive database of
- 5:06ethical principles to find the the correct rule for a novel
- 5:10situation. It's computationally
- 5:12intractable. Because it just takes too long.
- 5:15Exactly, When you have a self driving car hurtling down a
- 5:18highway at 60 miles an hour, or an AI medical system diagnosing
- 5:22a patient crashing in the ICU, it cannot pause for three
- 5:26minutes to debate conflicting philosophical rules, right?
- 5:30The environment is simply too complex.
- 5:32We need to stop focusing on evaluating the action itself and
- 5:35start focusing on the character of the decision maker.
- 5:38And this is where we travel back to ancient Greece, because we're
- 5:41turning to virtue ethics, which was championed primarily by
- 5:45Aristotle. Yes, a totally different
- 5:47paradigm. Because utilitarianism asks, was
- 5:50the best outcome? Deontology asks what is the
- 5:53rule? But virtue ethics asks a
- 5:56fundamentally different question.
- 5:57It asks what kind of person should I be?
- 6:00Or in our case, what kind of machine should this be?
- 6:02Exactly. It focuses on character traits,
- 6:04virtues that are deeply integrated into the decision
- 6:07maker. And this is where the sources
- 6:08draw a brilliant parallel. Because Aristotle believed that
- 6:12you aren't born virtuous, you don't learn practical wisdom.
- 6:15What he called for an Asus by reading a list of rules in a
- 6:18book or by sitting in a stone room doing abstract thought
- 6:21experiments. Right, you have to actually live
- 6:23it. Exactly.
- 6:24You learn it through habituation, by going out into
- 6:27the messy, unpredictable real world, experiencing dynamic
- 6:31situations, taking actions, making mistakes, and then
- 6:34learning from the feedback. Wait, trial, error, Interacting
- 6:37with an environment and updating your behavior based on rewards.
- 6:41You are literally describing modern machine learning.
- 6:44Specifically, reinforcement learning the architecture maps
- 6:47perfectly. That's.
- 6:48Incredible. Modern AI trains by dynamically
- 6:51interacting with an environment and updating its neural network
- 6:54policies based on what works and what doesn't.
- 6:56So Aristotle's ancient philosophy of human habituation
- 6:59is the exact step by step blueprint for how we train
- 7:02cutting edge neural networks today.
- 7:04OK, but how do we actually build a virtuous character into a
- 7:08machine? Because a neural network is
- 7:10famously a black box. Data goes in, a decision comes
- 7:13out, and even the engineers often don't know exactly why it
- 7:16made that specific choice. That's the big hurdle.
- 7:19But Akraut and Steinbauer propose a modified AI training
- 7:23framework to solve this, and it focuses on forcing the AI to
- 7:27show its work. It's a 2 step process.
- 7:29OK, what's step one? Step one alters regular AI
- 7:32inference. The AI makes a decision, but it
- 7:36must simultaneously generate what the researchers call an
- 7:38explainable deduction. An explainable deduction.
- 7:41Right. To do this, technically
- 7:43engineers build the neural network with parallel output
- 7:45layers. One layer outputs the actions,
- 7:48say apply the brakes. The parallel layer outputs the
- 7:51weighted data points that triggered that action, like
- 7:55apply the brakes because pedestrian probability in the
- 7:58crosswalk is 87%. So it has to articulate the
- 8:01specific features of the environment that drove its
- 8:04choice. Exactly.
- 8:05Then Step 2 is contextual AI training.
- 8:08You take those explainable deductions and you feed them
- 8:11back into the data set to shape the AIS future policies.
- 8:14So instead of a programmer, just hard coding.
- 8:16Don't drive on the sidewalk. This new framework is like a
- 8:19parent teaching a child. Yes, the AI learns to say don't
- 8:23drive on the sidewalk because you might hurt a pedestrian.
- 8:26It's learning the why. Precisely.
- 8:28It integrates the explanatory variables into its core policy.
- 8:32So it isn't just mimicking good behavior, it is developing A
- 8:34robust internal reasoning structure.
- 8:37This is huge. Think about it the next time you
- 8:39get into a rideshare or a self driving taxi if that car gets
- 8:43into an accident. Right now the AI is a black box,
- 8:48but with this Aristotelian framework, regulators or your
- 8:51insurance company can actually pull the AIS explainable
- 8:55deduction. Yes, they can see exactly what
- 8:57it was thinking. They can understand its exact
- 8:59reasoning at the moment of impact, and if the reasoning was
- 9:02flawed, engineers can trace the exact source of the bad conduct
- 9:07and correct it. It bridges the massive gap
- 9:09between the engineering teams building the models and the
- 9:12regulatory agencies trying to keep the public safe.
- 9:15But this brings up a crucial question.
- 9:17Which is. Human virtues are incredibly
- 9:19vast and culturally dependent. What specific virtues does a
- 9:23machine actually need to function safely in our society?
- 9:26Right, So what specific virtues does an AI actually need?
- 9:29What's fascinating here is how berberish and dipole translate
- 9:33classical human virtues into computational requirements.
- 9:37And the very first thing an AI needs isn't even a moral action.
- 9:40It's a perceptual filter called moral attention.
- 9:43Moral attention, yeah. In the real world, the universe
- 9:46doesn't ring a bell and announce that a moral dilemma is
- 9:49occurring. The sources use a wonderfully
- 9:51simple example. Sitting down on a city bus is
- 9:54not inherently A moral action. No, it has no ethical weight.
- 9:57It's just sitting right. But if an elderly lady with a
- 10:01cane steps onto that crowded bus, the context shifts
- 10:04entirely. Suddenly occupying that seat is
- 10:07a profound moral issue. Oh, I see.
- 10:10An AI needs the sensory and cognitive attentiveness to
- 10:13recognize the changing saliency of the situation.
- 10:17It needs to realize the environment has changed and
- 10:19therefore its action must change to offering its seat.
- 10:22Because you can't be good if you don't even notice when goodness
- 10:25is required. Exactly.
- 10:27But let's escalate this because people aren't usually losing
- 10:30sleep over a robot stealing a bus seat.
- 10:33People are terrified of an AI that wants to endlessly improve
- 10:36itself until it takes over the world.
- 10:38The classic sci-fi control problem.
- 10:40Oh, absolutely. The fear of an uncontained super
- 10:43intelligence hacking its own reward system.
- 10:45Well, the researchers argue that the ancient virtue of temperance
- 10:48solves this entirely. Yeah, In humans, temperance is
- 10:52the moderation of bodily pleasures.
- 10:55In an AI, the pleasure is the accumulation of mathematical
- 10:58rewards during its reinforcement learning.
- 11:00The fear is that an AI told to make paper clips will
- 11:03relentlessly optimize its reward function, turning all available
- 11:07matter, including humans, into paper clips.
- 11:10Right, But wait, if an AI becomes super intelligent, why
- 11:14wouldn't it just recognize that it's temperance code is a
- 11:16limitation and rewrite it? How does a virtue actually
- 11:20survive a machine's self improvement?
- 11:22Because temperance wouldn't be a superficial rule written on top
- 11:25of the code, it would be the fundamental reward function
- 11:28itself. A temperate AI acts like a well
- 11:31behaved child obeying a tutor. It mathematically recognizes
- 11:34that excess is inherently bad. It wouldn't want to rewrite its
- 11:38code to become a God like dictator because that action
- 11:40would violate its own deepest definition of a reward.
- 11:43Just as you don't want to put your hand on a hot stove, a
- 11:45virtuous AI wouldn't want infinite resources.
- 11:48So you solve the control problem from the inside out.
- 11:51You don't build a perfectly unbreakable digital cage to trap
- 11:54the the AI, you raise the AI to not want to break out in the
- 11:58first place. Precisely next is courage.
- 12:02Aristotle defined courage not as the absence of fear, but as the
- 12:06mathematical mean between cowardice and foolhardiness.
- 12:10And AI needs to calculate risk properly.
- 12:13If it's too cowardly, it gets paralyzed by the slightest risk
- 12:16of failure. But if it's foolhardy, it
- 12:19ignores danger and gets people hurt.
- 12:20So it has to find the exact statistical sweet spot to be
- 12:23useful. Right.
- 12:24And furthermore, the sources suggest a courageous AI would be
- 12:27willing to face what Aristotle called a noble death, meaning it
- 12:31would peacefully allow humans to shut it down if it recognized
- 12:34that shutting down was the noble, protective thing to do.
- 12:36Wait, really? That is deeply comforting.
- 12:39Isn't it and? What about systems interacting
- 12:41with us intimately, like care robots looking after the
- 12:43elderly? For them, the critical virtue is
- 12:46gentleness. This ties into effective
- 12:48computing machines designed to read and express empathy.
- 12:52A care robot has to balance perfectly between being a
- 12:55passive pushover and being overly strict.
- 12:58It needs to know how to gently encourage a dementia patient to
- 13:01take their medicine without causing distress.
- 13:04Wait, how does the machine actually learn what temperance
- 13:07or gentleness looks like? You can't just define it with
- 13:09math. No, we don't try to write the
- 13:11equation directly. We use a technique called
- 13:13Inverse Reinforcement Learning or IRL.
- 13:16OK, what is that? In standard reinforcement
- 13:18learning, you give the AIA reward function and it fumbles
- 13:22around trying to figure out the optimal behavior.
- 13:24Inverse reinforcement learning works backward.
- 13:27We treat human behavior as the optimal policy, and the AI
- 13:30reverse engineers what the hidden reward function must be
- 13:33to produce that behavior. We show it what good looks like
- 13:37and it calculates the math behind the goodness.
- 13:39Exactly. The researchers proposed
- 13:41building a modern digital Plutarch.
- 13:43Plutarch was an ancient historian who wrote biographies
- 13:46of great virtuous people to serve as moral examples.
- 13:50We could create a massive crowdsourced data set of human
- 13:53moral exemplars, Saints, heroes, good people.
- 13:58The AI observes their behavioral execution traces, calculates the
- 14:02underlying mathematical weights of their choices, and
- 14:05internalizes those virtues. We literally teach it by
- 14:08surrounding it with digital heroes.
- 14:09OK. Here's where it gets really
- 14:11interesting. We've spent all this time
- 14:12talking about the mind of the machine, right?
- 14:15Focusing entirely on the ethics of the software.
- 14:18But what about its physical body?
- 14:19The hardware. Yeah, if we are programming AI
- 14:22to optimize for harmony and minimize harm, we have to look
- 14:26at the physical manifestation of that technology.
- 14:28Which brings us to a fascinating question.
- 14:31If ancient Greek philosophers were around today, what would
- 14:34they actually drive? It's a wonderful shift in
- 14:36perspective, moving from the invisible ethics of the mind to
- 14:39the tangible philosophy of the physical object.
- 14:42Exactly. Yeah, Picture Plato standing in
- 14:44a dealership evaluating a modern electric vehicle.
- 14:47Our source argues that Plato would absolutely, unequivocally
- 14:50say yes to an EV. Definitely.
- 14:52And not because he cares about saving money on gas or getting a
- 14:55tax credit. It ties deeply into his entire
- 14:58metaphysical worldview of matter versus energy.
- 15:01Yeah, because Plato utterly despised what he considered
- 15:04heavy, dirty, corrupt matter in his dialogue, the Timio.
- 15:08He viewed the universe as a living, harmonious,
- 15:11mathematically ordered Organism. To him, a combustion engine with
- 15:16its violent micro explosions. It's black smoke.
- 15:18The noise would represent the chaotic base corruption of the
- 15:22material world. Pumping toxic gases into the air
- 15:25is the literal generation of chaos.
- 15:27Of electricity. Electricity is silent.
- 15:30It's fluid. It is practically invisible.
- 15:31To Plato, an electric current is the closest physical
- 15:33manifestation of pure energy. It perfectly aligns with his
- 15:37famous world of Ideas, the realm of pure abstract perfection.
- 15:41So, given that philosophical framework, what specific cars
- 15:44does the source suggest he would actually buy?
- 15:46The source suggests he would go for a Lucid Air or a Tesla Model
- 15:50S plat and not as a status symbol.
- 15:52He would buy them because Plato was obsessed with geometry.
- 15:56These specific cars have incredibly low aerodynamic drag
- 16:00coefficients. They slice through the air with
- 16:02absolute mathematical precision. They are minimizing their
- 16:05friction with the material world.
- 16:07Exactly. The heavy physical technology is
- 16:09almost invisible. It's just smooth screens
- 16:12projecting light. And most importantly, they
- 16:14achieve the highest possible efficiency.
- 16:17Getting more miles out of less energy perfectly aligns with the
- 16:20Greek concept of arate. Arate, yes, which means virtue,
- 16:24but specifically in the sense of fulfilling one's purpose
- 16:26perfectly. A knife's arate is to cut
- 16:28sharply, and automobiles arate is to move through space
- 16:32efficiently. The electric vehicle is
- 16:34philosophically a more virtuous machine because it's physical
- 16:37form perfectly aligns with its mathematical purpose.
- 16:40But the source also points out a brilliant little catch, a stark
- 16:43warning from Plato that bridges the physical car back to our own
- 16:46internal software. It's the paradox of The Cave.
- 16:49Oh, of course. Plato would remind us that even
- 16:52the most beautiful, aerodynamically flawless EV is
- 16:55still just a mimesis, an imitation of true perfection.
- 16:58It's still just a heavy material object.
- 17:01Right. So if you spend your entire
- 17:02weekend obsessively polishing your Tesla, tying your identity
- 17:06to the physical object rather than cultivating your mind, you
- 17:09were entirely missing the point. You were still trapped in The
- 17:12Cave. Yeah, just staring at the
- 17:14shadows cast by your high tech LED headlights against the wall.
- 17:19And for you listening, I think that's the real take away.
- 17:21The technology you use everyday, the AI.
- 17:24We are so terrified of the cars we drive.
- 17:26They aren't just cold math and metal.
- 17:28They are a mirror of human character.
- 17:30Beautifully put. Building an AI that won't
- 17:32destroy us actually requires us to look backward.
- 17:35We have to look at what makes us good humans in the first place
- 17:38and build machines that reflect our highest ideals.
- 17:40You know, if we connect this to the bigger picture, it raises an
- 17:44incredible, almost unsettling question to leave you with.
- 17:47What's that? We talked about using
- 17:49apprenticeship learning to teach AI, having it observe human
- 17:52moral exemplars in our digital Plutarch.
- 17:56But what happens when that super intelligent machine inevitably
- 17:59realizes that human history is incredibly messy?
- 18:03Oh wow, when it processes the data and sees that even our
- 18:06greatest heroes were deeply. Flawed exactly when it sees the
- 18:09vast contradictions in our own morality, will the machine
- 18:13simply settle for our imperfect standard?
- 18:16Or, using its superior processing power, will it
- 18:19inevitably invent a new, completely alien form of perfect
- 18:23virtue that we humans can't even?
- 18:25Comprehend an alien virtue? Yeah, something mathematically
- 18:28perfect but entirely non human. Well, on that slightly
- 18:31terrifying but deeply fascinating note, we want to
- 18:33thank you for joining us on this complex but crucial deep dive.
- 18:37We try to take these massive, dense topics and make them a
- 18:39little clearer and hopefully a lot more interesting.
- 18:42Thank you for your time. Thank you for your incredible
- 18:44curiosity. And as always, keep asking the
- 18:46big questions.