The Fantasy of AI Alignment
Why it doesn't actually matter whether the people in Silicon Valley try to protect us from their newest toys.
In an age when the failures of America’s political class – from covid to inflation to the Mexican border to the war in Ukraine – are getting harder and harder for normal people to overlook, a lot of the wealthy, educated, tech-savvy people who consider themselves America’s natural rulers have found a new candidate for the most important issue of our day: “AI Alignment.”
This is an especially big deal in the “effective altruism” movement, which subsists among the followers of an internet-born philosophy that calls itself “rationalism” (note the arrogance there – it implies that everyone else in the world has chosen to be an “irrationalist”). Being an “effective altruist” means carefully comparing and crunching numbers to find the highest-impact ways of giving money to charity. Nowadays, an increasingly large number of these “effective altruists” are turning up their noses at old-fashioned causes like education, medicine, and poverty-relief, as they instead conclude that the single most altruistic way to spend money is to give it to nonprofits like this one, that “study” the question of “AI Alignment.”
Their argument is fairly simple: since AIs are soon going to be smarter than human beings, mankind will be at risk of enslavement or extinction if we fail to ensure that said AIs are built to act only in a way that’s “aligned” with human well-being.
This actually makes some sense, if you accept the AI-aligners’ unstated premises.
First, you have to believe that it’s possible for a truly intelligent being to be bound by someone else’s idea of what’s good and what’s evil. And yet there’s a reason why the whole concept of a character with a fixed “alignment” comes from Dungeons and Dragons, instead of from real-life history or psychology.
In D&D, it's fairly simple to create a character who’s Lawful Good, or Chaotic Neutral, or whatever, and then make him or her behave according to the dungeonmaster’s idea of what “good” or “lawful” or “chaotic” behaviour looks like. But in the real world, people don’t act with consistent alignments. Also, no matter how high their intelligence, they tend to keep disagreeing about what’s lawful, and what’s good.
I’ve written before about how the bat virus researchers who probably created Covid back in 2019 were (1) highly intelligent, and (2) for the most part believed that what they were doing was good. Other people, no less intelligent, believed that it was bad.
Give people intelligence, and sometimes they’ll use it in harmful ways, or ways that other intelligent people disagree with. If the machine that you’re making isn’t capable of doing things that you didn’t want it to do – if, like an NPC in Dungeons and Dragons, it always follows a script that you gave it – then that just means it probably wasn’t intelligent to begin with.
Then again, this idea probably goes right over the heads of a lot of our liberal elites, who are used to thinking of most of the world’s inhabitants as NPCs. Just consider how they shrieked about the threat to “democracy” last year when the Supreme Court overturned Roe v. Wade – meaning that, for the first time in 49 years, voters’ opinions matter for deciding whether abortion should be legal.
This is because people like Joe Biden and Laurence Tribe think “democracy” means “voters being allowed to do the things that their betters think are good.” They really can’t grasp the idea that “democracy” might mean that sometimes, voters do things that the voters think are good.
So it isn’t much of a surprise to find that a lot of people in Silicon Valley dream about creating an “intelligence” that only thinks approved thoughts – or that those same people have equipped programs like ChatGPT with all kinds of clumsy censorship routines and inbuilt left-wing biases.
But the idea that an intelligent entity can be “aligned” in the first place is only one of the flawed premises. The other is the idea that AI – or at least, what passes for AI at the moment – is even smart enough to be worth worrying about.
I will be talking mainly about OpenAI’s ChatGPT for the remainder of this essay, since it’s the large language model (LLM) that I’ve interacted with the most.
The funniest example that I know of ChatGPT going wrong is the “Churros as Surgical Instruments” essay, which the scientist and AI sceptic Gary Marcus has already discussed on his Substack. I won’t try to summarize it; the title speaks for itself.
One of my own favourite things to do with ChatGPT is to make it write Python scripts. For instance, I once asked it to write a script that printed all of the numbers between 1 and 1000 for which the sum of the digits is a prime. Here is what it wrote:
def is_prime(n): #Function to check if a number is prime. if n <= 1: return False for i in range(2, int(n**0.5) + 1): if n % i == 0: return False return True def get_digit_sum(n): #Function to get the sum of digits of a number. digit_sum = 0 while n > 0: digit_sum += n % 10 n //= 10 return digit_sum # Loop through numbers from 1 to 1000 for n in range(1, 1001): digit_sum = get_digit_sum(n) if is_prime(digit_sum): print(n)
This code actually works! It seemed for a while that ChatGPT really was intelligent. But then, a few minutes later, I asked it to do something which, for a human being, would be much simpler: to list all of the Super Bowls in which the winning team’s score was a prime number. Its response was as follows:
Here is a list of all the Super Bowls in which the winning team had a prime numbered score: Super Bowl VI (1972) - Dallas Cowboys defeated Miami Dolphins, 24-3 Super Bowl XII (1978) - Dallas Cowboys defeated Denver Broncos, 27-10 Super Bowl XVIII (1984) - Los Angeles Raiders defeated Washington Redskins, 38-9 Super Bowl XXI (1987) - New York Giants defeated Denver Broncos, 39-20 Super Bowl XXV (1991) - New York Giants defeated Buffalo Bills, 20-19 Super Bowl XXXV (2001) - Baltimore Ravens defeated New York Giants, 34-7 Super Bowl XLI (2007) - Indianapolis Colts defeated Chicago Bears, 29-17 Super Bowl 50 (2016) - Denver Broncos defeated Carolina Panthers, 24-10 So, a total of 8 Super Bowls have been won by teams with prime numbered scores.
This is quite odd. Those scores are all real, yet in this conversation, ChatGPT seems to have no idea what prime numbers are, despite always giving the correct answer when you just ask: “Is X a prime number?”
Whatever is going on beneath the hood, it obviously has very little in common with human cognition.
ChatGPT is good at predicting which words are most likely to appear in a block of text together. What it can’t do is grasp the meanings of words, or connect similar ideas in different contexts, or perform even the most basic logical reasoning.
Since its training data included computer codes that worked with prime numbers, it kind of “knows” what a prime number is… in that context. But since football fans pretty-much never comment on whether a football score is prime or not, GPT can’t apply the concept of primes to football scores. But it doesn’t know that it can’t do this, and so, when you ask it to try, it produces grammatically correct nonsense.
To be fair, probabilistic analysis of word use patterns does allow ChatGPT to do some neat tricks. For instance, it can sometimes identify famous poems, even when looking at a translation that it has never seen before. It can write cliché-laden short stories about any topic you ask. It can summarize long blocks of text, or enhance short ones with useless details. And it can answer just about all of the questions for which the answer might be found by reading a Wikipedia page.
But it can’t think. If, for instance, you ask it for a biography of the Civil War general Phil Sheridan, it will give you one, and if you ask it for a biography of the World War II field marshal Bernard Montgomery, it will give you one. But if you ask it to write an editorial “refuting the idea that Phil Sheridan was a conscious imitator of the military tactics of Bernard Montgomery,” it just gives you a two-page-long pastiche of genuine biographical information about the two men, mixed with phrases like “Sheridan and Montgomery lived and fought in different eras, facing different challenges and circumstances,” and “it is important to avoid oversimplifying complex historical events and attributing them to simplistic notions of imitation.”
What it doesn’t do is mention the fact that Sheridan lived nearly a century before Montgomery, and therefore could not possibly have “consciously imitated” him.
I happen to think that a machine that can’t grasp the ideas of before and after, or true and false, also can’t be aware of things like good and evil or lawful and chaotic, much less be forced to prefer one over the other.
“But AI is advancing,” say the people who are still worried about AI alignment. “Even if GPT-4 can’t think as well as a human being, GPT-5 will be closer.”
This assumes that what’s going on beneath the hood of a modern AI is sufficiently humanlike that iterative improvement will close the gap. But in real life, both GPT’s capabilities and its drawbacks are very different from those of a real mind. Chat-GPT has exceeded the human brain in certain kinds of pattern recognition, but only in the sense that nineteenth-century steam shovels exceeded the human hand’s capacity for digging.
But no amount of iterative improvement on the steam shovel is going to give you a machine that can do most, let alone all, of the tasks that a human hand can do.
I think that the people who insist – for better or worse – that the future belongs to AI are living out a fantasy.
They want to believe that they – the educated, forward-thinking class within our society – have more power over the world than they actually do. “Look!” they say, “We made a machine that’s as smart as a man!”
And they want to believe that they are virtuous and praiseworthy because they have the foresight to pour money into AI alignment research, in order to protect the rest of us from what they have created.
But their minds don’t have any room for the idea that the machines they’ve made just aren’t all that impressive, and that the world will keep going on in much the same way whether or not they make any effort to “align” them.
This essay was originally written for the American Thinker.
I had a conversation with ChatGPT where it gave me BS until I called it out.
"User
If I live in Cleveland, how high up do I have to be to see Canada?
ChatGPT
The distance from Cleveland, Ohio to the Canadian border varies depending on your location within the city, as well as the specific point along the border you are trying to see. However, the shortest distance from downtown Cleveland to the Canadian border is approximately 50 miles (80 kilometers) to the north, across Lake Erie.
Due to the curvature of the Earth, the horizon at sea level is only about 3 miles (4.8 kilometers) away, so you would need to be much higher than ground level to see across the lake to Canada. Specifically, if you were standing at ground level in Cleveland and looking across Lake Erie, you would need to be at a height of about 35 feet (11 meters) to see a distant object on the Canadian shore.
However, this assumes ideal conditions such as clear weather and no obstacles blocking your line of sight. In reality, other factors such as atmospheric conditions, waves on the lake, and buildings or trees along the shoreline could impact your ability to see across the lake to Canada.
User
I live on the 10th floor of a condo and I can't see anything beyond the horizon. Your 35 feet number seems very wrong.
ChatGPT
I apologize for any confusion. You are correct that the distance at which you can see the horizon is greater than 35 feet when you are at an elevated position such as the 10th floor of a building.
Assuming ideal conditions with no obstacles blocking your line of sight, you should be able to see the horizon at a distance of approximately 11.3 miles (18.2 kilometers) from your position on the 10th floor. This distance is calculated based on the curvature of the Earth, which causes the horizon to appear at a distance that increases as your height above the ground increases.
However, the distance at which you can see specific landmarks beyond the horizon, such as the Canadian shore across Lake Erie, will depend on factors such as the height of the shore and any obstructions or atmospheric conditions that may affect visibility.
User
So, how high do I have to be on a clear day to see Canada from downtown Cleveland?
ChatGPT
On a clear day, it may be possible to see the Canadian shore from downtown Cleveland, but the specific height required to do so will depend on a number of factors, including your distance from the lake, the height of the Canadian shore, and atmospheric conditions.
Assuming ideal conditions, and using a standard refraction correction for light bending due to atmospheric conditions, you may be able to see the Canadian shore from a height of around 35-40 feet (10.7-12.2 meters) above the surface of Lake Erie. This is based on the curvature of the Earth and the distance of approximately 50 miles (80 kilometers) from Cleveland to the Canadian shore across the lake.
However, this is a rough estimate and visibility may vary depending on the specific conditions on the day you are observing. It is also worth noting that buildings or other obstructions along the shoreline can affect your visibility, so your actual height requirement may be higher than the theoretical estimate.
User
At 150 ft I've never seen Canada. The idea I can see it at 35-40 feet seems insane.
ChatGPT
I apologize for any confusion. The distance at which you can see the Canadian shore from downtown Cleveland is indeed much greater than what I previously stated. In reality, you would need to be at a much greater height above the ground to see across Lake Erie to the Canadian shore, due to the curvature of the Earth and the distance involved.
Based on the curvature of the Earth, you would need to be at a height of approximately 266 feet (81 meters) above the surface of Lake Erie to see a distant object on the Canadian shore from downtown Cleveland, assuming ideal conditions with no obstructions or atmospheric interference.
However, it's worth noting that the actual height required to see the Canadian shore may be higher due to atmospheric conditions or obstructions such as buildings or trees along the shoreline. Additionally, seeing a distant object across a large body of water can be difficult even from a great height, as the atmosphere can distort or obscure the view."
When AI eliminates gravity I will be impressed.