Click image for larger size.
"Prepare for Landing"
by Bill Bonner
Baltimore, Maryland - "AI is now not just an investment fad. It is also responsible for thousands of job losses. PennyGem: "AI Claims Another Victim as Amazon Executes World’s Largest Tech Layoff - 30,000 Eliminated In 1 Week"
In the stock market, AI marks the difference between a bull market and a bear market. It is also the difference between a capital investment boom...and a capital investment bust. And by adding capital investment figures to GDP, it has recently meant the difference between growth and recession.
But what does it really produce? The results so far aren’t good. An MIT study showed that the Mag7 companies had spent $560 billion on AI over the last 18 months. The payoff? These investments returned $35 billion in revenue, not profit. If the profit margin on revenues were 10% this would mean a rate of return of just 0.006%. So, it’s time to ask: AI...what are ye? Fish or fowl? Day or night? Good or bad?
First, Artificial Intelligence is very artificial. But there’s nothing very intelligent about it. Those huge data centers we hear so much about are huge for a reason, because they’re set up for electronic grunt work. In keeping with our attempts at reductio ad simplicitas (making things simple enough so even we can understand them) Large Language Models are merely word crunchers. They crunch zillions of them. They simply ask: what is most likely to come next? You say ‘dog.’ They give you ‘barks.’ That’s why what you get from ChatGPT, for example, seems so boring. It gives you what you would most expect.
But still, it can do some remarkable work. You ask: ‘Ten planes are preparing to land at JFK. What are the odds that one crashes?’ Here’s the answer from ChatGPT: To estimate the odds that one of ten planes will crash, we need to know the probability of a crash per landing. Let’s go step by step.
1. Estimate the crash probability per flight or landing.
According to aviation safety data:
The chance of a commercial airline accident (including nonfatal) is roughly 1 in 1 million flights.
The chance of a fatal crash is even lower — about 1 in 10 million flights (or 10−710^{-7}10−7).
Let’s use p=1×10−7p = 1 \times 10^{-7}p=1×10−7 as a conservative estimate for a crash per landing.
2. Probability that no plane crashes:
If each of the 10 planes lands independently:
P(no crashes)=(1−p)10P(\text{no crashes}) = (1 - p)^{10}P(no crashes)=(1−p)10
So:
P(no crashes)=(1−10−7)10≈1−10−6P(\text{no crashes}) = (1 - 10^{-7})^{10} \approx 1 - 10^{-6}P(no crashes)=(1−10−7)10≈1−10−6
3. Probability that at least one plane crashes:
P(at least one crash)=1−(1−p)10≈10−6P(\text{at least one crash}) = 1 - (1 - p)^{10} \approx 10^{-6}P(at least one crash)=1−(1−p)10≈10−6
That’s about 1 in a million odds that one of those 10 landings will crash.
4. Putting it in perspective:
Scenario:
Any 1 of 10 planes crashes
Approximate probability: 1×10−61 \times 10^{-6}1×10−6
Odds: 1 in 1,000,000
All land safely
Approximate probability: 0.9999990.9999990.999999
Odds: 99.9999%
✅ Answer: The odds that one of 10 scheduled planes will crash when landing at JFK are about 1 in a million — extremely unlikely.
What this shows us is that AI can put together logical arguments ...with nouns, verbs, and numbers just where they should be. This alone will be hugely important. There are millions of people who earn their living by doing these kinds of logical, routine, and ‘bounded’ activities. Truck drivers, for example. Most of their work involves only two binary decisions: fast or slow, right or left. And when you drive through a toll booth, you could stop your car and urinate on the booth...or simply take out your phone and have a conversation with your mother. But the common range of actions is much more restricted. You pay the toll. You get change. The barrier goes up.
There is nothing that takes much intelligence going on. Just math and routine language: “Here’s your change. Thank you.” The typical clerk gets a limited range of requests. And he has only a limited range of responses. No one turns to him for an opinion on the allegorical imagery in the "Faerie Queene".
Scene from "The Faerie Queene," by Edmund Spenser (1553-1599)
While the oddball challenge still arises, most routine work could be handled by some version of AI. This is true not only of toll booth clerks and paint salesmen; most of doctors’, lawyers’ and other professionals’ work could also be done - insofar that it is logical and predictable - by a creature without a heart.
An accountant, for example, is not even expected to have a heart. He’s not encouraged to let sentiment persuade his calculation...or to think ‘outside of the box.’ Instead, he’s meant to stay in the box. And when the task calls for adding two plus two, he’s supposed to get four...not five...not three. Not even 4.001. Just like a machine. A proper AI tool will get four exactly. Every time.
Which makes AI very unlike a real human being. Humans wander out of the box. This is how we get our art, music and literature. This is also how we get big disasters...and big breakthroughs. We drop the latex on the stove, by accident, and discover ‘vulcanized’ rubber. We go on vacation and leave a petri dish in the lab without a cover; when we return, we find that a blue mold has prevented growth of the bacteria. Penicillin, we call it.
So, in an effort to make AI more like human intelligence, programmers introduced a bit of randomness. Instead of always choosing the most likely next word...or even the most correct number...they allowed the machine to use some judgement of its own. This is what leads to the factual errors and ‘hallucinations’ that have troubled AI systems. Sometimes, they just make stuff up. Sometimes they come to the ‘wrong’ conclusion. Forbes: "Elon Musk Says Grok Will Be Fixed After Chatbot Sided with Sam Altman in Spat Over Potential OpenAI Lawsuit." The randomness feature is much appreciated by students. It makes their college essays look a little more like they wrote them themselves. But randomness is not cool if you’re landing a 737. More to come...
No comments:
Post a Comment