So had an interesting conversation with a friend who finished interviewing at Anthropic. The most notable part was the final stage where the mask dropped and the real face showed with very pointed questions.
The theme was "Effective Altruism" (EA). In this case, the idea is "If you build the most powerful AI that can control the world, then you can rule he world as a benevolent dictator".
The questions probed into the ethics and morality related to using dominance in AI to control the world. The goal was to ensure the candidate was okay with this idea, that they would be a obedient cog in the wheel and not develop a conscience later on.
For a technical guy, there are a lot of problems with the belief that if you build the most powerful AI, then you can rule the world, but lets put that aside for a bit and assume its possible. Assume that all it takes is more and more computational power to train more and more powerful AI until it reaches the threshold of cognitive capabilities necessary to rule the world. And lets assume that they are able to access the infrastructure necessary to run this monster and hook it up to all decision making systems so that it can rule over the world.
Thats. when the "EA" kicks in. This group, whoever achieves this superpower, will then be able to make the rules for the world and rule the world.
This is the belief system of Anthropic, and if we start from there, a lot of puzzle pieces will start fitting together. And since the puzzle itself always has two sides - the side the blackhats see (Black side) and the side the Whitehats see (White side), and these puzzle pieces need to fit both sides of the puzzle to make sense.
-
The push for datacenters and AI development by Trump admin. Putting aside the economic aspects of it, it now makes sense why Trump is so focused on this. Remember, its not enough to build the most powerful AI. Its also important to have the computational power to run it. If Trump (and We The People) control this infrastructre we control this powerful AI and what it can do. Turn off the infrastructure and off goes the AI.
-
Constant dooming about the datacenters. Makes perfect sense. If WHs controls the infrastructure, the enemy will never be able to rule the world. They would much rather prefer China win this race, so they will have a free hand to run it how they want. Hence spread doom and fear about data centers within the patriot community, and try and stump this race.
-
Push for Universal Basic Income - How do you keep people in line when the AI rules the world and you dont really need people? Keep them content with free money.
-
Of course the other answer for this question is simply get rid of extra humans - and that leads us to the real big push we have been seing - Push for depopulation
-
Why is the Left betting so heavily on commies like Mamdani, even though every logic screams that its a losing bid? Because a dictatorship ruled by AI will look like communism for all practical purposes. So while the techbros are building the AI-Emperor, the sleaze-bros will build the political side of it - communism.
-
This also explains why Hegseth fired some broadsides to Anthropic recently. And I think Elon is hedging the bets by making the Colossus deal with Anthropic
Now we are getting a full picture of this puzzle, with all seemingly unrelated observations coming into view well connected. How does it look from the Black side and the White side?
Black side: Even though they prefer CCP win this race, they still need to play nice with both Trump just in case. Whoever wins the race, they need the unlimited infrastructure provided by the winner to make their nightmares come true. If they think that this end goal is certain, they will expend all their energies towards this.
White side: Trump is using this greed for superpower AI as a carrot. He is getting these companies to agree to invest trillions to build the infrastructure. He is even providing them access to China as well in even more daring carrot dangling (hence the 17 CEOs going to China recently). The key for skynet to succeed is secrecy and shadows. Thats why Trump is doing everything openly. He is also leveraging the psyops for fear-mongering Trump's datacenter plans into helping immunize people against skynet so that even with all the infra and tech, built in the open, WeThePeople will never allow ourselves to lose control.
Not sure what this means. Computers, ie. silicon chips and programs, are dumb. Way dumber than humans. They can certainly compute things faster, but that doesn't make them smarter.
faster at following their programming, not more intelligent
As I said, "That begs the question of what "intelligent" means" -- because we all know machine "intelligence" is NOT the same as HUMAN intelligence. And machines are pretty stupid in many areas where humans shine. But machine "intelligence" IS very powerful, including in many ways ours is not. And therein lies the rub
To avoid getting stuck in mental tar pits with this topic there are a couple things I find helpful. One, whenever you are tempted to use the word intelligence, replace it with the word capability. Secondly, keep fresh in your memory a handful of keystone examples that cut through the abstract hand waving and cliches. Such as that old chestnut - computers only do what we program them to do. When surfing the edge of knowledge, thinking in cliches is an immediate wipe-out.
Here is one keystone example. Some years ago the U.S. Air Forces was testing an AI that one of their defense contractors wrote for the automated piloting of fighter jets and bombers. Of course the Air Forces wanted to see how it performed under simulated settings. So they contrived attack missions for the AI to participate in, and scored the software controlled planes on how many targets it took out and how quickly. The performance measure was something the software knew and trained on, because, of course, machine learning from training data is what transforms plain old software into AI (plus a few other ingredients). When the testers concluded that a mission was complete the controllers would "radio" all fighting elements, instructing them to return to base. We are now getting close to the kicker. The AI was allowed to analyze its own performance from the simulated missions, and then adapt. Guess what happened next? In subsequent missions, after taking off from the tarmac the first thing the AI did was circle back and shoot down its own radio control tower. The AI could then continue with the mission and clean the table. It had figured out that if it removed the one choke point that was restraining its actions it could rack up new high scores.
The moral of the story is twofold. In terms of capabilities AIs are optimization seekers. That is how they are designed. This attribute is especially potent when the AI has planning capabilities, and those capabilities are matched with the computational resources to search millions to billions of probable outcomes. The second half of the moral is that AIs are not humans, did not grow up in a family and inside of society, has no moral compass, and therefore has no real clue as to the assumptions of good behavior that by rights should constrain its actions. A human pilot does not need to be told to not shoot out the friendlies communications infrastructure. The AI, however, ran the math and found the loophole that led to a better "solution". It is this uncomfortable combination of search capabilities married to out-of-the-box thinking - let's call it that for sake of prose - that forces the designers to wrap their AI inside of so-called guardrails.
I think it is fair to note that it is in the design and QA of AI guardrails where companies and software engineers reveal their true allegiances. The guardrail code, though, is usually a closely guarded secret.
THAT is among the most important points about AI. They aren't even organic and thus have no empathy; basically, they are functional psychopaths, although they don't have the emotional damage that animates sociopaths (psychopaths aren't necessarily sociopathic, but that's another story). Programming is what we're relying on to give AI the necessary common sense and artificial empathy to make them useful without being seriously dangerous.
I don't believe that will turn out to be a good call, frankly.
Just as "It only takes one atomic bomb to ruin your whole day", it may only take one (planned or unplanned) AI event (hacking and disabling most of the electric grids on Earth? Designing and releasing an artificial viral or bacterial plague? etc) to destroy civilization or even end the human race. The electric grid thing would likely kill 90% the Americans in the first year, according to a study I've seen -- and that sounds very believable.
That sounds a bit over-the-top. Five years from now, I don't believe it will. OR, I'll be completely wrong and we'll all be safe in the better world that AI is building for us. I have no idea which way it'll go.
"computers only do what we program them to do" IS a cliche, and a thoroughly misleading one at that.
Decades ago, I learned that running a program through a new version of a programming language could, all by itself, change a behavior of the program I'd written.
You can say that in THAT case, the computer was STILL doing only what it was programmed to do -- but not what I, "the programmer", told it to do, but rather something that another programmer (who wrote the program or compiler) told it to do.
How many programmers are involved in writing the code for, say, ChatGTP? Do they all know every detail of what all the other programmers have done with the code?
And since AI ITSELF is now writing a significant amount of the new code, and it's unlikely that the programming team knows and understands those additions and which other parts of the code they connect and interact with . . . why would anyone think that AI is completely PREDICTABLE, even by its programmers, when it interacts with the unpredictable humans who make use of it and the unpredictable events of the world in which its actions take place?
Unexpected action by software has been a problem since the first electronic computers came on line, and even today, with millions of dollars spent on debugging software and then months or years-long beta campaigns by well funded companies, programs STILL often do things the programmers DID NOT specifically tell the software to do (in addition to the bugs they DID specifically program for, but didn't expect a particular situation would arise that would cause the problem).
I'm not expecting that to change.
Yes, and that's another problem. What are the guardrails of the military AI being crafted by the Chinese or other militaries? Or by ANY of the commercial AI companies? As you point out, we don't know. And even if we did, we'd be fools to think those guardrails will act as intended in all circumstances. For that matter, the guardrails in some cases may be malicious as written.
C++ is the undisputed world champion at that. I hate that language. I guess it is only fair to admit that has to do with the fact that I could never grok its arcane syntax plus semantics. I could never reach a level where anything about it ever felt natural. Unfortunately for me it is the dominant language of professionals for systems programming. It effectively locked me out of a career as a software engineer. So I had to find other means of gainful employment, such as emptying waste paper baskets.
In the spirit of better late than never, it's overdue to state that I enjoyed reading your reply. To the main point that you highlight at the outset, I'm paying attention to robotics. At two levels.
The first level is in the development of humanoid robots, which the Japanese, the South Koreans, and militaries around the world adore. So it's NOT not going to happen. Because what's coming is the merger of four things. One, the grounds-up training of perceptual systems of vision, auditory, sensory, and bipedal locomotion. Which two, will be married to LLMs, packaging high-level intelligence mimicry. Three, the physical form is wrapped in increasingly realistic fake skin and facial motor control. In what year the uncanny valley will be traversed I do not know, but it is getting closer. Which leads to four, the transition from "training" to "raising" the humanoid robots in factory, then office, then home environments. When this convergence nears we will approach something that earns the artificial intelligence moniker. Or at least earns serious consideration. As the warning goes, when the robots themselves ask to participate in the debate on stage in a round table discussion the answer is probably already determined.
The second level is swarm intelligence. Imagine the above but with thousands or more of them in high speed direct communication. Or a swarm of military attack drones all in one communication mesh, tied into an over-arching surveillance network. This prospects makes me jittery. My overall image of AI is that of hiking up a snowy ridge line in the Himalayan mountains. Yeah, reaching the peak might be a fantastic, exhilarating accomplishment. But one false step and you tumble to your death.
Speculative thought can keep going of course in all sorts of science fiction directions. But those are the two I'm alert towards in terms of near future.
There's still no intelligence. It's just the same software using statistics from previous results. There still is no thinking on the part of the software. It's still just running a program.
I agree that it depends on the definition of "intelligence". Seems like a hard word to define correctly/completely and I'm certainly not going to take a stab at it. However, my guess is that no reasonable definition would have the computer coming out ahead of the human in any "intelligent" task. As I said, computers can certainly compute faster, but I'm going to guess that a good definition of "intelligence" doesn't depend on how fast something is completed.
Intelligence is indeed a difficult word to define. Capacity for abstraction and complex thought is a handy concise formulation. But that restricts itself to the upper reaches of intellect. As a working definition I like: the ability to correctly assess the current situation and to successfully project future consequences. This covers a broad range, including animal intelligence.
Here's a really homey example of (low) intelligence that fits into my definition. I'm upstairs in my home office finishing my lunch. I go to the upstairs bathroom. The toilet paper runs low. Not being completely clueless so as not to run out, I go down to the the basement to retrieve another roll before the current one is empty. When I return upstairs I spot my empty plate sitting next to my desk. DOH! That was dumb. I should have grabbed the plate, dropped it off in the kitchen, then gone to the basement to fetch the toilet paper. If I were smarter I would have folded both tasks into one trip.
When it comes to the computer software, now near-universally called AI, I prefer the term Intelligence Mimicry. LLMs are a far ways from full-on organic intelligence, but still impressive in many tasks that exists at the surface level interface of human intelligence, e.g. question & answering, written responses. Today's frontier models do better than most humans at college entrance exams to graduate school. MMLU-Pro: A robust and challenging evaluation of a language model's multitask accuracy, benchmark and paper.
And if I may round out with a racist comment, the paper I linked has 17 authors contributing from three universities. All three universities are prestigious, 2 Canadian and 1 in the U.S., top-tier in computer science. Judging by the names the authorship consists of 2 Indians and 15 Chinese. Unless we reverse our education system there our culture is so screwed.
We'll have to agree to disagree. There is no such thing as Artificial Intelligence. It's just a new buzz word for neural nets 2.0 as neural nets really didn't take off.
Here's an example, and it's actually something that really happened a few months back. My granddaughter was over and I told here there were some cookies on the table and she could have some. I could tell by the way she was looking at the package, they happen to be Tate's cookies, she wasn't familiar with the opening/closing mechanism. I watched as she looked at the package for a few seconds and she then opened the tabs and unrolled the top so she could get at the cookies. She was six years old at the time. The only way a computer could figure this out is if the programmer programmed this problem/solution. As I said, computers are dumb, and they always will be.