BrickedKeyboard
Hi David. Reason I dropped by was the whole concept of knowing the distant future with too much certainty seemed like a deep flaw, and I have noticed lesswrong itself is full of nothing but ‘cultist’ AI doomers. Everyone kinda parrots a narrow range of conclusions, mainly on the imminent AGI killing everyone, and this, ironically, doesn’t seem very rational…
I actually work on the architecture for current production AI systems and whenever I mention approaches that do work fine and suggest we could control more powerful AI this way, I get downvoted. So I was trying to differentiate between:
A. This is a club of smart people, even smarter than lesswrongers who can’t see the flaws!
B. This is a club of well, the reason I called it boomers was I felt that the current news and AI papers make each of the questions I asked a reasonable and conservative outcome. For example posters here are saying for (1), “no it won’t do 25% of the jobs”. That was not the question, it was 25% of the tasks. Since for example Copilot already writes about 25% of my code, and GPT-4 helps me with emails to my boss, from my perspective this is reasonable. The rest of the questions build on (1).
It would be lesswrongness.
Just to split where the gap is :
- lesswrongers think powerful AGI systems that can act on their own against humans will soon exist, and will be able to escape to the internet.
- I work in AI and think powerful general AI systems (not necessarily the same as AGI) will exist soon and be powerful, but if built well will be unable to act against humans without orders, and unable to escape or do many of the things lesswrongers claim.
- You believe AGI of any flavor is a very long way away, beyond your remaining lifespan?
Consider a flying saucer cult. Clearly a cult, great leader, mothership coming to pick everyone up, things will be great.
…What if telescopes show a large object decelerating into the solar system, the flaw from the matter annihilation engine clearly visible. You can go pay $20 a month and rent a telescope and see the flare.
The cult uh points out their “sequences” of writings by the Great Leader and some stuff is lining up with the imminent arrival of this interstellar vehicle.
My point is that lesswrong knew about GPT-3 years before the mainstream found it, many OpenAI employees post there etc. If the imminent arrival of AI is fake - like the hyped idea of bitcoin going to infinity or replacing real currency, or NFTs - that would be one thing. But I mean, pay $20 a month and man this tool seems to be smart, what could it do if it could learn from it’s mistakes and had the vision module deployed…
Oh and I guess the other plot twist in this analogy : the Great Leader is saying the incoming alien vehicle will kill everyone, tearing up his own Sequences of rants, and that’s actually not a totally unreasonable outcome if you could see an alien spacecraft approaching earth.
And he’s saying to do stupid stuff like nuke each other so the aliens will go away and other unhinged rants, and his followers are eating it up.
Having trouble with quotes here **I do not find likely that 25% of currently existing occupations are going to be effectively automated in this decade and I don’t think generative machine learning models like LLMs or stable diffusion are going to be the sole major driver of that automation. **
- I meant 25% of the tasks, not 25% of the jobs. So some combination of jobs where AI systems can do 90% of some jobs, and 10% of others. I also implicitly was weighting by labor hour, so if 10% of all the labor hours done by US citizens are driving, and AI can drive, that would be 10% automation. Does this change anything in your response?
No. Even if Skynet had full control of a robot factory, heck, all the robot factories, and staffed them with a bunch of sleepless foodless always motivated droids, it would still face many of the constraints we do. Physical constraints (a conveyor belt can only go so fast without breaking), economic constraints (Where do the robot parts and the money to buy them come from? Expect robotics IC shortages when semiconductor fabs’ backlogs are full of AI accelerators), even basic motivational constraints (who the hell programmed Skynet to be a paperclip C3PO maximizer?)
- I didn’t mean ‘skynet’. I meant, AI systems. chatGPT and all the other LLMs are an AI system. So is midjourney with controlnet. So humans want things. They want robots to make the things. They order robots to make more robots (initially using a lot of human factory workers to kick it off). Eventually robots get really cheap, making the things humans want cheaper and that’s where you get the limited form of Singularity I mentioned.
At all points humans are ordering all these robots, and using all the things the robots make. An AI system is many parts. It has device drivers and hardware and cloud services and many neural networks and simulators and so on. One thing that might slow it all down is that the enormous list of IP needed to make even 1 robot work and all the owners of all the software packages will still demand a cut even if the robot hardware is being built by factories with almost all robots working in it.
**I just think the threat model of autonomous robot factories making superhuman android workers and replicas of itself at an exponential rate is pure science fiction. **
- So again that’s a detail I didn’t give. Obviously there are many kinds of robotic hardware, specialized for whatever task they do, and the only reason to make a robot humanoid is if it’s a sexbot or otherwise used as a ‘face’ for humans. None of the hardware has to be superhuman, though obviously industrial robot arms have greater lifting capacity than humans. Just to give a detail what the real stuff would look like : most robots will be in no way superhuman in that they will lack sensors where they don’t need it, won’t be armored, won’t even have onboard batteries or compute hardware, will miss entire modalities of human sense, cannot replicate themselves, and so on. It’s just hardware that does a task, made in factory, and it takes many factories with these machines in it to make all the parts used.
think:
Software you write can have a “belief” as well. The course I took on it had us write Kalman filters, where you start with some estimate of a quantity. That estimate is your “belief”, and you have a variance as well.
Each measurement you have a (value, variance) where the variance is derived from the quality of the sensor that produced it.
It’s an overloaded word because humans are often unwilling to update their beliefs unless they are simple things, like “I believe the forks are in the drawer to the right of the sink”. You believe that because you think you saw them their last. There is uncertainty - you might have misremembered, as your own memory is unreliable, your eyes are unreliable. If it’s your kitchen and you’ve had thousands of observations, your belief has low uncertainty, if it’s a new place your belief has high uncertainty.
If you go and look right now and the forks are in fact there you update your beliefs.
academic AI researchers have passed him by.
Just to be pedantic, it wasn’t academic AI researchers. The current era of AI began here : https://www.npr.org/2012/06/26/155792609/a-massive-google-network-learns-to-identify
Academic AI researchers have never had the compute hardware to contribute to AI research since 2012, except some who worked at corporate giants (mostly deepmind) and went back into academia.
They are getting more hardware now, but the hardware required to be relevant and to develop a capability that commercial models don’t already have keeps increasing. Table stakes are now something like 10,000 H100s, or about 250-500 million in hardware.
https://www.semianalysis.com/p/google-gemini-eats-the-world-gemini
I am not sure MIRI tried any meaningful computational experiments. They came up with unrunnable algorithms that theoretically might work but would need nearly infinite compute.
Personally I imagine him as a cult leader of a flying saucer cult where suddenly an alien vehicle is actually arriving. He’s running around panicking tearing his hair out because this wasn’t actually what he planned, he just wanted money and bitches as a cult leader. And because it’s one thing to say the aliens will beam every cult member up and take them to paradise, but if you see a multi-kilometer alien vehicle getting closer to earth, whatever it’s intentions are no one is going to be taken to paradise…
The one issue I have is that “what if some are their beliefs turn out to be real”. How would it change things if Scientologists get a 2 way communication device, say they found it buried in Hubbard’s backyard or whatever and it appears to be non human technology - and are able to talk to an entity who claims it is Xenu. Doesn’t mean their cult religion is right but say the entity is obviously nonhuman, it rattles off the method to build devices current science knows no method to build and other people build the devices and they work and YOU can pay $480 a year and get FTL walkie talkies or some shit sent to your door. How does that change your beliefs?
1, 2 : since you claim you can’t measure this even as a thought experiment, there’s nothing to discuss 3. I meant complex robotic systems able to mine minerals, truck the minerals to processing plants, maintain and operate the processing plants, load the next set of trucks, the trucks go to part assembly plants, inside the plant robots unload the trucks and feed the materials into CNC machines and mill the parts and robots inspect the output and pack it and more trucks…culminating in robots assembling new robots.
It is totally fine if some human labor hours are still required, this cheapens the cost of robots by a lot.
- This is deeply coupled to (3). If you have cheap robots, if an AI system can control a robot well enough to do the task as well as a human, obviously it’s cheaper to have robots do the task than a human in most situations.
Regarding (3) : the specific mechanism would be AI that works like this:
Millions of hours of video of human workers doing tasks in the above domain + all video accessible to the AI company -> tokenized compressed description of the human actions -> llm like model. The llm like model thus is predicting “what would a human do”. You then need a model to transform the what to robotic hardware that is built differently than humans, and this is called the “foundation model”: you use reinforcement learning where actual or simulated robots let the AI system learn from millions of hours of practice to improve on the foundation model.
The long story short of all these tech bro terms is robotic generality - the model will be able to control a robot to do every easy or medium difficulty task, the same way it can solve every easy or medium homework problem. This is what lets you automate (3), because you don’t need to do a lot of engineering work for a robot to do a million different jobs.
Multiple startups and deepmind are working on this.