@TheDag's banner p

TheDag

Per Aspera ad Astra

4 followers   follows 12 users  
joined 2022 September 05 16:04:17 UTC

				

User ID: 616

TheDag

Per Aspera ad Astra

4 followers   follows 12 users   joined 2022 September 05 16:04:17 UTC

					

No bio...


					

User ID: 616

I do wonder if we'll create a framework where places like OpenAI need to pay fraction of cents for each token or something. It would hit their profitability but would still make things fine if they achieve AGI.

Otherwise I agree that the open structure would be tough.

But there could be a utopia! Unlike Nuclear and other scenarios, I think it's likely this moves us far closer to a utopia, soon.

I am strongly of the opinion that since neoliberal PMC jobs are the easiest to automatic with AI, there will be incredibly strong regulation banning AI from taking the jobs of the PMC. The power to regulate is the power to destroy, and as incapable of actual productivity the PMC and their legion of bullshit jobs are, they know how to run a grift and bask in their own self importance.

I highly doubt this will happen. You talk as if the PMC is a giant union where everyone is aligned, which shows you don't understand the social context there and are clearly just poo-pooing your outgroup.

People in the PMC with power have capital, whether it's political, intellectual, or financial. The financial movers and shakers will not agree to regulating AI, at least until they have gotten their piece of the pie. Even if they do, it will take years and years to get everyone to agree on a framework.

You've also got the AI companies themselves. Altman has come out and said he doesn't think regulation at this stage is a good idea, and he's got an incredible amount of political and intellectual capital. Many people in government, for good reason, see Altman as one of the most important figure in the world right now. They don't want to piss him off.

The advent of generative AI heralds the single largest change in the structure of human society since the neolithic revolution (ie. the invention of agriculture and the settled society) 12,000 years ago.

I would actually argue this is a closer parallel to the cognitive revolution, or homo sapiens first discovery of culture, language, and general cognitive technology. The difference is that with the revolution from fire, or the agricultural revolution, or the industrial revolution, or even the internet, the AI revolution deals with intelligence and a paradigm of thinking itself. The Scientific revolution could also be a close contender, since it dramatically increase our ability to think and use our knowledge.

The only thing that seems certain is that it will radically reshape the life of every single human being on earth in the next 5-50 years.

Agree strongly here.

Currently, we're focused on the application of modern LLMs and other generative models to create media (writing, images, video etc) and to perform knowledge roles that involve a combination of text and data manipulation and basic social interaction (ie. the vast majority of PMC labor sometimes derogatorily referred to as 'email jobs'). But current models are so generalizable, and LLMs already appear to translate so well to robotics that even relatively complex physical labor is only a few years behind the automation of the PMC, especially given rapid improvements in battery technology and small motors, which are some of the other major bottlenecks for robotic labor.

The real step change in my opinion is once these models get good at things like drug discovery, mathematical proofs, and building models of physics. We have essentially been locked into a paradigm almost 100 years old in physics, and haven't found many fundamental changes in mathematical or chemical theory since then either, to my knowledge.

In the past, every time we had a major breakthrough in one of these fields it was enough to reshape the world entirely. Chemistry led to the industrial revolution, Newtonian Mechanics led to the scientific revolution. (or was the beginning, whatever.)

There is a (relatively persuasive) case to be made that the invention of agriculture led to a decline in the quality of life for the vast majority of human beings that lasted until the late 19th or early 20th century. It took 11,900 years for the neolithic revolution to pay quality of life dividends, in other words. We can only hope that the period of relative decline in quality of life is shorter this time round, or perhaps avoidable altogether.

As I mention above, I think the comparison to the agricultural revolution falls flat for a number of reasons. Admittedly most revolutions follow a pattern of short term negative issues with long term positive outcomes however.

At what level of 'smarts,' however, will an AI that is already training on how you do your job going to stop needing you around to do it?

At some point soon we will at least increase productivity by 1.5-2x per person. At that point why don't we collectively demand a 3 or 4 day workweek?

You might be interested in this post on LessWrong which discusses Scaffolded LLM's as natural language computers.. I'd be curious on @DaseindustriesLtd's take on this as well.

Key points:

What we have essentially done here is reinvented the von-Neumann architecture and, what is more, we have reinvented the general purpose computer. This convergent evolution is not surprising -- the von-Neumann architecture is a very natural abstraction for designing computers. However, if what we have built is a computer, it is a very special sort of computer. Like a digital computer, it is fully general, but what it operates on is not bits, but text. We have a natural language computer which operates on units of natural language text to produce other, more processed, natural language texts. Like a digital computer, our natural language (NL) computer is theoretically fully general -- the operations of a Turing machine can be written as natural language -- and extremely useful: many systems in the real world, including humans, prefer to operate in natural language. Many tasks cannot be specified easily and precisely in computer code but can be described in a sentence or two of natural language.

The LLM itself is clearly equivalent to the CPU. It is where the fundamental 'computation' in the system occurs. However, unlike the CPU, the units upon which it operates are tokens in the context window, not bits in registers. If the natural type signature of a CPU is bits -> bits, the natural type of the natural language processing unit (NLPU) is strings -> strings.

The RAM is just the context length. GPT4 currently has an 8K context or an 8kbit RAM (theoretically expanding to 32kbit soon). This gets us to the Commodore 64 in digital computer terms, and places us in the early 80s.


The obvious thing to think about when programming a digital computer is the programming language. Can there be programming languages for NL computers? What would they look like? Clearly there can be. We are already beginning to build up the first primitives. Chain of thought. Selection-inference. Self-correction loops. Reflection. These sit at a higher level of abstraction than a single NLOP. We have reached the assembly languages. CoT, SI, reflection, are the mov, leq, and goto, which we know and love from assembly. Perhaps with libraries like langchains and complex prompt templates, we are beginning to build our first compilers, although they are currently extremely primitive.

I find this framing extremely persuasive, and awesome in the true sense. If transformers can actually act as a new type of general purpose computer using natural language, the world will become strange indeed very quickly.

As @2rafa and others have mentioned, ML will be a step change in how human society creates value and interacts with the world more generally. Once we've achieved AGI, roughly defined as having an AI that can act at the level of an ordinary human, our ability to solve problems will drastically increase.

Intelligence is the generic solver for essentially any problem. Sam Altman himself has said that by 2030 he envisions a world where every product and service will either have integrated intelligence, or be angling towards that. This means that our phones, laptops, PCs, will all obviously be intelligent. However what most people don't realize is this technology will also effect our coffeemakers, stoves, thermostats, glasses, and practically every other technology you can think of. I'm sure adaptive clothing will exist soon with camoflauge like capabilities. People will be able to get realtime instructions into headphones telling them exactly how to complete each task.

Even these predictions only scratch the surface. If the true promise of AGI comes out it will also let us break through issues in hard mathematics, create brand new drugs, find extremely dense and powerful new materials. It will help navigate endless layers of bureaucracy, effortlessly pruning through the thousands of regulations that hold up large projects, helping us pinpoint ruthlessly where cost is added to solve the cost disease problem, and generally help unstick our public works. We could be building scintillating skyscrapers of filament-thin materials with bridges across the sky that glisten in the air, all in a decade. The future is truly difficult to even envisage, let alone predict.


In terms of comparisons to other revolutions, @2rafa says below:

There is a (relatively persuasive) case to be made that the invention of agriculture led to a decline in the quality of life for the vast majority of human beings that lasted until the late 19th or early 20th century. It took 11,900 years for the neolithic revolution to pay quality of life dividends, in other words. We can only hope that the period of relative decline in quality of life is shorter this time round, or perhaps avoidable altogether.

I agree that the agricultural revolution led to issues, a la Scott Alexander's review of Against the Grain.. That being said, I find the comparison of the AI revolution to agriculture as facile. Ultimately the reason the agricultural revolution proved bad for us was that we shifted our lifestyles from nomadic culture to a static culture - which inherently leads to problems of physical fitness, freedom, social control, and cultural institutions have to rapidly shift.

With the AI revolution, we have no idea how far it will go. The possibility space is far beyond what could have existed for any previous revolution. As doomers say, we could all die. We could all transcend our fleshly forms and become gods in ten years. China may create an ASI and lock us all into a totalitarian doom state forever.

The stakes here are far higher than the agricultural revolution, and I highly doubt our situation will parallel that trajectory.

At the end of the day if we can survive the AI revolution without any horrible outcomes of the x-risk or s-risk variety, I think it would be ridiculous to posit any sort of negative future. With intelligence at our fingertips, we will be able to finally achieve our potential as a species.

What do you have against Altman?

I’m glad you expect progress to continue. I do as well but I’m heavily biased so I try to temper my hopes.

I’m interested in putting money down but would need to include things like disability and other metrics. Raw unemployment is BS in the US at least.

Even an utter pittance would be more than enough to satisfy the average Joe if we get true AGI, even without self recursive improvement. We’re talking insane amounts of wealth here.

This is a good point actually. Depopulation may well happen but if it’s because everyone is rich and happy I don’t see the issue?

Diaspora is also close to my favored world, the journey of discovery is especially tantalizing. I also love the concept of the math mines.

I’m also skeptical on a personal identity level of totally copying the brain or uploading fully. I can’t quite bring myself to be sure that an uploaded me would be me, even with all the arguments about dreams unconsciousness etc.

I may choose to preserve my biological brain as long as possible for the safest form of identity transfer. Even once others have done it you can’t be sure they’re not just p-zombie copies. I imagine this could be quite controversial once these problems come up.

Isn’t this tax extremely regressive? As in the poor would be taxed a vast majority more % on their taxes than the wealthy?

Elites need people to feel elite to as well. I highly doubt they’ll kill us all. What is your proposed mechanism for this happening?

I must admit it is a beautiful reversal, and feels like quite a solid commuppance. I’m also part of the laptop class but I love seeing people have to eat their words.

The whole unemployment problem is artificial and ultimately fake

It is if you’re only concerned with whether or not there’s work to do. It’s a serious problem when you consider whether or not people can actually do the work, and how we coordinate that at a large scale.

There will always be work to be done, that’s in the nature of the universe. But how do we motivate people to work, organize their work, etc when jobs will be replaced far faster than they can be created?

You didn’t really engage with anything I brought up in my post, just hand waved it away by saying there will always be work. I agree, but explain to me how we get people to do that work?

Perhaps even if not every woman goes back into community building, we could focus more on child rearing or keeping up social connections?

I know in my personal life I wish I could spend more time with children teaching them, and I have plenty of friends I would’ve kept in touch with were it not for the demands of time. I see a noble and crucial position for women in repairing and maintaining the social fabric.

Frankly I’m more worried about how men will cope when most of the useful work is taken away.

It’s harder for me to scoff at the robot rapture after seeing the way these GPT-4 instance can be scaffolded to make a more general purpose system.

As you say, even if we don’t get anymore fundamental breakthroughs on the LLM scaling side, I’m convinced the ecosystem of tools and connecting LLM instances together will easily get us to AGI.

Our reference class for something like AGI should be more like the transition into industrial capitalism, except much faster, and on a much larger absolute scale.

I absolutely agree! This is why I am pushing for having a plan now rather than just letting it rip as we did during the Industrial Revolution. There was an incredible amount of unnecessary pain and suffering during that transition which we could’ve prevented with the correct amount of foresight.

What have we learned from that revolution? Hopefully we at least won’t force parents and children to work 12 hours a day six days a week in horrific conditions. Does an early UBI help?

Perhaps a slower rollout would be desirable? What lessons have we learned from history, now that we know the general direction of a technological revolution?

This is a very persuasive argument, the more I think about it the more I tend to agree. Useful economic work is a minority. Having personal experience at a startup it boggles the mind to realize some people really do 10x or 100x the work of others on short time frames.

I suppose my only concern would be our ability to grapple with the speed of change.

The sad truth is that we’re little meat sacks with very simple brains

However I take issue with this dreary view of things. Our very simple brains, when networked properly, were able to build AGI. Even if that is our greatest accomplishment, we’ve done some incredible things along the way as well. Just stopping here would satisfy the Natural Law crowd, inherent dignity in being human, etc.

That being said, I don’t see why we can’t augment ourselves to keep pace or at least stay in the orbit of an ASI. I’d imagine at a certain point most will choose to do that - we are a species of technologists after all. We want to feel relevant, and it’s hard to feel relevant if you can’t understand what’s going on.

It's pathological to fear the separation from our current regime of incentives.

As always your descriptions are on point, and I don’t disagree that our current situation is far from ideal assuming we have working AGI. I don’t have any sort of Stockholm syndrome about the status quo, in fact you could certainly put me in the effective accelerationism camp. (As much as that’s a coherent camp.)

My concerns are twofold - one, what if AGI does not ‘work as intended?’ What if we reach scaling limits soon, and only replace vast swathes of white collar work but don’t get into the recursive ASI machine god territory you seem to take for granted?

In that situation we could rapidly have a large class of people who are dispossessed, both economically and from a perspective of purpose, with our society. In the past these labor revolutions have primarily touched the underclass, but now we’re talking intelligent, well connected, moneyed classes being hit by massive layoffs all at once. Even if we do get ASI in 10-15 years it will be a brutal transition period.

Second, what if we do get ASI but TPTB succeed in regulating/controlling the systems to the degree where we do get a UBI, but everyone not owning a piece of the machine lives in gray square tenements the rest of their earthly lives?

Just because ASI holds promise doesn’t mean it automatically tears us from the embrace of ‘the festering undying corpse of our industrial civilization.’ Those poor incentives could well haunt us into the next era. I see no guarantee that ASI always decides to break our chains.

Nice post. I’d argue that the distancing from oratory also had a lot to do with the lack of belief in God and losing the rituals of the church after the world wars.

The problem of evil looms large over Christianity when the evil is so palpable and undeniable.

Oops, my ignorance is showing. Should've asked GPT-4.

Oh yeah, it's brutal. All the 100 hour days....