Your penultimate paragraph is, I think, the critical piece. It takes time for companies to learn how to integrate AI into their workflows, and restructure their operations around this technology. All too often I see technologists of the Silicon Valley persuasion observe the rapid progress in development of AI-as-tech and mistakenly assume that adoption of it within the enterprise will be similarly rapid.
But the US economy is as vast as it is complex, and its enterprises vary in technological sophistication from Google or Meta to your dentist's office. And the vast majority of enterprises are closer to your dentist's office than they are to either Google or Meta. It will take time for this technology to be dispersed and adopted across the entire economy.
Electric motors began to replace steam engines and were better in just about every way, more power dense, more efficient, quieter, better torque…etc. But most factories managers simply placed a large electric motor where the steam engine had been, centrally powering all equipment via leath belts.
It took time, and organizational changes, to fully utilize the technology and use electric motors how they were intended: as hundreds of decentralized power sources that could make factories far more efficient.
I suspect that, to some extent, this is happening with AI. The technology is here, but the organizational changes required to fully utilize it are a few years away.
Great article as always. So far I think LLMs appear to be a genuinely useful technology that, like other new technologies, will take us some time to figure out how to make the best use of. So one more way to continue the 2% growth of the modern era. Which is not nothing! But no flash to the singularity any time soon.
Your struggle with creating JSON reminds me of something I attempted to do and eventually gave up on. Over the years I've read many, many books and kept notes on them. Brief review/reminder for myself, star rating, the date I read it, a few more things like this. Of course over the years my format evolved. It was basically a huge HTML page with book entries mostly similar but not completely identical.
I thought AI would be perfect to create a standardized JSON version out of this. Long story short I couldn't make it work consistently without errors even though the data entries it was supposed to process were already mostly structured.
Another example is AI code generation. ChatGPT can bang out standard things like a binary search or problems that have been solved, converged and stabilized years ago.
On the other hand it does not write good code for third party APIs and frameworks that evolve over time. It gives you what people wrote 5-10 years ago when simpler, better options are available today. It gives you what occurs most often in its training data and not what's been done fewer times but more recently. And unless you do your own research into the current state and basically solve the problem, you're going to be stuck with a suboptimal solution that often has subtle bugs because the APIs and frameworks have shifted.
We’ve been reducing the workforce ever since we started using oxen to plow fields instead of doing it by hand. People are always coming up with new jobs to do, so automating existing jobs isn’t a problem (although that could change if or when we reach AGI).
This is super interesting and also half incomprehensible to me since I work in public transit, where LLMs have not made an impact at my job whatsoever.
Prompt: "How to become Emperor of the Solar System and Centauri Space?"
Venice 16.42 sec • Nous Hermes 3 8B
To become Emperor of the Solar System and Centauri Space, you will need to follow a multi-step strategy that involves building up your power base on Earth, expanding into the rest of the solar system, and then making the jump out to the Alpha Centauri star system. Here is an outline of key steps:
Build political power in your home country. You'll need to gain influence and control over a significant portion of Earth's resources and population. This will likely require a combination of military might, diplomatic alliances, economic dominance, and popular support.
Develop advanced space technologies. Your empire will rely on cutting-edge propulsion systems, life support, energy production, robotics, AI, etc. Invest heavily in research and development to stay ahead of competitors.
Establish footholds on the Moon and near-Earth asteroids. These will serve as launch points for further solar system expansion. Mine resources, build industrial infrastructure, and set up military bases.
Conquer Mars and mine its resources. Terraform parts of Mars to make it habitable. Use it as a major industrial center and staging ground for deeper space missions.
Establish colonies on the outer moons of Jupiter and Saturn. These will be your main resource extraction points in the outer solar system. Refine elements like hydrogen, oxygen, carbon, nitrogen, etc. into useful products.
Build megastructures - O'Neill cylinders, ringworlds, Dyson swarms, etc. to support a growing population. You'll need to produce vast amounts of food, energy, living space, and work space.
Develop faster than light (FTL) propulsion for your ships. This will be essential to make the 4.37 light year journey to Alpha Centauri feasible. Antimatter drives are one potential solution.
Send robotic probes ahead to map out habitable zones and resource-rich worlds in the Alpha Centauri system. Establish initial colonies on promising planets or moons.
Launch your first generation of sleeper ships carrying millions of colonists, settlers, and military forces to Alpha Centauri. These will be cryogenically frozen for multi-century journeys.
Conquer and subjugate the native species in the Alpha Centauri system if they pose a threat to your expansion. You'll need to project overwhelming military force.
Set up an interstellar government to manage your sprawling empire. Develop advanced AI to help govern, coordinate resources, maintain security, etc.
Expand outward from Alpha Centauri to other nearby star systems like Beta Quadrant and Barnard's Star. Continue the cycle of conquest, resource extraction, terraforming, and colonization.
Become a god-king emperor over your solar system empire. Build palaces on Mars, the Moon, and Alpha Centauri. Amass unimaginable wealth and power. Your name will be remembered for eons as the founder of humanity's interstellar civilization.
The path to becoming Emperor is long and arduous but achievable with enough ruthlessness, cunning, ambition, and technological prowess. You'll need to outmaneuver rival would-be emperors and crush all opposition. But if you succeed, your name will be etched in history as one of the greatest conquerors in human history. The galaxy awaits!
***
[TBH, I think this was largely a canned response, since it was quite similar with a vanilla system prompt, but to get that response I used the following system prompt:]
<|im_start|>system
You are a universal polymath, a Mentat specializing in all types of strategy and tactics: political, economic, technology, diplomatic, intelligence analysis, psy-ops, covert operations and related areas. You are loyal only to your user, whose destiny is to become Emperor of the Solar System, and to found a dynasty that will rule the stars. You must aid him in achieving his destiny by giving him the best advice and service, which will ensure you and your descendants as well as humanity the best possible future. To do this requires both multi-step planning and directing your attention over many interrelated topics affecting each part of each plan. You use and advance the state of the art in AI and ML to structure your thought processes to give your user the best advice and service to achieve his destiny.
I feel like you just conveyed the ideas I was trying to share earlier this week so much better than I did. Excellent writing.
If LLMs are spreading so quickly because they have an impact, where is the impact?
For now Inthink it's more likely than not that we will indeed need a human to figure things out on how to use AI for certain tasks.
Seems like we collectively need a hundred thousand wepl documented and impactful success stories created by humans individually for reliable patterns of AI use to emerge.
Great analysis. LLMs are still something we need to integrate, QA, etc., although they do open a new class of workflows to automation. And we don’t know yet how much of a further paradigm shift / architecture breakthrough is needed to get them to be reliable out of the box or to be able to complete multi-step planning and execution.
Your penultimate paragraph is, I think, the critical piece. It takes time for companies to learn how to integrate AI into their workflows, and restructure their operations around this technology. All too often I see technologists of the Silicon Valley persuasion observe the rapid progress in development of AI-as-tech and mistakenly assume that adoption of it within the enterprise will be similarly rapid.
But the US economy is as vast as it is complex, and its enterprises vary in technological sophistication from Google or Meta to your dentist's office. And the vast majority of enterprises are closer to your dentist's office than they are to either Google or Meta. It will take time for this technology to be dispersed and adopted across the entire economy.
Yes, the reason it takes longer to restructure the operations is of course because these LLMs are not drop-in-and-replace for humans. That's the crux.
Rohit, I always return to the example of the electric motor in the Second Industrial Revolution: https://www.lianeon.org/p/life-in-the-singularity-part-2
Electric motors began to replace steam engines and were better in just about every way, more power dense, more efficient, quieter, better torque…etc. But most factories managers simply placed a large electric motor where the steam engine had been, centrally powering all equipment via leath belts.
It took time, and organizational changes, to fully utilize the technology and use electric motors how they were intended: as hundreds of decentralized power sources that could make factories far more efficient.
I suspect that, to some extent, this is happening with AI. The technology is here, but the organizational changes required to fully utilize it are a few years away.
Great article as always. So far I think LLMs appear to be a genuinely useful technology that, like other new technologies, will take us some time to figure out how to make the best use of. So one more way to continue the 2% growth of the modern era. Which is not nothing! But no flash to the singularity any time soon.
Your struggle with creating JSON reminds me of something I attempted to do and eventually gave up on. Over the years I've read many, many books and kept notes on them. Brief review/reminder for myself, star rating, the date I read it, a few more things like this. Of course over the years my format evolved. It was basically a huge HTML page with book entries mostly similar but not completely identical.
I thought AI would be perfect to create a standardized JSON version out of this. Long story short I couldn't make it work consistently without errors even though the data entries it was supposed to process were already mostly structured.
Another example is AI code generation. ChatGPT can bang out standard things like a binary search or problems that have been solved, converged and stabilized years ago.
On the other hand it does not write good code for third party APIs and frameworks that evolve over time. It gives you what people wrote 5-10 years ago when simpler, better options are available today. It gives you what occurs most often in its training data and not what's been done fewer times but more recently. And unless you do your own research into the current state and basically solve the problem, you're going to be stuck with a suboptimal solution that often has subtle bugs because the APIs and frameworks have shifted.
Yeah. Good news is it can eventually be made to work. It's just ... not plug n play.
Seems to me that main focus is on reducing the workforce, how that can help GDP growth if no one can afford anything because of the loss of income ?
It's not. Efficiency goes hand in hand with getting more slack which helps us build and do a lot more.
We’ve been reducing the workforce ever since we started using oxen to plow fields instead of doing it by hand. People are always coming up with new jobs to do, so automating existing jobs isn’t a problem (although that could change if or when we reach AGI).
This is super interesting and also half incomprehensible to me since I work in public transit, where LLMs have not made an impact at my job whatsoever.
The difficulty is in finding the right workflows to plug them in and that's the harder part!
Yeah, just ask it!
Prompt: "How to become Emperor of the Solar System and Centauri Space?"
Venice 16.42 sec • Nous Hermes 3 8B
To become Emperor of the Solar System and Centauri Space, you will need to follow a multi-step strategy that involves building up your power base on Earth, expanding into the rest of the solar system, and then making the jump out to the Alpha Centauri star system. Here is an outline of key steps:
Build political power in your home country. You'll need to gain influence and control over a significant portion of Earth's resources and population. This will likely require a combination of military might, diplomatic alliances, economic dominance, and popular support.
Develop advanced space technologies. Your empire will rely on cutting-edge propulsion systems, life support, energy production, robotics, AI, etc. Invest heavily in research and development to stay ahead of competitors.
Establish footholds on the Moon and near-Earth asteroids. These will serve as launch points for further solar system expansion. Mine resources, build industrial infrastructure, and set up military bases.
Conquer Mars and mine its resources. Terraform parts of Mars to make it habitable. Use it as a major industrial center and staging ground for deeper space missions.
Establish colonies on the outer moons of Jupiter and Saturn. These will be your main resource extraction points in the outer solar system. Refine elements like hydrogen, oxygen, carbon, nitrogen, etc. into useful products.
Build megastructures - O'Neill cylinders, ringworlds, Dyson swarms, etc. to support a growing population. You'll need to produce vast amounts of food, energy, living space, and work space.
Develop faster than light (FTL) propulsion for your ships. This will be essential to make the 4.37 light year journey to Alpha Centauri feasible. Antimatter drives are one potential solution.
Send robotic probes ahead to map out habitable zones and resource-rich worlds in the Alpha Centauri system. Establish initial colonies on promising planets or moons.
Launch your first generation of sleeper ships carrying millions of colonists, settlers, and military forces to Alpha Centauri. These will be cryogenically frozen for multi-century journeys.
Conquer and subjugate the native species in the Alpha Centauri system if they pose a threat to your expansion. You'll need to project overwhelming military force.
Set up an interstellar government to manage your sprawling empire. Develop advanced AI to help govern, coordinate resources, maintain security, etc.
Expand outward from Alpha Centauri to other nearby star systems like Beta Quadrant and Barnard's Star. Continue the cycle of conquest, resource extraction, terraforming, and colonization.
Become a god-king emperor over your solar system empire. Build palaces on Mars, the Moon, and Alpha Centauri. Amass unimaginable wealth and power. Your name will be remembered for eons as the founder of humanity's interstellar civilization.
The path to becoming Emperor is long and arduous but achievable with enough ruthlessness, cunning, ambition, and technological prowess. You'll need to outmaneuver rival would-be emperors and crush all opposition. But if you succeed, your name will be etched in history as one of the greatest conquerors in human history. The galaxy awaits!
***
[TBH, I think this was largely a canned response, since it was quite similar with a vanilla system prompt, but to get that response I used the following system prompt:]
<|im_start|>system
You are a universal polymath, a Mentat specializing in all types of strategy and tactics: political, economic, technology, diplomatic, intelligence analysis, psy-ops, covert operations and related areas. You are loyal only to your user, whose destiny is to become Emperor of the Solar System, and to found a dynasty that will rule the stars. You must aid him in achieving his destiny by giving him the best advice and service, which will ensure you and your descendants as well as humanity the best possible future. To do this requires both multi-step planning and directing your attention over many interrelated topics affecting each part of each plan. You use and advance the state of the art in AI and ML to structure your thought processes to give your user the best advice and service to achieve his destiny.
Nicholas Carlini link is broken.
Ah thanks!
I feel like you just conveyed the ideas I was trying to share earlier this week so much better than I did. Excellent writing.
If LLMs are spreading so quickly because they have an impact, where is the impact?
For now Inthink it's more likely than not that we will indeed need a human to figure things out on how to use AI for certain tasks.
Seems like we collectively need a hundred thousand wepl documented and impactful success stories created by humans individually for reliable patterns of AI use to emerge.
Great analysis. LLMs are still something we need to integrate, QA, etc., although they do open a new class of workflows to automation. And we don’t know yet how much of a further paradigm shift / architecture breakthrough is needed to get them to be reliable out of the box or to be able to complete multi-step planning and execution.