If you want to try to interact with a mockup of a persuasive vending machine, here is Vendy, a little GPT I threw together. It knows that the solution to any problem is probably buying a refreshing lemonade, and is here to help: https://chat.openai.com/g/g-LMszzSJYv-vendy-the-friendly-vending-machine
Thanks for another great article and all that you do and share in educating all of us interested in this space...as much as I enjoy the productive benefits of AI it's clear, no matter what spin people are putting on it, it will be replacing A LOT of jobs. I'm just wondering how you think things may unfold and any suggestions you might have for people worried about a future with AI (if you haven't seen Jon Stewart's take on Daily Show last week, definitely worth watching).
This is going to be a great leveler. My hunch is that we can look to other recent levelers to see how they effected the markers they entered, to get a sense for how AI will effect things in the next say 5 years.
One example that comes to me is Shopify and other similar e-commerce builder companies. The net effect seems to have been a commodification of small web businesses. Where before there were significant barriers to entry for building specialized sites, now there are almost none, leading to 1) the proliferation of e-commerce sites, and as a result, 2) very high levels of competition that compete away most of the previous excess profits. Good for consumers (except that the lower profit levels push more sellers into fraud or borderline fraud), and bad for individual e-commerce companies. 3) a proliferation of hucksters offering “tricks” to boost otherwise low profits.
Widespread SEO is another “leveler”, that by dropping barriers to entry for site rankings resulted in a race to the bottom for the web.
I’m just thinking out loud basically. My hunch is that this will strip the difficulty from many jobs, commodify them, allowing more people in hut dropping compensation levels significantly. If agents get good enough to completely replace individual workers in a significant portion of the white collar workforce, then all bets are off and after a lot of screaming and shouting and political chaos we will probably have some form of UBI. I also wonder if training AIs on video (which seems imminent) will revolutionize robotics a lot faster than people currently expect, and also take the emphasis away from advanced sensors and physical tech by making the software extremely flexible
Good observations, Shawn. On the robotics question, for the past few months Alexander Kruel's daily Substack "Axis of Ordinary" has regularly featured arXiv papers on just such training of robots by video alone. The state of the art seems quite advanced, but I think it's waiting for a good, practical demonstration. I think Paul Allen's Coffee Cup Challenge will do the most to convince people: an autonomous robot that enters the front door of a modern, American suburban home, finds the kitchen, and makes a cup of drop coffee without spills or breaking anything. The day that happens, robotics adoption will increase at an exponential rate.
Astonished by your opening. I really don’t see evidence that any model had doubled or more GPT-4 in capabilities over the last thirteen months.
Your extrapolation is at odds with recent data. (Update: I see the paper you have linked to, but I think we have entered a period of diminishing returns since mar 2023; also the paper is about amount of compute required, not capabilities.)
“When life gives you lemons, don’t make lemonade. Make life take the lemons back! Get mad! I don’t want your damn lemons, what the hell am I supposed to do with these?” - Cave Johnson
In seriousness though, I’m excited to try vendy as well as to read Ethan’s new book. Really enjoying my time with different AI tools these days… using them as research accelerators, ideation tools, fun (making an original story with pictures along with my daughters), and today I used a proprietary GenAI bot to help me write a python script at work (I’m at best a very basic python user). The script worked first try, and after another couple of shots of refinement it does everything we needed it to efficiently. It would have taken me days to do what I did in 30 minutes today because of AI. So exciting!
It's cruel that you provided this A# B# thing challenge and then not providing the solution. I wasted 15 minutes on this, my brain bleeding because I cannot do these tasks, my heart triggered from severe math trauma from school and feeling dumb, and then finally thinking I MIGHT have solved it but I can't check it. Honestly. Also, there is nothing worse than math teachers or programmers saying "here's an AI test that even kids can solve". I couldn't then and I can't now. And this has made me so angry that I think I should devote my research career on investigating why so many humans like me need the same kind of storytelling to get math as AI does. Actually that's a great idea. I'm less angry now thank you. Also, love your blog.
If anyone could tell me whether #A B# B# was correct that would be great.
I'd bet you have the right answer, since I got the same answer in two different ways. It seems it doesn't matter if you do all the possible computations for (adjacent #'s of) a row at one go (e.g. the starting setup has two), compared to making just one computation/move at a time. Making multiple moves at one go seems to make the process go faster in terms of total moves needed. In this example, it's 6 moves (multiple moves per turn) vs. 7 moves (single move per turn).
Wonderful post, Ethan. Yet one more example of why I am always excited to get this Substack in my inbox.
As fast as the models advance, I suspect there will be substantial friction in implementing them in the workplace. That is, I expect the effect of AI on the economy (GDP, productivity, employment) to be muted for the next 3-4 years. A friend of mine, a software development manager at a midsize SaaS company, is actively working to replace his low performing employees with AI. He has been running parallel experiments where he tasks an employee and also GPT3.5 to write code from detailed specs. For six months now, GPT3.5 has outperformed them considerably. His bosses, however, are resistant: laying people off is hard, and (ironically) expensive. Plus there is hesitance to trust a new technology.
What will happen, I suspect, is that people will simply not be hired. Positions will stay vacant, then be eliminated, with AI taking up the slack, either directly via agents or indirectly via covert or overt use by current employees. This effect will be especially pronounced in government, where union rules making lay offs of low performers essentially impossible.
The first sign of this, I think, is what is happening with management consulting. Ethan's study with Boston Consulting Group may have been a pebble that started a landslide: if the consultant with AI is better than just the consultant - and faster - how good could the AI be on its own?
Interesting post. Did the guy end up paying 10k to somebody? I have seen the power of prompting first hand. Building a lot of new abilities thx to emerging capabilities in my passion project uni.com.ai
Great read! AI breakthroughs remind me of people breaking records in sports:
● Nobody thought it was possible to run a marathon under 2h. Until someone did
● Nobody thought it was possible to deadlift 500kg. Until someone did
Crazy how things are going. Not sure how I feel about this. But I don't think anybody can stop this train now. I just hope our society doesn't turn into the movie *I, Robot*
I’m genuinly curious… When you say “we don’t know what these systems are capable of” and in the same breath write “multiple people figured out how to get AI to solve the problem within the day using prompts alone”, I feel you are begging the question here.
First of all, it looks like we’re reverse engineering based on a known outcome we’re trying to achieve. And only with very specific prompting that outcome is achieved, which means the underlying thing isn’t a result of generalization.
Secondly, who’s achievement is this really? The system, or the cleverness of the folks that find the right combination of buttons to press to make the model dance?
Another great article. Read it and immediately went out and bought...
LEMONADE.
buy his book ;-)
If you want to try to interact with a mockup of a persuasive vending machine, here is Vendy, a little GPT I threw together. It knows that the solution to any problem is probably buying a refreshing lemonade, and is here to help: https://chat.openai.com/g/g-LMszzSJYv-vendy-the-friendly-vending-machine
Thanks for another great article and all that you do and share in educating all of us interested in this space...as much as I enjoy the productive benefits of AI it's clear, no matter what spin people are putting on it, it will be replacing A LOT of jobs. I'm just wondering how you think things may unfold and any suggestions you might have for people worried about a future with AI (if you haven't seen Jon Stewart's take on Daily Show last week, definitely worth watching).
this continues to be the most terrifying column i regularly read on any subject
This is going to be a great leveler. My hunch is that we can look to other recent levelers to see how they effected the markers they entered, to get a sense for how AI will effect things in the next say 5 years.
One example that comes to me is Shopify and other similar e-commerce builder companies. The net effect seems to have been a commodification of small web businesses. Where before there were significant barriers to entry for building specialized sites, now there are almost none, leading to 1) the proliferation of e-commerce sites, and as a result, 2) very high levels of competition that compete away most of the previous excess profits. Good for consumers (except that the lower profit levels push more sellers into fraud or borderline fraud), and bad for individual e-commerce companies. 3) a proliferation of hucksters offering “tricks” to boost otherwise low profits.
Widespread SEO is another “leveler”, that by dropping barriers to entry for site rankings resulted in a race to the bottom for the web.
I’m just thinking out loud basically. My hunch is that this will strip the difficulty from many jobs, commodify them, allowing more people in hut dropping compensation levels significantly. If agents get good enough to completely replace individual workers in a significant portion of the white collar workforce, then all bets are off and after a lot of screaming and shouting and political chaos we will probably have some form of UBI. I also wonder if training AIs on video (which seems imminent) will revolutionize robotics a lot faster than people currently expect, and also take the emphasis away from advanced sensors and physical tech by making the software extremely flexible
Good observations, Shawn. On the robotics question, for the past few months Alexander Kruel's daily Substack "Axis of Ordinary" has regularly featured arXiv papers on just such training of robots by video alone. The state of the art seems quite advanced, but I think it's waiting for a good, practical demonstration. I think Paul Allen's Coffee Cup Challenge will do the most to convince people: an autonomous robot that enters the front door of a modern, American suburban home, finds the kitchen, and makes a cup of drop coffee without spills or breaking anything. The day that happens, robotics adoption will increase at an exponential rate.
Bought the book....it's awesome....TY!
Thank you!
Agree. Have already recommended it to anyone who will listen to me 😂
Astonished by your opening. I really don’t see evidence that any model had doubled or more GPT-4 in capabilities over the last thirteen months.
Your extrapolation is at odds with recent data. (Update: I see the paper you have linked to, but I think we have entered a period of diminishing returns since mar 2023; also the paper is about amount of compute required, not capabilities.)
“When life gives you lemons, don’t make lemonade. Make life take the lemons back! Get mad! I don’t want your damn lemons, what the hell am I supposed to do with these?” - Cave Johnson
In seriousness though, I’m excited to try vendy as well as to read Ethan’s new book. Really enjoying my time with different AI tools these days… using them as research accelerators, ideation tools, fun (making an original story with pictures along with my daughters), and today I used a proprietary GenAI bot to help me write a python script at work (I’m at best a very basic python user). The script worked first try, and after another couple of shots of refinement it does everything we needed it to efficiently. It would have taken me days to do what I did in 30 minutes today because of AI. So exciting!
It's cruel that you provided this A# B# thing challenge and then not providing the solution. I wasted 15 minutes on this, my brain bleeding because I cannot do these tasks, my heart triggered from severe math trauma from school and feeling dumb, and then finally thinking I MIGHT have solved it but I can't check it. Honestly. Also, there is nothing worse than math teachers or programmers saying "here's an AI test that even kids can solve". I couldn't then and I can't now. And this has made me so angry that I think I should devote my research career on investigating why so many humans like me need the same kind of storytelling to get math as AI does. Actually that's a great idea. I'm less angry now thank you. Also, love your blog.
If anyone could tell me whether #A B# B# was correct that would be great.
I'd bet you have the right answer, since I got the same answer in two different ways. It seems it doesn't matter if you do all the possible computations for (adjacent #'s of) a row at one go (e.g. the starting setup has two), compared to making just one computation/move at a time. Making multiple moves at one go seems to make the process go faster in terms of total moves needed. In this example, it's 6 moves (multiple moves per turn) vs. 7 moves (single move per turn).
Thank you Mika <3
I got A# as the final result, not that I know that is correct.
Thank you! We will never know who is right :D
Wonderful post, Ethan. Yet one more example of why I am always excited to get this Substack in my inbox.
As fast as the models advance, I suspect there will be substantial friction in implementing them in the workplace. That is, I expect the effect of AI on the economy (GDP, productivity, employment) to be muted for the next 3-4 years. A friend of mine, a software development manager at a midsize SaaS company, is actively working to replace his low performing employees with AI. He has been running parallel experiments where he tasks an employee and also GPT3.5 to write code from detailed specs. For six months now, GPT3.5 has outperformed them considerably. His bosses, however, are resistant: laying people off is hard, and (ironically) expensive. Plus there is hesitance to trust a new technology.
What will happen, I suspect, is that people will simply not be hired. Positions will stay vacant, then be eliminated, with AI taking up the slack, either directly via agents or indirectly via covert or overt use by current employees. This effect will be especially pronounced in government, where union rules making lay offs of low performers essentially impossible.
The first sign of this, I think, is what is happening with management consulting. Ethan's study with Boston Consulting Group may have been a pebble that started a landslide: if the consultant with AI is better than just the consultant - and faster - how good could the AI be on its own?
Interesting post. Did the guy end up paying 10k to somebody? I have seen the power of prompting first hand. Building a lot of new abilities thx to emerging capabilities in my passion project uni.com.ai
I think it's "every [5 to 14]", not "ever [5 to 14]" . . .
Your persuasion example immediately reminded me of the persuasive elevator in Hitchhikers' Guide to the Galaxy!
Great read! AI breakthroughs remind me of people breaking records in sports:
● Nobody thought it was possible to run a marathon under 2h. Until someone did
● Nobody thought it was possible to deadlift 500kg. Until someone did
Crazy how things are going. Not sure how I feel about this. But I don't think anybody can stop this train now. I just hope our society doesn't turn into the movie *I, Robot*
I’m genuinly curious… When you say “we don’t know what these systems are capable of” and in the same breath write “multiple people figured out how to get AI to solve the problem within the day using prompts alone”, I feel you are begging the question here.
First of all, it looks like we’re reverse engineering based on a known outcome we’re trying to achieve. And only with very specific prompting that outcome is achieved, which means the underlying thing isn’t a result of generalization.
Secondly, who’s achievement is this really? The system, or the cleverness of the folks that find the right combination of buttons to press to make the model dance?
I’m reading the second paragraph of the introduction to open my AI presentation to the staff at my K-5 elementary school.