I ran through a calculation in detail in a different post (The Exponential Growth of Progress) that estimated time spent farming in the year 1800 verse today. In the year 1800 when the global population was 1 billion, humans collectively spent around 830 billion hours growing food. Today humans spend approximately 110 billion hours growing food - even though our population is nearing 8 billion. So, we spend ~700 billion less hours to grow food even though our population is up near 8X.
Put differently, our cost to grow food was 830 hours per person per year in the year 1800 vs. 13.75 per person per year today - approximately a 60 fold improvement.
Another anecdote I love to tell is comparing the cost of the various components of an iPhone - were you to have purchased them in the late 1980s - verse today. Mostly due to the compute (CPU/GPU) and storage (memory), you would have needed to spend in excess of $3 million to buy everything that you get when you purchase a $1,000 iPhone today. Further, you would have needed a 5,000 square foot warehouse to store the memory and computing power, and somewhere on the order of 1 million square feet to store the sum total of human information - which is now accessible from your pocket.
These are some of my favorite examples of the miracle of productivity gains. Here’s another more recent example.
Below is a link to a video on Youtube that demos a new software which allows the user to upload an image of a person or character, and then automatically creates a video of that person performing any action that has been uploaded to the library.
Keep this capability in mind.
Many companies are working on something similar for voice, including Microsoft:
Put them together, and what we have is an easy way for any person to take an image of any other person, and then make them say or do anything. This technology will be able to replace actors, it will enable people to turn any person they have a crush on into their own personal porn star, and it will enable anyone with an internet connection to make a video of their least favorite politicians doing or saying heinous things.
Were you to try and create a “deep fake” - as a random person with no technical ability 5 years ago, you would have needed to spend literally thousands of dollars to get the task finished (cost to hire someone with the relevant expertise). And, once you were done, if you wanted to make edits or create a second deep fake you were looking at a minimum of $50-$100 per hour and probably 5-10 hours of time for each iteration.
Within the next 12-24 months you will probably be able to do the same thing with a $20 monthly subscription.
Put differently, what previously would have cost on the order of $2,000+ will now cost $20 - a 100X difference.
What is Agentic AI?
ChatGPT and other generative technologies are in the process of unlocking what I will call Agentic AI.
Traditional AI did things like:
Filter spam
Decide which search results to show you on Google
Help design efficient computer chips
Match you with advertisers who are selling products you are likely to buy
For non-technical consumers (like me) - traditional AI’s usefulness mostly hid behind the scenes. Further, our relationship with things like Google’s search had the following user-experience: Ask a question » Get an Answer » YOU Act on the answer.
“Where can I buy a new pair of Nike Sneakers?” » Google provides links » Human clicks on links and determines what they want.
“How do I create a pivot-table in Microsoft Excel?” » Google provides links » Human clicks on links and then follows the instructions.
Using our example above of automating images, you could have Googled: “How do I create a deep fake?” Google would have provided links to courses on writing code, doing graphic design, using Adobe products to make illustrations - and so on…The information was there to teach you how to do it - but it would have required hundreds if not thousands of hours to master the expertise required.
ChatGPT and other giant machine learning programs are now capable of understanding the world. For $20/mo you can interact with a voice that is indistinguishable from that of a human (which did fool my grandmother), which is capable of doing everything from writing code to carrying on a lovely conversation reminiscing about past experiences (another thing I did with my grandmother - I told ChatGPT where she was from, where she lived, and then asked it to have a conversation with her - continually asking questions to keep the conversation going - and it worked beautifully). If you upload a picture of your computer’s home screen and then tell ChatGPT you want to do some task (for example, open a web browser to search for shoes) - it will be able to tell you what icon to click on and then provide step by step instructions to accomplish your goal.
We do NOT need to make any more advances on the intelligence side for Agentic AIs to become incredibly useful and ubiquitous. While intelligence will continue to improve rapidly - all we need to enable Agentic AIs to proliferate is simple. We need to create:
User interfaces
Improve efficiency and cost of required compute
Plug them into everything (APIs, the internet, etc)
That’s basically it. None of the above three tasks are hard - and literally thousands of companies from giants like Microsoft to start-ups like the ones creating the automation tool above are already rapidly addressing them.
GPT-4 and Co-Pilot (from Microsoft) are examples of Agentic AIs fast heading to ubiquity within the world of software developers. You can ask them to write scripts to crawl web pages (for example, if you were trying to find companies that use certain key words), you can upload a bunch of computer code and ask them to check for bugs, and more recently you can even upload images with flow charts describing the basic features of a website and have them spit out the HTML, CSS, etc to turn that into a reality.
Something I have done countless times over the years is ask Google how to do “X” in Microsoft Excel. For example, “How can I create a spreadsheet that automatically pulls data from a different spreadsheet?” Or, “What’s the formula that spits out text if an answer is less than zero and spits out a calculation if the number is above zero?”
With Agentic AI, you will simply be able to talk to your spreadsheet and have it do the work for you. Even better, once Agentic AIs get access at the operating system level of computers, you will be able to do things like tell them to download the monthly financials from Quickbooks and update your “Year-to-date” spreadsheet that calculates cash-flow.
On the consumer side, you will be able to do things like ask the AI to find the least expensive pair of Nike Sneakers Model “Jordan Runners Red” - and it will search the internet until it does.
Instead of asking computers how to do something - we’ll just instruct them to do those things on our behalf. And, because the interface will be voice - the productivity gains will not be limited to programmers - they will flow through to every single activity performed in the digital world.
I spent some time thinking about specific jobs that I was familiar with and then estimating how much time would be saved if they had access to Agentic AIs, specifically I thought about:
Purchasing manager (e.g. sourcing materials for construction)
Financial analyst (internal, buy/sellside)
Sales manager
Software developer
Let’s work through a hypothetical estimation of time savings for #1 above.
Purchasing managers frequently have to source hundreds of raw materials/components, each of which might be supplied by numerous vendors. Which vendor they purchase from will depend on things like:
Price
Minimum purchase quantities
Local availability
Shipping costs
Delivery date
Once a decision has been made about where to purchase from shipping has to be arranged. This usually requires logging into various portals and checking rates and when the carrier has space available. Very frequently something will have changed. For example, maybe there is a new sales rep, or prices have been updated, or the supplier needs to check whether they have local stock or if they’ll need to bring it in from across the country.
So, the purchasing manager will need to send out myriad emails collecting all of this information and then make a decision about what to buy, how much to buy, where to buy it, how to ship it, etc.
With Agentic AIs the purchasing manager might simply instruct the computer as follows: “Reach out to all suppliers of 2X4s, get quotes from those willing to sell 50 units or less but only if they’re able to ship within the next 48 hours.”
One of the suppliers sends an “out of office” response and another responds that the sales rep for their territory has changed and the new one will have to provide pricing. The AI without being prompted sends a follow up email to the new sales rep and to the “other point of contact” for the person who was out of office.
Once it has gathered all of the necessary information, the AI comes back with the best option, then the purchasing manager tells it to: “Send them a PO and make a note on the PO that we want to handle our own shipping, blind copy Bob from Shipper A and Jack from Shipper B to get quotes and estimated delivery dates”.
Having gone through the above process countless times myself, I can guarantee that the time saved would be on the order of 50%.
After running through similar exercises across a variety of professions, it seems intuitively obvious that Agentic AI will be capable of saving the average desk-worker at least 5 hours per week even before the Agents are capable of communicating at an operating system level.
My best guess is that we’ll have this level of Agency within 2-3 years.
What would 5 hours of savings per week mean for people and businesses?
For perspective, the workforce is estimated to have grown by around 8 million people over the past ten years. As the unemployment rate hasn’t changed much - we can infer that the economy was capable of placing these workers.
The total US workforce is approximately 164 million and 89 million of those jobs are white-collar. If we were to assume that all productivity gains flowed through to those 89 million jobs then it would be as if we only needed 89 X (35/40) = 77.875 million workers to accomplish the same amount of work.
I can’t envision a scenario where new jobs are created fast enough to make up for 11 million jobs becoming unnecessary in the next 2-3 years. The number is simply too big. For perspective, total job losses during the 2007-2009 great recession were about 9 million.
HOWEVER - I do not expect job losses to begin en-masse - at least not yet.
Ever since ChatGPT and Microsoft’s Code-Pilot came out I’ve been asking software developers how much time they’re saving. Most responses I get are between 2 and 5 hours. The second thing I’ve been asking is what’s happening with that additional free time? Are they able to get more done and hence do more for the company?
To date, I have had only one person tell me that they’re passing through the productivity gains to the business. In all other cases the developers are just enjoying having more time to spend with their families, play computer games, or do whatever it is they want to do. The companies are not the ones benefiting from this increased productivity. To be transparent - my sample is biased because most of the developers I know work at least part time remotely.
I should also point out that an argument could be made that the companies will benefit by having happier employees under less stress. All else equal happier employees are less likely to leave, more creative, and more efficient during the hours they’re still putting in.
I’m still trying to puzzle through what this actually means for the future of work, productivity at companies allowing remote work vs. those that don’t, etc. One thought I’ve had is that it will lead to a widening performance gap between companies that can attract and retain people who are going to work 50+ hours per week no matter what - and those where employees mostly prioritize quality of life.
If you are a top programmer there is only ONE reason you would work for a company like General Motors - you prioritize quality of life and want to work 25-30 hours a week and take full advantage of your vacation time.
If you work for Tesla or SpaceX (known to have the hardest working employees) you are much more likely to be mission focused and pass through the productivity gains to Tesla.
When will the workforce participation rate begin to shrink?
By the time Agentic AIs hit the market in the next 2-3 years they will be far more intelligent than they are now. Hence, the productivity gains they’ll be capable of will likely be greater than 5 hours per job - which was an already conservative estimate.
It may take another 2-3 years for a majority of companies (big and small) to figure out how to effectively integrate them into their workflows and adjust their expectations around employee output. Once integrated, companies will have a much better idea of the amount of work they should be able to get out of their employees using the new tools, and even if the tools keep on improving they will be able to capture at least the 5 hour productivity boost that became possible 2 years prior.
My takeaway is that we are at most 6 years from having the labor force participation rate start to shrink, never to rise again. And we could be there in only 4 years.
Personally, I can’t envision a scenario where new jobs are created beyond 6 years at a rate that isn’t overwhelmed by the job losses caused by the proliferation of Agentic AIs.
That said, for all of history this viewpoint has been wrong - so take that view with a grain of salt.
Agentic, Independent, Embodied
I want to close this post with some thoughts on independence and embodiment. Independence is what I’m calling the ability to execute broad objectives with limited guidance or oversight. For example, say that you wanted to start a social media account and grow a following with the objective of earning advertising revenue for sponsored posts. An AI with agency would be able to follow a command like: “Come up with a funny post about Father’s day and post it to my account”. But an AI with Independence would be capable of coming up with posts, responding in clever ways to posts of other users, determining whether it made sense to advertise a certain post, monitoring social media for topics that had recently gone viral and incorporating that into the content it creates - and so on. All you would need to do is tell it to build a social media account around humor and make as much money as possible - it would infer the rest.
A digital AI with independence would also be able to do things like orchestrate a DDOS attack to take down a website, or potentially follow a command like: “Start trying to hack into XYZ computer system and don’t stop until you succeed”. People like DeepMind co-founder Mustafa Suleyman have expressed fears that this level of what I’m calling independence could be less than 2 years from being a reality.
Embodiment means giving the AI a physical body with the capability to move around the real world. Until this past year there was no point trying to figure out how to build a physical shell to house a general purpose robot (meaning, one that could follow commands like: “clean the kitchen” or “prepare those items for shipping”, or “sort through that pile of stuff and throw away anything that looks like trash”). Now that the “brain” of the robot exists I believe we’ll see billions of dollars flow into developing the skeletons/bodies in the near future. To be clear, when I say the “brain” I’m referring to the intellectual part of it - not the part which handles movements/dexterity/etc - those bits I’m considering as still being necessary under the “skeleton/body” development.
It’s likely that putting Agentic AIs into bodies (humanoid or other) will end up being an easier problem to solve than self driving because they don’t need to operate in a world where edge cases cause fatalities.
When Agentic, Independent AIs become embodied -everything will change. Job losses will occur at scale, inequality will rocket as capital once again dominates labor. Thankfully, this will also cause cost-deflation across the vast majority of goods and services, from housing to food to transport to end-of-life care. My guess is sometime in the 2030s the big companies with $5-10Trillion valuations will be forced to embark - in conjunction with the government - on a marketing campaign that will successfully convince most people that an era of abundance is coming when no one will need to work to enjoy a quality of life higher than the average person has today. This will placate the jealousy of the population at large as they see the number of billionaires rocket, and as the world’s first trillionaire’s are created.
Great article, Ben! And a little frightening to imagine all the ways these things might play out.....
Thanks for sharing your thoughts. Highly interesting and deeply unsettling.