The year is 1956. You are a researcher who works at International Business Machines, the world’s tabulant machines, which has recently diversified in the new field of electronic computers. He has commissioned to determine what purposes, exactly, his clients are using the huge IBM mains.
Register here to explore the big and complicated problems facing the world and the most efficient ways to solve them. Sent twice a week.
The answer turns out to be quite simple: computers are for the army and only for the military. In 1955, the previous year, with much, the largest source of individual income for the IBM computers division was the Sage project, an initiative of the Department of Defense that takes in IBM with the creation of a computer system capable of providing early warnings throughout the United States if Soviet bombers with nuclear weapons attack the country. That brought $ 47 million in 1955, and other military projects raised $ 35 million. Meanwhile, programmable computers sold to companies brought $ 12 million.
You send a memorandum to your boss explaining that the impact of computers on society will be mainly to give the United States an advantage in the Soviet in the Cold War. The impact on the private sector, on the contrary, seems less. Remove yourself in your chair, you turn on a cigarette and reflect on the glorious future of the Industrial Defense Complex.
Of course, I would be totally wrong, not only in the distant future but in the very immediate. This is how the income of each of the IBM computer divisions in 1952 to 1964, compiled by the veteran of the company Emerson Pugh in his book in his book IBM building:
Emerson Pugh, IBM building.
Just two years after 1956, the Ble Computers program sold to private companies had matched SAGE as a source of income. The year after that, the private sector was bringing as much as the military in general. By 1963, not even a decade after the 1955 data he was looking at, the army seems to be a rounding error next to IBM’s private income revenues, which have grown to account for most of the company’s income in the United States.
What can we learn from how people use AI at this time?
This week, the impressive teams of Openai economists and Anthrope launched large and designed reports on people who are using their AI models, and one of my first thoughts, “I wonder what an IBM reports how people like.” (Discllosure: Vox Media is One of Several Publishers That has signed partnership aggregates with openai. Our Reporting Remains Editorialy Independent. Also, Future Perfect is Feded in Part By The Bemc Foundation, WHOE MAJOR FUNDER WAS to Donly an enormor an energy has Havey have an enerry have an enerry have an enerry have an an an an enhropic Alsye Haverly Aleye Haverly Aleye Haverly Andyty Aleye Haverly Andyty a Antoly Antoly Antoly Antoly Antoly Antoly Antoly Antoly Antyty Andytytistor;
To be clear: the level of attention that the teams of the companies of AI put in their work are many, many orders of magnitude greater than that of our fictional IBM analyst. Income is not the best measure of interest and real use of the client; Everyone knew in 1955 that computers were quickly improving and their uses would change; IA companies have access to a real -time variety of data on how their products are used that would have caused the Watson family to work with IBM.
That said, I think IBM’s example is useful to clarify what, exactly, we want to get out of this data child.
The reports of AI companies are more useful to give us a snapshot in time, and a story recently about a couple of years, or what the child’s child is being Chatgpt and Claude. It is possible that I have read my colleague Shayna Korol in the perfect newsletter for the future of Wednesday, presenting the operai findings, and I also recommend the co -author of the study and the summarized publications of Harvard professor David Deming. But some great and non -trivial things that I have learned from the two reports are:
- Absorption is shooting: Chatgpt has gone from 1 million users registered in December 2022, 100 million people who use it at least weekly by 2023, to more than 750 million weekly active users now. If the number of messages sent to him continues to grow to the current rhythm, there will be more chatgpt consultations than Google searches at the end of next year.
- Both Operai and Anthrope discover that the richest countries are using AI more than poor (it is not a surprise there), but OpenAi finds intriguingly that middle -income countries such as Brazil use chatpt almost as much as.
- The most important use cases for Chatgpt were “practical tips” such as how coughing or tutoring/teaching (28.3%consultations), editing or translating or generating text (28.1%) and search style information consultations (21.3%). Anthrope uses different descriptive categories, but finds that people who use Claude.AI, the chatgpt type interface for their models, commonly use it for computer and mathematics problems (36.9% use), while a use of use of use use of use use of use use use use use of use use.
But I am greedy. I don’t just want to know the descriptive facts of the first order on how these models are used, only although those are the son of the questions these documents, and the internal data that OpenAi and Anthrope collect more generally, can answer. The questions that I really want to answer about the use of AI and its economic ramifications are more like:
- Will human labor and complements or substitutions in five years be? Ten years? Twenty?
- Will salaries upload because the economy is still a bottleneck about things that only humans can do? Or will they collapse to zero because those bottlenecks do not exist?
- Will AI enable the creation of “geniuses in data centers”: ia agents who carry out their own scientific research? Will this lead to the stock of scientific knowledge about the world to grow faster than ever? Will that lead to explosive economic growth?
Many people ask these questions, and an impressive amount of theoretical work has been done in economics that is already on the subject. I have found this set of conferences and paper offers on the issues of economist Philip Trammell very useful.
But that theoretical work is mainly in the form of “What are some concepts that we could use to make sense of what is happening or will happen shortly?” – It’s the theory, that’s the point! – And so leaves a greedy and impatient man as Myelf without good answers, or even partially good conjectures, in the previous questions. It is a place where I want a good empirical research to give me a sense of which theoretical frameworks correspond to the earth’s reality.
My fear is that, for reasons that explains the parable of IBM, the empirical details about how AI can now be cheated on us how it will be used in the future, and in this regard they must be important effects in our lives. If you cryogenically freeze our IBM analyst in 1956 and resurrected them today to analyze OpenI and anthropic reports, what would you say about the most speculative questions above?
They could point out the fact that the Chatgpt study found half of all messages correspond to a fairly small number of “work activities”, as traced by the Labor Department, such as “documenting/recording information” and “making decisions and solving problems.” Those are great categories, but people have to do much more in their work that does not fall under them. Our IBM analyst could conclude that AI is only automating a fairly small part of work tasks, which means that human work and AI will complement each other in the future.
On the other hand, the analyst could look at the anthropic report that found the cases of “automation” use (where only aging tells him and does all the task, perhaps with the pericodic human feedback) are the proni -tropic bacon of Trelbon, a confidential committee. Specific routines enabled to claude that the cases of use of “increase” (where he asks Claude or to learn, etc., and work in concert with him). The increase still constitutes a majority of use on the Claude.AI website, but automation participation is also growing there. Our analyst could see this and conclude that AI and human work will end as substitutions, since Claude users use it less as a crisp than as an agent who works on his own.
All these conclusions would be, I think, premature to the point of the recording. That is why, for their credit, the authors of the OpenI and Anthrope reports are very careful about what they do and do not know and can and cannot infer their work. They do not claim that these findings can inform us about the medium effects or the long term of AI on labor demand, or the distribution of economic growth, or professionals that will be most affected by ai-just, although that is precisely verse.
Why AI is different from corn (I promise that this makes sense)
So allow me to end up focusing on something that the reports tell us that it is, I think, important. One of the oldest findings in the economy of innovation is that new technologies have often, often “spread” through the economy.
The classic role here is Zvi Griliches in 1957 about the spread of hybrid corn. Hybrid corn was not a specific product, but a particular approach to reproduce corn seeds optimally for specific soil in specific areas. Once a few farmers in a state adopted hybrid corn, ASA and the sub -direor handle seemed to be incredibly fast. Look those curves S!
ZVI Griliches, “Hybrid Cornillo: an exploration in the economy of technological change”
But although the diffusion within the individual states was rapid, the diffusion between the states not. Why did Texas need a decade after the increase in hybrid corn in OWA to realize that you could greatly increase yields? Why did it seem to reach a much lower roof or a use of 60-80%, compared to Universal Upake in Iowa? You also see these delays when you look at cases such as electricity and data sets that cover a wide range of inventions.
Something that anthropic and Operai data tell us quite clearly is that dissemination delays for ia are, according to historical standards, very short. The adoption of this technology has a fast bone, in fact faster than previous online products such as Facebook or TickTok, much less hybrid corn.
It adapts to Purphto’s general technologies such as electricity or calculation of years or decades to spread through the economy, which limited its benefit for a while, but also gave us time to adapt. We are probable that we do not have that time this time.