[BostonCard]

Pasted from a post I made internally back in June when Amodei first started making this prediction.

So, I’ve been thinking about the interview a lot of late and while I may be one of the people completely sticking my head in the sand on this, I think Amodei has fallen prey to a few cognitive traps here.

First, extrapolation of recent progress.  What has happened in the last five years in the AI space is nothing short of astounding.  However, extrapolating that level of progress going forward is not certain, and innovation tends to happen in fits and starts, with periods of revolutionary change followed by long periods of evolutionary change.  Think of the sea change when the internal combustion engine became available, and comparison to the previous technology (horses).  However, in the ensuing 150 years since Otto first developed the Otto Cycle Engine, progress has been extensive but much more about refining and improving the engine than evolutionary changes.  While I am not an expert in the field, I get the sense that since GPT-3 came out five years ago, subsequent improvements have been more of the evolutionary variety than continued revolution and it appears that progress may be slowing, wih computational limits and training data availability becoming bottlenecks.

Two, it is often said that the future is already here, it’s just unevenly distributed.  I think Amodei mistakes capability with deployment.  In other words, he looks at his models and says “AI can already do [X} and will soon be able to do [Y}“, but it is a very long way from knowing that the technology could do it to the technology actually doing it.  Companies deploying AI will need time to develop their requirements, chose amongst competing systems, implement pilot programs to test them, get them adopted by users, all of which tends to happen pretty slowly, especially in large companies with legacy systems.  As I mentioned, it was five years ago that the paper describing GPT-3 was published, but the introduction of [our internal AI system] and the roll out of Gemini has only recently happened.  And while the [internal communications channel] is really a hive of activity and it is great to see people experiment and deploy it for their use cases, I think we are a very long way from the type of activity that would make 20% of employees redundant.  To give a good example of the diference between capability and deployment, consider the electronic switchboard and telephone operators.  This article [https://www.richmondfed.org/publications...ic_history] is quite interesting.  It took only 16 years after Alexander Graham Bell patented the telephone before an automatic switchboard was first deployed.  18 years later, in 1910, only 300,000 subscribers out of 11 million had automatic service, and employment in the number of operators continued to increase from around 178,000 in 1920 to about 342,000 in the middle of the century, and was still 250,000 in 1960, 68 years(!) after the first electronic switchboard was deployed.

Lastly, and I think the biggest fallacy that Amodei falls prey to is the lump of labor hypothesis.  That is the idea that there are a number of fixed tasks to be done, so that if you replace them (for example, outsourcing to another country, using immigrant labor, or automating them) that the inevitable result is unemployment.  Time and again, this fallacy has been shown to be incorrect.  Not only do these improvements historically not lead to increased unemployment, improved productivity is what drives employee wages.  There is another good example in history, the development of the ATM.  It turns out that while many people predicted that the widespread deployment of ATMs would lead to fewer jobs for bank tellers that they were replacing, paradoxically employment of bank tellers actually increased, as shown below.



Why?  Well, while the number of tellers per branch went down substantially, the number of bank branches expanded dramatically, because now it was cheaper to run a branch than it was prior to ATMs.  The nature of the job changed a bit, and tellers were relieved of routine stuff such as depositing checks and withdrawing cash (which was now done by ATMs), and were able to be used to build relationships and open accounts, activities that were much more valuable to the banks.

What does it mean for a company like [my company]? Let’s imagine that due to AI all of us can all of a sudden be 20% more efficient in what we do.  I suppose one thing that [the company] could do is lay 20% of the company off and return the money to shareholders.  But there is not a fixed pot of drugs to be developed, and an alternate use of the 20% of excess productivity is to work on 20% more programs.  With a 20% improvement in efficiency, suddenly some programs that may not have been viable to pursue because aNPV was marginal, might all of a sudden be viable because it requires 20% fewer resources to pursue.  Maybe we pursue an additional indication for a drug already in our portfolio in parallel instead of gating it.  The point is that if we were 20% more efficient, it might not mean that 20% of us are let go, but rather that we have a chance to work on 20% more interesting programs.  In fact, as with the bank teller example, if we are 20% more efficient overall, the cost/benefit balance of adding employees becomes favorable, since their extra productivity may make taking on the extra FTE to be worth it.  Now, of course, it is never that simple, and AI will have variable effects on productivity, meaning that the effect will be unevenly distributed, which means there will be different productivity bottlenecks over time, so the relative job composition in the company will probably change.  I hope that impact is minimal, and I certainly hope that my job is made more productive rather than redundant by AI, but it is a fine line between the two.

Apologies for the rambling response; I’m curious what other people think.  Do I have my head in the sand about AI?  Am I too pessimistic about the pace of deployment?


BC

[BostonCard]

Pasted from a post I made internally back in June when Amodei first started making this prediction.

So, I’ve been thinking about the interview a lot of late and while I may be one of the people completely sticking my head in the sand on this, I think Amodei has fallen prey to a few cognitive traps here.

First, extrapolation of recent progress.  What has happened in the last five years in the AI space is nothing short of astounding.  However, extrapolating that level of progress going forward is not certain, and innovation tends to happen in fits and starts, with periods of revolutionary change followed by long periods of evolutionary change.  Think of the sea change when the internal combustion engine became available, and comparison to the previous technology (horses).  However, in the ensuing 150 years since Otto first developed the Otto Cycle Engine, progress has been extensive but much more about refining and improving the engine than evolutionary changes.  While I am not an expert in the field, I get the sense that since GPT-3 came out five years ago, subsequent improvements have been more of the evolutionary variety than continued revolution and it appears that progress may be slowing, wih computational limits and training data availability becoming bottlenecks.

Two, it is often said that the future is already here, it’s just unevenly distributed.  I think Amodei mistakes capability with deployment.  In other words, he looks at his models and says “AI can already do [X} and will soon be able to do [Y}“, but it is a very long way from knowing that the technology could do it to the technology actually doing it.  Companies deploying AI will need time to develop their requirements, chose amongst competing systems, implement pilot programs to test them, get them adopted by users, all of which tends to happen pretty slowly, especially in large companies with legacy systems.  As I mentioned, it was five years ago that the paper describing GPT-3 was published, but the introduction of [our internal AI system] and the roll out of Gemini has only recently happened.  And while the [internal communications channel] is really a hive of activity and it is great to see people experiment and deploy it for their use cases, I think we are a very long way from the type of activity that would make 20% of employees redundant.  To give a good example of the diference between capability and deployment, consider the electronic switchboard and telephone operators.  This article [https://www.richmondfed.org/publications...ic_history] is quite interesting.  It took only 16 years after Alexander Graham Bell patented the telephone before an automatic switchboard was first deployed.  18 years later, in 1910, only 300,000 subscribers out of 11 million had automatic service, and employment in the number of operators continued to increase from around 178,000 in 1920 to about 342,000 in the middle of the century, and was still 250,000 in 1960, 68 years(!) after the first electronic switchboard was deployed.

Lastly, and I think the biggest fallacy that Amodei falls prey to is the lump of labor hypothesis.  That is the idea that there are a number of fixed tasks to be done, so that if you replace them (for example, outsourcing to another country, using immigrant labor, or automating them) that the inevitable result is unemployment.  Time and again, this fallacy has been shown to be incorrect.  Not only do these improvements historically not lead to increased unemployment, improved productivity is what drives employee wages.  There is another good example in history, the development of the ATM.  It turns out that while many people predicted that the widespread deployment of ATMs would lead to fewer jobs for bank tellers that they were replacing, paradoxically employment of bank tellers actually increased, as shown below.



Why?  Well, while the number of tellers per branch went down substantially, the number of bank branches expanded dramatically, because now it was cheaper to run a branch than it was prior to ATMs.  The nature of the job changed a bit, and tellers were relieved of routine stuff such as depositing checks and withdrawing cash (which was now done by ATMs), and were able to be used to build relationships and open accounts, activities that were much more valuable to the banks.

What does it mean for a company like [my company]? Let’s imagine that due to AI all of us can all of a sudden be 20% more efficient in what we do.  I suppose one thing that [the company] could do is lay 20% of the company off and return the money to shareholders.  But there is not a fixed pot of drugs to be developed, and an alternate use of the 20% of excess productivity is to work on 20% more programs.  With a 20% improvement in efficiency, suddenly some programs that may not have been viable to pursue because aNPV was marginal, might all of a sudden be viable because it requires 20% fewer resources to pursue.  Maybe we pursue an additional indication for a drug already in our portfolio in parallel instead of gating it.  The point is that if we were 20% more efficient, it might not mean that 20% of us are let go, but rather that we have a chance to work on 20% more interesting programs.  In fact, as with the bank teller example, if we are 20% more efficient overall, the cost/benefit balance of adding employees becomes favorable, since their extra productivity may make taking on the extra FTE to be worth it.  Now, of course, it is never that simple, and AI will have variable effects on productivity, meaning that the effect will be unevenly distributed, which means there will be different productivity bottlenecks over time, so the relative job composition in the company will probably change.  I hope that impact is minimal, and I certainly hope that my job is made more productive rather than redundant by AI, but it is a fine line between the two.

Apologies for the rambling response; I’m curious what other people think.  Do I have my head in the sand about AI?  Am I too pessimistic about the pace of deployment?


BC