Fermi Paradox of GenAI

I assume most people reading this are familiar with the Fermi Paradox: “the discrepancy between the lack of conclusive evidence of advanced extraterrestrial life and the apparently high likelihood of its existence.” This paradox has spawned a host of putative explanations as to why we don’t see life everywhere in the galaxy, from the nature of intelligent life is to destroy itself or alien life is too incomprehensible, to extraterrestrial life is rare or nonexistent (e.g., the Hansen Great Filter).

With every new release of state of the art GenAI products, the capabilities of these models seems to grow. It is now all but impossible to go through the day without seeing incredible high profile stories about how AI is soon to take over the drudgery of our lives, collaborate to make us all better workers and more creative, or even displace many of us from our jobs. Admittedly, there has been astounding progress in GenAI on many different fronts. GenAI is mastering tough benchmarks, acing PhD level exams, helping code, outperforming doctors, and creating high quality pictures and movies. However, I do not (yet?) see GenAI taking over in the business world, or really even with a strong foothold towards a market-wide takeover.

When Will We See the GenAI Breakout Performance

GenAI has an Adoption Paradox at the moment; To pattern it after the Fermi Paradox, there is a discrepancy between the lack of widespread adoption of GenAI in economy and, for many believers, its apparently high level of capabilities and intelligence. Why is this a paradox? Well, over the past year or more, each new model release has been accompanied by amazing demonstrations and dramatic improvements in performance on benchmarks, including most recently, OpenAI’s latest model release, o3 and its 87.5% performance on the ARC-AGI benchmark.

So, those who believe current GenAI is at a performance level where it has already achieved human level or higher performance on economically meaningful tasks, it’s reasonable to wonder why we do not see GenAI being adopted everywhere -- right now, or in the near future.

Adoption Hesitancy or Interview Failure

What are the Fermi Paradox-like explanations then for why we do not yet see signs of substantial and widespread GenAI adoption across the economy? Maybe there is hesitancy or inertia that is delaying the use or adoption of GenAI. A survey of 644 companies reported by Gartner earlier this year found that the top reason given for their hesitancy in adopting GenAI was “the difficulty in estimating and demonstrating the value of AI projects.” This can be spun by supporters and detractors to support their view of GenAI market penetration. Maybe more time is needed for companies to find productive uses or overcome their bias or hesitancy to use AI. However, “difficulty in estimating and demonstrating the value” of a new technology is, well, a very common reason that technologies and companies fizzle.

Clearly, GenAI companies could attack this issue head on and demonstrate the cost-effectiveness and utility of their product by using specific job or function based benchmarks, or even using these products in their own companies. I have no doubt that AI coding assistant companies use their own products at work, all day, every day.

Another possible explanation for the lack of adoption at the moment is that while GenAI is progressing by leaps and bounds, state of the art GenAI is just below the threshold of capabilities or performance to really get a foothold in business. Maybe 87.5% on the AGI benchmark isn’t quite there, but somewhere above 90% will be enough. Again, this clearly could be true and only time will tell. GenAI models are obviously capable of performing certain economically valuable tasks, but maybe their level of performance is not yet high enough to overcome barriers to adoptions is the wider economy (e.g., error rate or performance relative to humans).

The GenAI Company GenAI Hiring Spree

It is very easy to imagine that a combination of these issues and potentially others is at play. Some combination of lack of value proposition, confidence in the technology, trust in the results, and general industry inertia may all combine to slow the general adoption of GenAI in the market. Again, though, I would expect that other than capability of the technology, companies in or near the GenAI industry would themselves be very early adopters of the technology to the extent its performance made sense. This is exactly what I imagine is happening in companies making GenAI-based coding assistants; I have little doubt that these companies are making economically valuable use of their own products. Maybe a first place to investigate to see if the adoption paradox is real, is GenAI companies themselves.

Relationship Between Model Performance and Market Penetration

It is interesting to imagine what a realistic market adoption and penetration profile for AI as a function of their performance and cost. As discussed above, clearly some portion of the adoption is based on factors not directly related to model performance (e.g., inertia opposing new technology or longer term evidence of positive ROI). However, the longer businesses and people are exposed to GenAI, the more I would expect adoption decisions to be based on performance rather than, say, lack of trust or fear of AI in general. There are likely areas of utility for models, such as coding, picture or video generation in which users and companies may have a higher comfort level in adopting GenAI technology because the output can be assessed directly and often quickly by the user. There may be many areas of adoption of GenAI models that fit this pattern.

Current GenAI Does Have Use Cases

I want to be clear about what I am saying and more importantly, what I am NOT saying:

I am NOT saying current GenAI has no use cases

I am NOT saying people don’t use GenAI productively every day

I am NOT saying there will not be big breakthroughs in AI in the future

I am NOT say GenAI will never be generally useful

My main point is that the current GenAI use cases do not yet have the combination of scale, cost saving, and productivity enhancement that matches the hype or the level of investment in the technology. It’s easy to cast this observation aside by saying wait 6 months, or a year or two years. Yes, I agree, there could be some advance in the next months or years that changes everything. However, what happens in 2 years when GenAI has made inroads into some markets as a better, more useful expert system, but is still not on track to be a productivity resource in the general market? Will people still say, wait 2 more years or 5 more years? I think the economically important question is: how useful will GenAI be for the economy vs the money and time invested in it to see that return?

Conclusion

So, to circle back to the beginning, my point is that there is a discrepancy between the lack of widespread adoption of GenAI in the economy and, for many believers, its apparently high level of intelligence and utility. We will see how this gets resolved over the next few years. If GenAI represents true general intelligence, it should quickly win over the market as an economically viable tool. In the absence of that, genius level performance on benchmarks may not be enough to justify continued patience.