Reaction to "AI as a Normal Technology"

This is a reaction on the article: https://knightcolumbia.org/content/ai-as-normal-technology.

The normal technology thesis is particularly convincing to us since it decouples technological and research development from the real-world economical impact of the technology in terms of applications, use-cases and adoption.

Technological progress does not necessarily imply the economic impact, the second one depends on applications, business cases and value, integration into the current process and eventually adoption.
However, the process of development and adoption happen on different speeds, AI evolves on the pace of research while adoption happens on the pace of humans.

This leads to the feeling that organizations are always behind, each couple of months there is a new model or a new capability, and there is constant pressure to adapt to the changes, yet adoption happens much slower since organizational change needs the change of processes, responsibilities, incentives, and potentially even cultural shifts.

Adoption is one of the key bottlenecks of AI

As far as we see it now, adoption is the bottleneck rather than technology itself.

AI models might work really well in the lab conditions, however adoption is the moment when AI meets the real world, which is diverse, uncertain, heterogeneous and full of exceptions. Most importantly, AI cannot adopt itself, humans adopt AI.

This is well seen in the field of software development: coding became much faster thanks to AI, however software product development did not speed up at the same pace, the bottleneck shifted from implementation towards making decisions on what should be implemented and what would bring value to the business.

The key question is increasingly becoming not "how do we build this" but "what should we build and why".

That is why we completely agree with the thesis that there is an important distinction between invention, innovation, adoption and diffusion, better technology doesn't imply useful products.

We also think there might be a bottleneck in the area of invention, since the transformer architecture, most of the developments in frontier AI were made through scaling of the computational power and the amount of data rather than through the discovery of a new fundamental invention. The fact that some Chinese models became competitive with lower amounts of computational power suggests that more sophisticated research might bring substantial gains.

High-risk applications will adopt AI slowly

We also agree that adoption will be slower in areas where mistakes have severe consequences.

AI cannot be adopted in such domains as healthcare, law, finance and others in the same manner as AI for writing assistants or coding tools. Even if AI becomes much more reliable than today, I think that people will still be cautious, not necessarily because of lack of correctness, but for human bias and ethics.

For instance, an AI judge that is statistically better and less biased than a human judge, would probably be considered morally wrong by many people to sentence another human being.
Thus, the key problem is not only reliability, but also ethics and responsibility.

AI may accumulate power indirectly

There is a certain point, in regard to which we are less convinced of the argument about "normal technology".

we agree that AI does not automatically get any formal authority, but it is capable of accumulating power indirectly.

People already outsource parts of the cognitive work to AI. A person with decision-making authority may increasingly start relying on AI to analyze the situation, to propose a course of action and finally recommend a decision. Formally the person will make a decision, but the real source of this decision will become unclear.

This is already visible in the software development workflows: business analyst uses AI to write a user story, developer uses AI to implement it and QA uses AI to generate test cases. The role of people in this workflow decreases and most of the cognitive work is being outsourced.

The more people outsource their critical thinking, the more power AI will accumulate.

AI being strong for augmentation, rather than replacement

The paper talks about how current AI systems look like they are better at augmenting professionals rather than replacing them.

This might be true, but there is no guarantee that this will stay the case.
Optimists like to draw connections to the industrial revolution, where lots of manual labor was replaced with other jobs that were nicer for humans in the long term. This point is somewhat implicitly made through the quote "humans will adapt, and will focus on tasks that are not yet automated, perhaps tasks that do not exist today".
There are two concrete problems with these points that we would like to discuss.

1. There was lots of human suffering in the transition period

When the industrialization started factory owners got lots of leverage over employees.
This leverage was used to mistreat the proletariat by working them an inconsiderable amount of hours, ignoring any safety regulation and not paying them enough to live.

These conditions were only improved with violent and deadly protests, which eventually lead to things like a minimum wage, the 40-hour work-week, etc.

If there are new jobs for humans (point two tries to put that into perspective) a transition into those jobs is in no way guaranteed to be beneficial to workers. Historically the divide between the poor and wealthy grew considerably in those times (industrialization, invention of the commercial microchip, the internet).

2. There is no guarantee that new jobs will be created

A point that is often left out when discussing the topic of a post-AI job market is that there is no guarantee that LLMs will create new jobs. While that was the case for previous technological advancements, this one is different in nature.

All industrial shifts replaced an aspect of human labor and automated it (like physical labor in the industrial revolution, algorithmic thinking with the invention of the computer), but always left humans with capabilities that couldn't be replicated with the technology of the time.

We argue that the last capability where humans are needed is reasoning and thinking. Most white-collar work relies on a human doing the thinking so he can operate the machines that automated away all other aspects of his job.

This non-replaceable skill of reasoning and thinking about things is now the area of automation, shown by the rise of "meat-proxies", a group of workers that just pass information between AI-agents and act as an (inefficient) proxy of information. Those workers often work in jobs where a lower level of reasoning is needed and where LLMs are already more capable in most situations to take the wheel on reasoning.
An analogy for this is "The horse didn't get a new job after the invention of the automobile, it became obsolete".

AI is not a new species

On the other hand, we completely reject the idea that AI should be treated as a new species.

From my perspective, AI remains a tool, although a very powerful one. A human being using a tool can usually be more capable than the tool alone.

Similar to electricity or the Internet, AI is a horizontal technology that is capable of transforming many domains. But there is an important difference between AI and other horizontal technologies, AI operates directly via language.

Language is intrinsically human. Being capable of communication in natural language, AI generates anthropomorphism and creates a perception that AI is more threatening than other technologies.

The largest risks still come from humans

Overall, we believe that the largest risks come from the way how people use AI.

AI does not require any intentions in order to create a threat, a powerful tool can be dangerous when used by malicious people.

Discrimination is an interesting example, AI is learning from the historical data, however the historical data contains historical biases. Therefore we want AI to learn from the past, but not to reproduce all of the values from the past.

That is why we expect AI not only to learn patterns, but to incorporate values of the modern society.

AI creates a false impression of competence

Another issue we would like to discuss is the way AI affects the learning process.

In my understanding, knowledge resembles a tree with branches. AI will let you develop the knowledge in a particular area really fast.

However, you will run into problems when you start adding more branches to the knowledge you have just obtained. In case you did not fully assimilate the knowledge of the topic, it may become difficult for you to distinguish right answers from mistakes made by the AI.

In this case, it will create a fake impression of competence in the topic.

Where we disagree the most

Recursive self-improvement

It is in the area of recursive self-improvement.

We do not think that the concept of AGI is very useful. But we do think that a transition to the state of AI capable of self-improvement is possible, and this transition will bring a new challenge in the form of control problem.

We are therefore convinced by the idea of AI as normal technology, but with a condition, as long as humans are the drivers of its development and adoption.

The key questions is not only how capable AI becomes. It is also how much thinking, judgment and responsibility humans outsource to AI.

Self regulation

The paper talks a lot about how AI companies will be forced to self regulate through the forces of the open market, but later disproves it by precedent, talking about the meat-packing industry, railroad expansion, etc.

The way the industry stands at the moment, it seems that an oligopoly of a few competitors (currently OpenAI, Anthropic and Google) will dominate the market, partially by being the only players with access to enough compute.

This will (historically shown) allow very reckless behavior, as consumers don't have alternatives.

Additionally: The consumers wont even be the one that incur the harm of reckless roll-outs, which is different than the comparisons that the authors drew to the autonomous driving industry.
A consumer using e.g. an Anthropic model for programming tasks wont swap to a competitor (of which there are few) because someone thousands of kilometers away used it in harmful ways, like mass-posting propaganda.

Owning the new means of production need more capital

The maybe most important topic the report never even considers is the fact that a few AI companies now own the primary means of production (at least for software).

Even without any significant advances in the technology the means of production (or the enabling of productivity) have shifted from resources that can be owner by workers (software, hardware, etc.) to include LLMs that smaller businesses cannot own, as they are incredibly capital intensive (even assuming that the competition doesn't need to train their own frontier model).

A startup can no longer just buy their employees a laptop for work, they now need to pay for access to AI models that they can never own themselves. This leads to a situation where small companies are reliant on large AI companies for their productivity, causing a direct disadvantage in any business model against the large AI companies.

For example: Your small business operates some sort of SaaS. If any of the AI companies ever want to expand laterally into that market they have a collection of direct advantages:

1. They know everything about your business model from the data that you sent them. They even know what you are building next.
2. Their inference is much cheaper, leading to a more efficient roll-out of features.
3. When you build features you pay for their inference and give them even more capital to perfect your competition.

All of these disadvantages are likely to lead to a loss of market share for you, while enriching the AI company in many separate ways. Long-term this can lead to an AI company forming lots of subsidiaries that have cheap inference, for the sole reason of disrupting every market to enrich themselves.