AI is not mid - a response to Dr. Cottom’s NYT Op-Ed
In a recent
New York Times op-ed, Dr. Tressie McMillan Cottom critiques what she calls the "mid tech" of artificial intelligence - technologies that promise transformation but deliver only mediocre improvements. She argues that AI's current applications are mundane: generating meal plans, managing calendars, and writing emails "that no one wants."
Dr. Cottom makes several important points. She's absolutely right to be skeptical of the breathless hype around AGI and super-intelligence that will make our world unrecognizable overnight. She's also correct in her criticism of tech proponents who suggest AI could replace education and expertise entirely.
However, I think Dr. Cottom's framing misses something important about what's happening with AI right now. Even the current "mid" AI technology will be more transformative than it initially appears.
As someone who studies ecological dynamics, I find it helpful to view technological change through that lens.
Imagine a marble rolling on a hilly landscape. The marble will naturally settle into valleys or low points - these are what we call "attractor states." The hills and barriers between valleys are "constraints" that keep the marble in its current valley. When a new technology changes the constraint landscape (reshaping the hills and valleys), the marble may end up rolling down into an entirely new valley, far away from its previous position. In this way technologies that may appear insignificant at first, when they interact with the complex dynamical system of our world, can cause powerful changes to society at large. In Ecological Dynamics parlance, this is a phase transition into a new attractor state.
The history of smartphones offers a perfect case study in how technologies reshape constraint landscapes and drive emergent organization.
The smartphone changed the constraint landscape in subtle but profound ways. It made information immediately available from any place. It integrated multiple sensing and communication capabilities (GPS, camera, microphone, touchscreen) into a single, relatively cheap device.
Initially, these changes manifested in simple ways - checking email on the go, getting directions without a paper map. In the early days, it would have been easy to dismiss them as something that would only “mildly reshape the way white-collar workers work”. Looking at the initial iPhone capabilities in isolation, you might reasonably conclude it was "mid tech" - just an incremental improvement on existing devices.
But as adoption increased and the device spread throughout society, we observed more substantial shifts around this technology. More users led to more apps, which attracted more users, which incentivized better mobile infrastructure, cheaper hardware, etc…
Many industries reorganized around the fact that smartphones were now ubiquitous. Entirely new behaviors and services emerged that couldn't have been predicted from the initial technology: location-based dating, real-time ride-sharing, social media centered around photo sharing. Music and entertainment industries were reshaped around smartphone consumption - leading to an emphasis on streaming, podcasting and consumption of media through smartphone screens and speakers.
Society gradually settled into fundamentally different patterns of social organization, communication, and information consumption that would have been impossible in the pre-smartphone landscape, and all of this was the result of a relatively small change in the actual technology.
It's critical to note that these phase shifts aren't inherently positive. Attractor states don't care about whether they're a net benefit to society or not - they're simply where the marble rolls when the constraint landscape changes. Smartphones allow one to navigate a new city without ever feeling lost. They make your uncle is a bit less likely to bluster since a fact-check is readily available in everyone’s pocket. They also brought us swiping dating culture, doom scrolling and a host of other problems.
I think about recipe sites as a simple example of what's changing. For a while, some people could have a business by bundling things that people wanted (recipes) with things that were beneficial for the content creators (ads, subscriptions, attention). Now people can get AI to get the recipe without having to look at the ads or scroll past the life story.
Perhaps a more impactful example is reading a research paper. One can now have an AI agent summarize the findings, rewrite sections while explaining certain terminology, or go and ingest a set of related papers and generate a comparative analysis.
Personally, I already feel differently about picking up a technical book — "why can't I just shove this into an LLM so I can have a conversation about it?"
Transformer-based AI has fundamentally changed how we interact with information:
Unbundled and non-linear information consumption: Being able to skip, summarize and compare content - we're no longer bound to consume information in the packages it was created.
Reducing barriers to information: AI creates summaries, explanations, and targeted answers from dense material that previously required specialized knowledge to interpret. It also substantially reduces language barriers.
Bridging the gap around image and audio processing and generation: Transforming from visual or audio information to text and back is now seamless, which makes these new modalities far easier to integrate with technologies like search.
Human language driven tool use: AI agents can increasingly "do what I mean" rather than just "do what I say," representing a profound shift in human-computer interaction. We no longer have to communicate with software in terms of pre-defined instructions. Within certain domains, AI agents can interpret a human instruction and appropriately apply a set of tools to accomplish a given task. An immediate application is home and phone assistants that actually work.
Interacting with information is never going to be the same. This is already changing search, advertising, copyright, and many other fundamental aspects of our society.
An example of a phase transition happening in real-time is within the hiring process. Within Software Engineering hiring, this transformation has occurred virtually overnight.
Hiring teams now get hundreds of AI-generated resumes that are custom-tailored for each role. While many early applications of AI to this space are still quite unsophisticated and
can be distinguished from real resumes, it is already possible to generate a high-quality resume using the current technology and an appropriate prompt, so getting a useful signal out of resume screening is becoming more and more difficult. A hiring team can't manage to phone-screen hundreds of candidates; we simply don't have the resources.
Meanwhile, take-home coding tasks or coding interviews, which have historically been small, self-contained, routine problems, are now a lot less useful as a signal of technical expertise since AI is exceptionally good at solving these sorts of challenges.
The entire tech hiring industry (and likely hiring for other industries) is rapidly re-organizing around this new reality. The challenge of establishing that the candidate is a real human and genuinely has the credentials they claim to have is being felt acutely by everyone currently trying to hire for tech. Technical evaluation of candidates is shifting drastically to "can you talk about your code" and away from "can you produce a technical artifact." All of this places more emphasis on other signals like networks, referrals, degree-awarding institutions, and may lead to a resurgence of hiring agencies that can outsource this work.
This transformation isn't just about making the hiring process slightly less efficient or slightly more efficient - it's about the entire system reorganizing into a qualitatively different state with new behaviors, practices, and institutions emerging.
As a software engineer, I've experienced another constraint removal firsthand.
Coding assistants dramatically lower the barrier to entry for new programming languages, frameworks and code bases. Before AI, shifting to a new programming language or framework required learning syntax, studying documentation, and making a considerable investment becoming productive. Now, I can start programming in a new language right away. Initially, I am heavily relying on AI for generating valid syntax, fixing syntax mistakes I make, and providing in-context explanations. This was somewhat possible before through liberal use of stackoverflow and google, but the new reality is that you can learn much more effectively in the context of the thing you’re actually trying to do, instead of relying on more general resources.
Here's where I think Dr. Cottom's warning is important: you can only get so far with critical reasoning skills and AI alone. As she pointedly notes, "A.I.'s most revolutionary potential is helping experts apply their expertise better and faster. But for that to work, there has to be experts."
AI is susceptible to dated information and "unknown unknowns." As an experienced engineer, I can notice when AI goes off-track, but a novice might get an illusion of understanding while missing key information. I still think AI can significantly accelerate expertise acquisition, especially for autodidacts who are using AI to learn under mentorship from human experts.
Inflated promises of superintelligent AI solving all of humanity's problems deserve skepticism, as do the claims that AI will make education and expertise obsolete. Dr. Cottom is right to challenge these narratives.
However, I think labeling the technology as "mid" is simply not true. Even existing AI capabilities are reshaping the constraint landscape in ways that will likely transform society in fundamental ways.