Rick Rosner On, 21: AI Harmonization, Regulation, Coding Agents, and the Future of Human Adaptation

Author(s): Scott Douglas Jacobsen

Publication (Outlet/Website): Personal SubStack

Publication Date (yyyy/mm/dd): 2026/08/09

In this interview, Scott Douglas Jacobsen and Rick Rosner discuss rapid AI development, institutional adaptation, and the concept of technological harmonization. They examine shifting leadership at Google DeepMind, AI competition, university policies, synthetic-media disclosure under the EU AI Act, coding agents, semiconductor diversification, energy efficiency, and human oversight. Rosner also considers how digital technologies may reshape relationships and communication, while Jacobsen argues that societies can respond through iterative policy, infrastructure, and cultural adaptation over time.

Scott Douglas Jacobsen: Demis Hassabis, the Nobel Prize-winning AI researcher, is moving away from the day-to-day management of Google DeepMind to become its chairman and Alphabet’s chief scientist. Koray Kavukcuoglu will take over day-to-day leadership of Google DeepMind while continuing as Alphabet’s chief AI architect.

At the same time, Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le are leaving Google to establish a separate AI research company called Discovery Loop. So there is a significant reshuffling of AI leadership at Google, with Hassabis moving from operational management toward longer-term scientific strategy and research.

Why do you think one of the world’s leading AI researchers would make this shift from day-to-day management toward a broader scientific and strategic role?

Rick Rosner: From what I have seen lately, and you may know better than I do, the pace is increasing. It was already fast, but new models are now coming out at a faster rate. Each new generation can feel unsettlingly more capable than the previous one.

The different models also have different strengths and areas of emphasis. I do not think the field can even be reduced to just a small handful of dominant models or companies anymore.

You have China in the mix too, but there are also smaller companies trying to find niches for themselves. It is a turbulent landscape. Everyone is scrambling for an advantage, trying to find somewhere to stand where they will not be swept away by the next wave of technology. They want to be the ones doing the sweeping rather than the ones being swept away.

There is also the question of whether we even want to be doing all of this. AI has not become less risky over time. If anything, the potential risks increase as the systems become more capable. At the same time, there is increasing economic pressure to disregard those risks because companies and countries feel they have to keep moving or risk falling behind.

People often look for historical analogies for AI, and one comparison is the automobile. Cars transformed American cities over a few decades. Before widespread automobile adoption, horses were essential to urban transportation, and large cities had to deal with enormous quantities of manure, urine, carcasses, feed, and stable waste. New York City had tens of thousands of working horses around the turn of the twentieth century, not anything close to 800,000 horses on its streets every day. As automobiles, trucks, buses, and electric transit spread, the role of urban horses declined dramatically during the first decades of the twentieth century.

If you are going to use automobiles as an analogy for AI, though, the technological transition feels much faster. It is as if decades of technological change were being compressed into a few years. Things are moving extraordinarily quickly.How nervous are you about AI?

Jacobsen: Why would I be? I think we are in a period of harmonization with the technology.

Rosner: So you think society can keep adapting to what AI can do and incorporate it?

Jacobsen: Did we ever keep up with the printing press in the sense of reading everything it produced? How many books exist? How many lifetimes would you need to read all of them?

Rosner: No individual keeps up with every technological system or every product of one. You do not keep up with every development in mining technology either. Millions of new books are published globally each year, depending on how a “book” and a publication are counted. That works out to thousands per day, not hundreds of thousands per day. The scale is still far beyond what any individual could absorb.

Jacobsen: It is a discouraging number, depending on how you look at it.

Every technology occupies particular niches, and sometimes developments within those niches become transformative. People often call the current period the Fourth Industrial Revolution. I would qualify that. The term is established, including at the World Economic Forum, but it describes a convergence of digital, biological, physical, and AI technologies rather than an industrial revolution identical in character to the earlier mechanization-driven revolutions.

What is particularly important with AI is the possibility of increasingly autonomous and self-improving systems. “Recursive self-improvement” is one of the concepts people discuss, although present-day large language models do not independently redesign themselves into progressively more capable systems in the strong sense associated with that idea.

What we are seeing instead is the integration of language models with tools, memory, planning systems, multimodal capabilities, and other components to create more general-purpose AI systems that can operate with increasing autonomy in open environments. We are not yet at artificial general intelligence in any universally accepted sense.

That is why I am not especially afraid of the technology itself. I see us as being in a harmonization period in which institutions have to adapt their policies, norms, incentives, and educational practices.

At a panel with educators in Ukraine, for example, a study was presented suggesting that only a minority of universities surveyed had formal policies governing AI use in relation to plagiarism and other forms of academic misconduct. Among students at institutions that did have policies, awareness of those policies was incomplete.

The exact percentages would need to be verified against the original study before publication, but the broader issue is clear. There can be two gaps. First, institutions may not have adequate AI policies. Second, even where policies exist, students may not know what those policies require.

Universities therefore have to develop clear rules, communicate them effectively, and design assessment systems that make responsible AI use possible while discouraging academic misconduct. Detection alone cannot solve the problem. Current AI-detection systems can produce false positives and false negatives, so institutions also need assessment design, transparency requirements, oral verification where appropriate, process documentation, and clear standards for permissible and impermissible AI assistance.

It is similar with a lot of other issues. Are universities terrified of AI? Not necessarily. They are developing policies, adapting their practices, and, in the terminology I am using, harmonizing with the technology.

We are in a period of harmonization. How much electrical power do we need? How much computing capacity does each kind of task require? Where does it make sense to devote substantial resources to artificial intelligence, and where is AI unnecessary?

AI will affect nearly every area of life, but the effects will be asymmetric across different sectors and asynchronous in their timing. Different areas will change at different rates and to different degrees. It is going to be a very bumpy ride if technological civilization continues toward something resembling a Kardashev Type I civilization.

Rosner: I was wrong about the number of books being published. I was off by at least an order of magnitude. Estimates vary enormously depending on whether you count traditional publishing, self-publishing, new editions, ebooks, print-on-demand titles, and publications in different countries. But even estimates in the thousands or tens of thousands of new titles per day are incredible.

I have read a lot of books in my life, and the larger point remains that no individual can remotely keep up with humanity’s total publishing output.

Now consider one area where, when all of this shakes out, there may simply be less fucking. That is already happening for a variety of reasons. People have more alternatives competing for their attention.

In the 1970s, if you wanted entertainment, there were far fewer options available on demand. There were no modern home video games, no streaming services, no smartphones, and no internet. Today, a large video game can occupy somebody for dozens of hours. Multiplayer games can keep people engaged for hundreds or even thousands of hours. There were recreational substitutes in the past, of course, including sports, hobbies, television, reading, music, and outdoor activities, but today’s entertainment environment is dramatically more abundant and personalized.

There is simply much more extremely compelling entertainment available now.

There are also economic and cultural factors affecting dating, relationships, and sexual activity. Forming relationships can feel more complicated. Housing and other living costs can make independent adult life harder to establish, particularly for younger people. Those factors interact with changing social norms, dating apps, delayed marriage and parenthood, and digital entertainment.

As AI becomes increasingly integrated with entertainment, the experiences available to people will probably become even more personalized, interactive, and compelling. That could provide an increasingly attractive substitute for some forms of socializing, dating, and sexual activity.

So one possible consequence of the kind of technologically advanced civilization you are describing is less banging.

That has potentially good and bad aspects. It is not only about people having less sex. It may also involve changes in how comfortable people are with face-to-face interaction.

People my age sometimes make fun of Gen Z for supposedly having weaker conversational skills or giving people the stereotypical blank stare. I would not treat that as a universal characteristic of Gen Z. There is research and ongoing debate, though, about how heavy smartphone use and digitally mediated communication affect face-to-face interaction. Digital communication can supplement relationships, but extensive screen use can also displace some opportunities to practice in-person social skills.

You can even see generational anxieties about conversation reflected in advertising. Progressive has a long-running series of commercials featuring Dr. Rick, a fictional expert who teaches younger homeowners how to avoid “becoming their parents.” A recurring joke is that the characters engage strangers in unnecessary conversations or display other stereotypically parental social habits, and Dr. Rick tries to stop them.

The joke works because social conventions have changed. Talking extensively with strangers, chatting with service workers, or striking up conversations in public can be interpreted differently across generations and situations.

That does not necessarily mean younger people get less from verbal interaction. Communication habits have shifted toward texting, messaging, social media, and other digital channels, while face-to-face interaction remains important. Recent research also suggests that online and in-person communication involve somewhat different social skills rather than simply representing a single scale from “good” to “bad” communication.

If you are older, you can find those changes irritating or alienating. Or you can see them as a natural adaptation to the technologies people now use.

I am not necessarily the best at conventional small talk, but I am super fucking chatty. This whole enterprise of ours over the past 12 years has basically been built on chatting.

I learned, not especially well, from people who are professionally chatty. Jimmy Kimmel and Adam Carolla both have substantial backgrounds in radio. In radio, especially live talk radio, you learn to avoid long stretches of dead air. Talented broadcasters are constantly listening to the conversation and preparing the next thing they can contribute as the discussion changes.

Radio itself is not dead, although traditional terrestrial radio has lost some of its cultural dominance as audiences have fragmented across streaming, podcasts, satellite radio, and other digital media. Podcasts are a somewhat different medium. There are millions of podcasts and an enormous range in quality. Not everybody making one has the conversational training or performance skills of somebody like Kimmel or Carolla.

These are all changes that follow from technological change, which is what you were getting at earlier. If you were a horse person in 1910, perhaps you were sad to see horses gradually disappear from everyday urban transportation. On the other hand, you might have recognized that replacing large numbers of working horses with mechanized transportation solved some serious urban sanitation and animal-welfare problems. Technology changed, and the environment around it changed too.

Jacobsen: Since August 2, 2026, important transparency provisions of the European Union’s AI Act have applied. Under Article 50, people generally must be informed when they are directly interacting with certain AI systems unless that fact is obvious from the circumstances. Providers of generative AI systems must also make AI-generated or manipulated outputs detectable in a machine-readable format where technically feasible. Deployers face specific disclosure obligations for deepfakes and certain AI-generated or manipulated material concerning matters of public interest.

Meta also launched Muse Code in beta on August 5, powered by Muse Spark 1.2. It is designed to write and debug code, handle long software-engineering tasks, and coordinate multiple sub-agents working in parallel. It also keeps a persistent activity log, which allows interrupted work to resume without starting the entire task again. It is entering the same broad market as tools such as OpenAI Codex and Anthropic’s Claude Code.

Can we address the first issue first, the European transparency rules for AI-generated content?

Rosner: Europe has often taken a more precautionary regulatory approach to digital technology than the United States.

Cookies are a good example. European data-protection and electronic-privacy rules helped make cookie-consent notices ubiquitous on European websites years ago. Those notices subsequently became common on U.S.-accessible websites too, especially when companies operated internationally or were subject to state privacy laws. So it is not that cookie notices only appeared in the United States within the last year or two. Europe simply moved earlier and more comprehensively in regulating certain forms of online tracking.

There is a longstanding difference in regulatory philosophy. The United States has often placed greater emphasis on commercial freedom and post-hoc enforcement, whereas European regulation has tended to impose more explicit obligations on businesses before certain harms occur.

Calvin Coolidge is often quoted as saying, “The business of America is business,” although his actual 1925 address was more nuanced than that shorthand suggests. The quotation captures something real about the strong pro-business tradition in American political culture, even if it oversimplifies Coolidge’s argument.

That difference matters with AI. Europe is saying there can be real harms when people are systematically deceived by synthetic media, so certain forms of AI interaction and generated content should be disclosed.

I am reasonably sophisticated about detecting AI-generated material, but if I am not paying attention and I am scrolling through X or some other platform, I can still occasionally be fooled. It might be a cute animal clip, a political video, or something else designed to look plausible.

Political clips are particularly important. If I see a video that makes a politician I already dislike look terrible, I have to be especially careful because confirmation bias makes it easier to accept material that reinforces what I already believe. With increasingly capable generative systems, somebody can fabricate or alter video, audio, or images to make a public figure appear to have done or said something that never happened.

So requiring disclosure for certain categories of AI-generated material is not unreasonable. The technology keeps improving.

A couple of years ago, AI-generated images often contained obvious artifacts. Those problems have diminished rapidly. Systems are increasingly capable of producing images that can be difficult to distinguish from photographs without close inspection or provenance information.

There have also been informal benchmark-style prompts that people use to track improvements in image-generation systems, including prompts involving animals and other subjects that historically exposed anatomical, spatial, or compositional errors. As the models improve, those formerly difficult examples become less useful because the systems learn to produce increasingly convincing results.

The prompt was something like an otter riding on a passenger jet, sitting in its seat and typing on a computer. A couple of years ago, you could put that into an image generator and get something vaguely recognizable, but the details would often be wrong. The animal’s paws might be malformed, the passenger compartment might not make sense, the airplane window might have the wrong shape, and the computer, tray table, and surrounding geometry could all be distorted.

Then, within months, the results became dramatically better. You could get a convincing image of an otter sitting on an airplane and typing on a computer.

More recently, the same general idea was apparently turned into an entire generated video sequence. Instead of simply showing the otter typing, the system produced multiple shots and changed camera angles. It showed what the otter was typing, shifted to another location, and introduced a team of military-style otters moving through an underground utility tunnel. They reached a locked door, while the otter on the plane appeared to be remotely working on the problem. The lock opened, and the team moved through.

The important point is not the particular otters. It is that a relatively simple prompt can now produce something resembling a short cinematic sequence, with multiple shots, recurring characters, implied causality, and a rudimentary narrative. That represents a substantial increase in generative-video capability over a short period.

So, given how quickly the technology is improving, people should have some way of knowing whether what they are looking at is synthetic.

Unless you are in America, where we have historically been somewhat more willing to let people make money by fooling other people.

Thoughts on that? You also mentioned coding agents.

Jacobsen: They are becoming an important area of competition among Meta, OpenAI, Anthropic, Google, and other companies.

AI companies are also investing more heavily in their own semiconductor capabilities while continuing to rely on outside suppliers and cloud infrastructure. Anthropic, for example, announced in August 2026 that it was building an internal chip-design team, while still planning to use hardware and infrastructure from AWS, Google, NVIDIA, and AMD. Meta is similarly expanding its MTIA custom-silicon program while maintaining partnerships with NVIDIA, AMD, AWS, Arm, Broadcom, and others.

It looks like a form of infrastructure diversification. Companies do not want their entire AI strategy to depend on one chip architecture, one supplier, or one cloud provider. That can reduce exposure to shortages, supply-chain disruptions, pricing pressure, and limitations associated with a particular kind of hardware.

Rosner: Right. There is also an opportunity to optimize entire computing systems around particular workloads instead of treating every processor as a generic interchangeable component.

Current AI data centers consume enormous amounts of electricity, generate substantial heat, and, depending on the cooling system, can consume significant amounts of water. As AI infrastructure expands, you would expect a lot of work on efficiency.

The human brain performs an enormous amount of information processing using roughly the power consumption of a dim light bulb. That does not mean silicon systems can simply imitate the brain’s energy efficiency, because biological and electronic computation work very differently. But it does show that useful computation does not inherently require today’s levels of energy consumption per operation.

AI hardware should become progressively better at producing more computation for each unit of energy, with improvements in chips, memory, interconnects, cooling, algorithms, model architectures, and workload specialization.

There are fundamental physical limits. Landauer’s principle, for instance, establishes a theoretical minimum thermodynamic cost associated with logically irreversible operations such as erasing a bit of information. It is more precise to describe this as a lower bound on information erasure than to say there is a fixed minimum amount of energy required simply to “flip a bit.”

Present computers operate enormously above that theoretical limit, so there is still considerable room for improved efficiency before fundamental thermodynamic limits become the dominant constraint.

Coding creates another set of problems. Finding errors in AI-generated code can sometimes be difficult because a developer may receive a large block of code that they did not construct incrementally themselves.

When a human programmer writes software directly, they often develop some understanding of the program’s structure while building and debugging it. With an AI coding agent, you can say, “Give me code that does this,” and receive hundreds or thousands of lines very quickly.

The output may be mostly correct, but a small number of errors can still matter. The difficulty is that reviewing a large amount of plausible-looking code can impose its own cognitive burden. Human oversight remains necessary, but the volume and apparent quality of AI-generated work can make that oversight increasingly demanding.

That connects to what you have been calling harmonization. The initial technology produces one set of problems, then institutions and engineers develop systems to mitigate them. Those solutions introduce their own second-order problems, which require another round of adaptation.

Mammography is a good example. AI systems can assist radiologists in identifying suspicious areas and, in some studies, improve diagnostic performance. One 2025 study found that radiologists using AI support achieved better overall discrimination than when reading mammograms unaided. Other large studies suggest that carefully designed human-AI screening pathways can improve sensitivity while reducing the number of human reads required.

But the relationship between human and machine is not simply, “AI reads the mammogram and a person checks it.” Human reviewers can develop automation bias, and the optimal workflow depends on how the AI is integrated into screening. If an AI system becomes highly accurate, rare errors can also become psychologically difficult to detect because people may increasingly expect the machine to be correct.

So you end up designing a second-generation system around the interaction between AI and human reviewers, including escalation rules, uncertainty thresholds, independent reads, auditing, and procedures for resolving disagreement.

Jacobsen: That is what harmonization looks like. Each technological capability creates a new layer of practical, institutional, and human problems, followed by another layer of attempted solutions.

Rosner: And your view is that none of this necessarily becomes substantially more apocalyptic than the other large-scale risks humanity already faces?

Jacobsen: Right. Good for today.

Rosner: Thank you for your time, your questions, and your tolerance. I will see you tomorrow, probably around two or three.

Jacobsen: Okay.

Rosner: All right. Thank you.

Last updated May 3, 2025. These terms govern all In-Sight Publishing content—past, present, and future—and supersede any prior notices.  In-Sight Publishing by Scott  Douglas  Jacobsen is licensed under a Creative Commons BY‑NC‑ND 4.0; © In-Sight Publishing by Scott  Douglas  Jacobsen 2012–Present. All trademarks, performances, databases & branding are owned by their rights holders; no use without permission. Unauthorized copying, modification, framing or public communication is prohibited. External links are not endorsed. Cookies & tracking require consent, and data processing complies with PIPEDA & GDPR; no data from children < 13 (COPPA). Content meets WCAG 2.1 AA under the Accessible Canada Act & is preserved in open archival formats with backups. Excerpts & links require full credit & hyperlink; limited quoting under fair-dealing & fair-use. All content is informational; no liability for errors or omissions: Feedback welcome, and verified errors corrected promptly. For permissions or DMCA notices, email: scott.jacobsen2025@gmail.com. Site use is governed by BC laws; content is “as‑is,” liability limited, users indemnify us; moral, performers’ & database sui generis rights reserved.