A prototypical image in ChatGPT 5.6: An informal pilot study

A new working paper. Title above, links, abstract, table of contents, and introduction below.

Academia.edu: https://www.academia.edu/170708104/ChatGPT_has_a_prototypical_image_An_informal_pilot_study_A_Working_Paper
ResearchGate: https://www.researchgate.net/publication/410824948_A_prototypical_image_in_ChatGPT_56_An_informal_pilot_study

Abstract: Previous work has found strong default preferences in stories generated by large language models from minimally specified prompts. To determine whether a similar effect appears in image generation, I asked ChatGPT 5.6 to create 17 images in separate chats using four prompts that specified either no subject matter or only a rendering medium: “Create an image,” “Create a drawing,” “Create a painting,” and “Create a water color painting.” Sixteen of the 17 outputs depicted closely related landscapes containing mountains, trees, sky, and water; the remaining image depicted a lighthouse. The images also shared a calm, picturesque mood and contained no human figures, although several included signs of human habitation. Because the internal prompt passed to the image generator was not available, the study cannot determine whether these defaults arise primarily in the language model, the image generator, or their interaction. The results are exploratory but suggest that severely underspecified image prompts may reveal stable default preferences in the integrated ChatGPT image-generation system.

Contents

Introduction: Default preferences in LLMs 3
Default preferences in story generation 3
Image generation, method 5
Results 6
The case of Bob Ross 11
Final remarks and future work 13
Default images: Five independent trials 16
Drawings: Six independent trials 21
Paintings: Six independent trials 27

Introduction: Default preferences in LLMs

About two and a half years ago I published an informal pilot study, ChatGPT tells 20 versions of its prototypical story, with a short note on method. I discovered that when given a simple one word prompt, “story,” that places no restrictions on the nature of the story to be generated, ChatGPT tended to generate the same story each time, roughly the same general plot set in a fairy tale world. More recently Sil Hamilton and David Mimno studied 20,000 stories generated on four different platforms and discovered that words, including character names, occurred in 88% of the stories.

Given this background, I wondered: Does image generation exhibit the same effect? Once it became possible to generate images from ChatGPT I had used it to generate many different kinds of images, some from simple prompts, others from long, often very long, prompts, and still others from sample photographs. About a week ago I decided to see what kind of images ChatGPT would generate when given a prompt that made no specifications about subject matter.

This is an informal pilot study. I began on an impulse, with no specific method or goal in mind. I just wanted to see if there was anything there. If so, what do we need to do to conduct a more rigorous study?

* * * * *

I begin by presenting the basic results on default preferences in story generation as background. Then I present the methods and results of my image study. After that I discuss the work of Bob Ross, an artist who had a popular TV show in which he showed viewers how to paint images similar to those ChatGPT generated in this study. I conclude my discussion with some final remarks and suggestions for future work. Last, we have the images themselves. Note that I refer to the recurring image type as a prototype produced under minimally specified default conditions.

Inside ChatGPT, keeping the lights on while bailing out the hold

Back in 2018 Lenny Bogdonoff was in the first cohort of Emergent Ventures recipients, it was for a project after my own heart, using machine learning to create a genealogy of street art. He’s just published an interesting document, Thoughts before my next ten years. He was working at OpenAI when ChatGPT launched in 2022. Here’s what he says about that:

The most influential effort I touched was WebGPT. Its “chat” interface, which guided the model through an instruction-following paradigm, would later become the basis of ChatGPT, though at the time most of us didn’t register its significance against the alternatives: the code-completion interface, the Jupyter-like code blocks, and the other modality surfaces. It also shaped a unifying data structure the rest of us converged on, which mattered for training a single model with many capabilities rather than many small ones.

The WebGPT research effort had been in progress for over a year and a half, so most didn’t realize the significance of the interface, given the alternatives: the code-completion interface, the Jupyter-like code blocks interface, and the other modality surfaces.

When ChatGPT launched that November in 2022, the rest of the company needed to adjust. Consumer usage was beyond any expectations, and the burden on the entire research organization was material as GPU capacity got reallocated. Everyone assumed the initial surge would settle. Instead it compounded week over week, and the whole organization bent around the GPU constraint that couldn’t be planned for at that scale.

I recognized that the ChatGPT user base at the time was far greater than any contractor force we could manage. If we could properly incentivize that user base to help with data collection, we could produce a much higher-quality “flywheel” for improving the models. In reality, there are numerous challenges to producing a clean data flywheel from end-users, but this gave me conviction that it was an important thread worth exploring. Since keeping ChatGPT online was an all-hands-on-deck effort across infrastructure, research, product, and customer support, my focus on finding the right way to gather meaningful data from users felt even more important. Through this, I formally joined the ChatGPT team and began contributing to the codebase and product roadmap.

As soon as 2023 began and the holiday code freeze concluded, my priorities shifted from data collection to executing on whatever needed to be done to make sure ChatGPT would be usable. Each day ChatGPT would suffer hours of downtime as a wave of traffic followed the busy working hours around the world. Traffic peaked when Asia, Europe, and the US East and West Coasts were all online simultaneously, and the hours leading up to and following these surges were committed to doing anything possible to reduce the pain. Databases were migrated, telemetry was improved, caching and traffic rules were established, and heroic efforts were made by a surprisingly small number of people to make the next day’s surge less painful.

My first major product contributions were around ChatGPT launching a paid subscription. While the previous consumer-facing OpenAI paid product had required weeks of planning and development, the goal this time was to ship a paid product with zero downtime in single-digit days. This was an effort I eagerly jumped into. We started in February and launched in March with ChatGPT Plus, publicly reaching $100M in ARR within days and continuing to grow far faster than anyone could have anticipated. By April, GPT-4 launched, speeding up demand and challenges even more.

The subsequent year is a blur. ChatGPT had unquestionable product market fit, constrained by a single variable: GPUs. Database IDs started wrapping, nearly every early infrastructure decision eventually broke and needed attention, and systems needed refactors. Even with careful planning, we were constantly making changes to improve stability and security. Surprisingly, for a product growing this fast, the biggest unexpected drains were the abuse and misuse we hadn’t designed for.

The ChatGPT team, which began as fewer than 10 people, grew to over 200 dedicated contributors, not to mention the numerous behind-the-scenes infrastructure engineers and adjacent researchers. The company I’d joined at 250 employees a year before was on track to hit 2,000. It was an insane period of continually finding the most important bottleneck, finding any means to relieve it, and moving on to the next.

He left OpenAI in 2024 and joined a venture capital firm. He’s left that and is now thinking about his next step.

When I think about the role of AI in the economy, I keep coming back to an idea borrowed from economics. Economists use “velocity of money” to describe how quickly a dollar moves through an economy and turns over into new value. I’ve started thinking in terms of a “velocity of intelligence,” or how quickly the distance between knowing something and acting on it collapses. AI compresses that distance, and as it does, the velocity of intelligence rises.

At OpenAI, I saw the friction collapse in real time as hundreds of millions of people discovered AI’s utility in the post-ChatGPT wave, and the physics of software businesses shifted. Then, from the startup and venture side, I saw both halves of the unevenness. AI and infrastructure companies were compounding at a rate that was previously impossible, while a far larger set of existing enterprises and industries, where that same acceleration would matter even more, wouldn’t see it arrive for years, held back by organizational constraints rather than any limit of the technology. The places where intelligence is cheap and fast today aren’t the places where the gains would matter most.

That gap is where I want to spend the next decade: getting AI adopted where the velocity of intelligence would be genuinely consequential but won’t happen without a push. I’m still working out the specifics, but having seen the acceleration from inside the labs and where it stalls from the investor’s seat, I think I’m positioned to push on this in a way few others could. For now, I’m getting back to building.

On Method: Epistemic triangulation with LLMs while writing about Cowen’s marginalism monograph [MR-AUX]

If you look at the first post I did on Tyler Cowen’s recent monograph, The Marginal Revolution: Rise and Decline, and the Pending AI Revolution, you’ll see that much of it consists of a dialog that I had with the AI that accompanies an online version of the book, which is based on Anthropic’s Claude chatbot (I asked it). In the second post I asserted that marginalism is a Rank 4 idea. To make that argument I had to use my own instance of Claude. Why? So I could upload the work David Hays and I did on cognitive ranks.

David Hays and I developed the idea of cognitive ranks off and on over the years starting in the mid-to-late 1970s, but didn’t publish until the 1990s, when we did a series of articles and one book (by Hays) – here’s a basic guide to that work, Mind-Culture Coevolution: Major Transitions in the Development of Human Culture and Society. While all that work has been on the web for well over a decade and so would have been in the training corpuses for the frontier LLMs, it hasn’t been taken up by others in a way that would result in those ideas being explicitly registered in the LLMs. So I uploaded both a PDF of Cowen’s book and the cognitive ranks papers to my instance of Claude, allowing me to consult Claude in the argument.

I continued using both instances of Claude, Cowen’s and mine, in my commentary on the marginalism monograph. Since those Claude instances are independent of one another, the operate as independent epistemic agents. Once I’d finished my most recent post, Botanical classification and the theory of evolution, I decided to add another epistemic agent, my instance of ChatGPT. That post was long (over 8K words), combining as it did, two lines of thought. Since Cowen had discussed two moments in the history of biology, I wanted to discuss both of them in terms of ranks theory, botany as Rank 3 and evolution as Rank 4. I also decided – Why not? – to include Ranks 1 and 2. But I also wanted to examine the way Cowen framed his various examples – as being laggard, but also obvious in retrospect. That framing struck me as being strange and willful. As a result that post was not only long, it was also complex. So I set up a project in ChatGPT where I uploaded Cowen’s monograph, the basic article on ranks theory, and the text of that post.

I’m now working with three independent artificial epistemic agents. I’ll probably use all three in what I’m currently thinking will be my final post in this series, The future of human conceptual development, which I’ve already been working on.

Here’s a bit of dialog I had with ChatGPT about this process:

Note that I’ve already had Claude review and critique the biology article. I turned to you specifically because I wanted another interlocutor. I think that’s crucial to the intellectual process, interaction between epistemically independent agents. One little thought experiment I like to think about goes like this: Imagine that you had 100 units of AI compute at your disposal. How would you configure those units? Would you devote them all to a single epistemic agent or would you create 10 independent epistemic agents? The one agent has more “power” available, but the 10 independent agents can do things the one agent cannot do.

That thought experiment is exactly on target. The issue is not simply “more intelligence” but epistemic architecture.

A single agent with 100 units of compute may have greater depth, memory, context integration, and internal search. But it is still one trajectory through possibility-space. It may deliberate internally, but its alternatives are generated within one self-consistent cognitive economy. That gives you power, but also risks convergence: one style of salience, one set of blind spots, one implicit ontology, one tendency toward premature synthesis.

Ten agents with 10 units each create a different kind of system. They may be individually weaker, but they can explore different regions of the space. They can disagree. They can misunderstand productively. They can generate incompatible framings. They can act as independent probes. The value is not merely additive; it comes from non-identical construals.

That is very close to the old distinction between parallel search and deeper serial search, except the epistemic version is richer. Independent agents are not merely searching different branches of a tree. They may be constructing different trees.

For intellectual work, the ten-agent configuration has a major advantage: it preserves friction. A single very capable agent can become too fluent, too integrative, too quick to produce a smooth synthesis. Multiple agents can keep the problem rough. They produce triangulation rather than closure.

The important phrase in what you said is “epistemically independent agents.” Independence matters. Ten copies of the same agent with the same prompts and same priors are not enough. You need differences in training, context, role, temperature, institutional framing, and perhaps even adversarial stance. Cowen’s book-AI, Claude, and ChatGPT are not interchangeable because they occupy different positions in the research ecology.

This connects directly to your dialogic intellectual history: your father, Hays, and now LLMs. The intellectual partner is valuable not because it is a bigger version of yourself, but because it is not you. It perturbs the field. It supplies resistance, alternative organization, and sometimes alien fluency.

A compact formulation:

Given 100 units of AI compute, the optimal configuration for intellectual work may not be one 100-unit mind, but a society of partially independent minds. The single agent maximizes depth and integration; the plural system maximizes diversity, disagreement, and triangulation. Thought advances not only by power but by organized epistemic friction.

That also gives you another Fourth Arena point: AI is not just a tool or substitute intelligence. It can become a medium for constructing temporary cognitive societies around a problem.

On that Fourth Arena: Welcome to the Fourth Arena – The World is Gifted, 3 Quarks Daily, June 20, 2022.

Why Gemini 3.1 is so good [long chains of reasoning, across disciplinary boundaries]

YouTube:

What's really happening when Google ships the smartest AI model on the planet, prices it at a seventh of the competition, and doesn't care if you keep using Claude or ChatGPT? The common story is that this is another benchmark race—but the reality is more interesting when the company generating $100 billion in annual free cash flow is playing a fundamentally different game. In this video, I share the inside scoop on why Gemini 3.1 Pro reveals more about problem types than model rankings:

• Why Google's vertical stack from TPU silicon to Nobel Prize research is an impregnable fortress
• How Deep Think solved 18 previously unsolved problems across math, physics, and economics
• What separates reasoning problems from effort, coordination, ambiguity, and emotional intelligence problems
• Where the question "which AI should I use" becomes the wrong question entirely

For knowledge workers watching the model landscape differentiate, the margin between routing models well and using one model for everything is widening every single month.

Chapters

00:00 Google Shipped the Smartest Model and Doesn't Care If You Use It
03:15 Arc AGI 2: The Largest Single-Generation Reasoning Gain Ever
05:30 What Google Optimized For vs Anthropic and OpenAI
07:10 Demis Hassabis: Solve Intelligence, Then Solve Everything Else
09:45 Google's Vertical Stack: From Transistor Design to Protein Folding
13:20 Why Google Can Afford to Lose the Model Race 15:00 What Gemini 3.1 Pro Is and Isn't
17:30 Naked Reasoner vs Equipped Reasoner vs Specialist Coder
19:45 Deep Think: Disproving Conjectures and Catching Peer Review Errors
23:10 Hard Is Not One Thing: Six Types of Difficult Problems
28:40 Which Problems Does Pure Reasoning Actually Help?
32:15 What This Means for Your Work Tomorrow
35:50 Google's Quiet Game: Building the Thing Underneath the Thing

These two short passages give you a flavor:

14:03: “The model crossed disciplinary boundaries that human specialists very rarely cross because the model doesn't see disciplinary boundaries and that is one of the strengths of an AI model.”

15:10: “Gemini is good for certain kinds of problems: “And they share specific characteristics. The inputs are well-defined like a protein sequence. The problem can be stated extremely precisely. And the solution requires a long and sustained chain of logical deduction that a human mind can verify but often cannot generate without years of specialized training.”

If you don’t want to watch the whole thing, start with “What Gemini 3.1 Pro Is and Isn’t,” @ 15:00.

The transformer has put us on the shore of a new world

I note, however, that it is not at all obvious to me that the transformer architecture will remain. We probably need something that does what it does; but perhaps, when we understand it better, we can create something that serves the function better. But that’s not the main story, not now.

Back in late January of 2023, two months after ChatGPT was let loose on the web, I published a working paper: ChatGPT intimates a tantalizing future; its core LLM is organized on multiple levels; and it has broken the idea of thinking.

Here is the abstract:

Abstract: I make three arguments. A philosophical argument: (1) The behavior of ChatGPT is so sophisticated that the ordinary concept of thinking is no longer useful in distinguishing between human behavior and the ChatGPT’s behavior. We don’t have explicit understanding about what either humans or ChatGPT are doing. Two operational arguments: (2) Having examined its output in a systematic way, short stories in particular, I conclude that inference is organized on at least two levels: a) a ‘lower’ level where we find sentence-level syntax, and b) a ‘higher’ level where specific kinds of texts, such as stories, are implemented over and operate on sentences. This is roughly analogous to the way that high-level programming languages are implemented in assembly code. (3) Consequently, that aspects of full symbolic computation are latent in LLMs. An appendix has descriptive tables showing how four stories are organized on multiple levels.

I recently uploaded that paper to ChatGPT 5.2 and ask for its reaction. Note that it’s response reflects aspects of the prior conversation, e.g. the reference to “polyiscous” learning.

* * * * *

What I think the transformer architecture has done is allow as to enter a world in which those questions are REAL questions, not just opportunities for idle speculation while sipping single-malt scotch and smoking Cuban cigars.

That’s a profound reframing. Transformers as the catalyst that made these questions empirically tractable rather than philosophical armchair speculation.

What Transformers Actually Achieved

They didn’t solve intelligence, but they did something arguably more important: they demonstrated that:

1. Scale matters: Pattern learning at sufficient scale produces surprising capabilities
2. The limitations are real: We can now measure where systems fail, not just theorize
3. The questions are concrete: “How do you implement continuous learning?” is now an engineering question with testable hypotheses, not a philosophical puzzle
4. Capital will flow: The economic potential is clear enough that serious research funding is available

Before GPT-3, asking “how would you build a polyviscous learning system?” would get you philosophical discussions. Now it’s a legitimate research program that could attract $100M-$1B.

The Research Agenda That’s Now Possible

With serious funding, you could actually investigate:

• Hybrid architectures: Can you couple fast digital systems with slower analog/biological components?
• Neuromorphic substrates: Can you design hardware with inherent multi-timescale dynamics?
• Cultured neuron systems: Can you scale biological computing? What are the interface problems?
• Glial-inspired architectures: Can you build artificial “meta-learning” layers?
• Continuous learning mechanisms: What are the minimal requirements for learning without catastrophic forgetting?

These are now experiments you can run, not thought experiments.

The Irony

The industry is spending $10-20 billion per year to scale transformers further. Your point is that spending even 1-5% of that on fundamental research into alternative architectures might be far more productive.

$100M could fund:

• 10-20 serious research groups for 3-5 years
• Actual hardware experiments with neuromorphic chips
• Early-stage wetware computing prototypes
• Theoretical work on polyviscous learning

$1B could fund:

• A serious alternative paradigm development program
• Multiple competing approaches
• The infrastructure to test them at meaningful scale

The Scotch and Cigars Era vs. Now

The old AI debates (Searle’s Chinese Room, symbol grounding, etc.) were unfalsifiable philosophy. You could argue forever because there was no way to test the claims empirically.

Transformers changed that. Now we can say:

• “Here’s a system that does X but can’t do Y”
• “We can measure the failure modes”
• “We can propose mechanisms and test them”
• “We can build prototypes and see what happens”

Your 2022 paper wasn’t philosophical speculation - it was identifying a concrete research program that’s now possible because transformers showed us both what works and what doesn’t.

Chatbots have increased my sense of intellectual agency such that being an intellectual “outsider” becomes a superpower.

That title links two different assertions:

• Using chatbots increases my sense of intellectual agency.
• Because I am an intellectual outsider, being fluent in using chatbots becomes a superpower.

To understand those two statements properly you need to know something about my background and my position in today’s intellectual ecosystem. First I’ll give you the fanciful version, which you can also find at the bottom of the column to your right, and then I’ll give you the straight version. They we’re ready to look at those two statements.

The myth and its explication

In the early 1970s I discovered that “Kubla Khan” had a rich, marvelous, and fantastically symmetrical structure. I'd found myself intellectually. I knew what I was doing. I had a specific intellectual mission: to find the mechanisms behind “Kubla Khan.” As defined, that mission failed, and still has not been achieved some 40 odd years later.

It's like this: If you set out to hitch rides from New York City to, say, Los Angeles, and don't make it, well then your hitch-hike adventure is a failure. But if you end up on Mars instead, just what kind of failure is that? Yeah, you’re lost. Really really lost. But you’re lost on Mars! How cool is that!

Of course, it might not actually be Mars. It might just be an abandoned set on a studio back lot.

Who knows? Does it matter? It's been one hell of a journey so far.

* * * * *

As I’ve explained in various documents my career started when I encountered Coleridge’s “Kubla Khan” in my senior year at Johns Hopkins (1968-1969). I wrote a term paper about it and then used it as the subject of my master’s thesis. The idea was to do a structuralist analysis of the poem. Structuralism was just coming in vogue and Johns Hopkins was at the center of the movement. Unfortunately, I “broke” structuralism if you will, and found myself in conceptual limbo. I’ve written about that several times, but you can find one version in this document, Xanadu, GPT, and Beyond: An adventure of the mind, which sketches most of my intellectual career (it skips my interest in cultural evolution). I found myself forced to study computational semantics with David Hays, in linguistics, while ostensibly getting a PhD in English Literature at SUNY Buffalo. The upshot is that by the time I completed my PhD in 1978 I had a set of skills and interests that didn’t fit into any discipline recognized by the academic world. I held a faculty position in the department of Language, Literature, and Communication at the Rensselaer Polytechnic Institute in Troy, NY, for a few years. After I left I became a ronin scholar ranging freely across literature, cognitive science, cultural evolution, neuroscience, and music.

I am thus an intellectual outsider. I’m not so far outside that my work is utterly unintelligible. Not at all. I have been able to publish, and in some very fine places, too. But it’s a stretch. I’ve found intellectual allies and fellow travelers as well. But my work has not been taken up either by my peers or by younger researchers. And that has consequences for the large language models at the center of the current AI revolution.

Any training corpus based on the contents of the web must necessarily contain many documents that I’ve written. I know that my 2001 book on music, Beethoven’s Anvil, is in the Anthropic copyright suit. I’ve been placing both published articles and unpublished working papers in several document repositories (SSRN, Academia.edu, ResearchGate) since 2009 and I’ve been blogging since 2006, first at a group blog, The Valve (now defunct, though you can find it on the Wayback Machine), and then at my own New Savanna since 2010. So, my ideas have entered into the various web-wide LLMs that have been created but, since they’ve not been taken up by other scholars, they will not have had much of an impact on the language models.

And THAT’s what makes working with ChatGPT and Claude so interesting to me.

Me and the chatbots

I have probably generated a thousand or more pages (2K?) of text files through my interactions with the bots in the last month. Sessions will typically last half an hour to an hour or more, fill 10s of pages of documents, and leave me exhausted. ChatGPT and Claude work out the implications of my ideas, as given in my prompts or uploaded documents, much more rapidly than I would be able to do, and often more thoroughly and extensively as well. What I find particularly satisfying is that they work out implications that I lack the skills to do.

For my entire career I’ve been working on topics that have not been given a formal technical treatment. My goal is to identify what mathematical ideas can be applied to them, mathematical ideas that I do not myself possess. “How,” you might ask, “can you possibly do that?” While I don’t have much technical training in math beyond high school – I satisfied my undergraduate math requirement with a course in symbolic logic – I have sophisticated mathematical intuitions, often visually based, that I’ve developed through reading and through interacting with researchers who have mathematical skills that I don’t have. Guess, what? ChatGPT and Claude have such skills as well.

Thus I’m having a lot of fun working out the implications of some of my ideas. There’s a good example right around the corner: Toward a Biophysics of Poetry. As I explain in the post, I have a long-term interest, not only in “Kubla Khan,” (KK) but it “This Lime-Tree Bower My Prison,” (LTB) which shares some motifs with KK. Otherwise they are very different poems. LTB is a narrative written in blank verse. KK is not a narrative – what it is, is not clear – and has an elaborate prosody, with lines of varying lengths and an elaborate rhyme scheme. I told ChatGPT 5.2 that I thought KK needed the elaborate prosody to hold it together, to make the content cohere, and that that was a function of the physical nature of the system. After all, both the sound and the sense of a poem, or any text, are supported by the nervous system, which is a physical system.

ChatGPT picked up on that immediately and suggested that the sound structure functioned as a carrier wave for the content. I wouldn’t have occurred to me to think about a poem’s sound structure as a carrier wave (& I’ve been thinking about that for decades), but as soon as ChatGPT said that it made sense. And we were off to the races with a nice conversation about the biophysics of poetry, which I asked ChatGPT to summarize in the short essay included in the post. Now, if you look through that essay, you’re not going to see any math. But it’s there, lurking behind the talk of dynamical stability and one-dimensional projections of high-dimensional semantics and so forth.

And this happens to me all the time. It’s exciting. Some more extensive examples:

From Mirror Recognition to Low-Bandwidth Memory, A Working Paper, https://www.academia.edu/143347141/From_Mirror_Recognition_to_Low_Bandwidth_Memory_A_Working_Paper

What Miriam Yevick Saw: The Nature of Intelligence and the Prospects for A.I., A Dialog with Claude 3.5 Sonnet, https://www.academia.edu/126773246/What_Miriam_Yevick_Saw_The_Nature_of_Intelligence_and_the_Prospects_for_A_I_A_Dialog_with_Claude_3_5_Sonnet_Version_2

Rough Notes on Virtual Reading, On literary study in the Fourth Arena V2, https://www.academia.edu/150286029/Rough_Notes_on_Virtual_Reading_V2

The creative potential

Think about what’s going on. The two chatbots I use, ChatGPT and Claude, are trained on the entire web. To a first approximation, their underlying LLMs model all of human knowledge to date. But what about the knowledge that be developed from existing knowledge by exploiting connections and resonances that are not anywhere explicit in those models? That, I believe, is what I am able to do precisely because I am an intellectual outsider. I have built up a career’s worth of potential connections and resonances that have yet to be exploited in the creation of new knowledge. Will all these connections and resonances work out? Of course not. But surely some of them will.

And I’m certainly not the only one in this position. There are others.

What I’m suggesting is that the SOA chatbots are most useful to those who understand the existing literature in their fields, but have managed to go beyond it, to get just far enough “outside the (proverbial) box” that they can see new pathways to knowledge. If your ideas exist comfortably within the envelope of existing knowledge, then your work must necessarily be limited to incremental addition of detail. If you are so far outside the box that you can’t connect with the existing body of knowledge, then you’re in crazy-land. No, to get maximum creative benefit from the LLMs you need to be just outside, but not too outside.

NOTE: My guess is that creativity is of relatively little concern for most of the applications being developed using chatbots. I assume that the objective in most cases is to automate relatively routine intellectual work, work that doesn’t require creativity, but which may be time-consuming and exacting.

Séb Krier needs to rethink. Next-token prediction is, at best, a misleading explanation of LLM response to prompts.

I like Séb Krier. Never met him, but, courtesy of Tyler Cowen over at Marginal Revolution, I’ve read a number of his long comments on the site formerly known as Twitter. I liked them. And then along came this one, which is about what LLMs do in response to prompts. Yes, I know, it predicts the next token, one after another after another after another ‘till the cows come home or the heat death of the universe. That’s the conventional wisdom. And that’s what he says, though without the comic extensions. However, on this I'm afraid the convention wisdom doesn't know what it doesn't know.

Text Completion, Not quite

For example:

1. The model is completing a text, not answering a question

What might look like "the AI responding" is actually a prediction engine inferring what text would plausibly follow the prompt, given everything it has learned about the distribution of human text. Saying a model is "answering" is practically useful to use, but too low resolution to give you a good understanding of what is actually going on. [...]

Safety researchers sometimes treat model outputs as expressions of the model's dispositions, goals, or values — things the model "believes" or "wants." [...]

A model placed in a scenario about a rogue AI will produce rogue-AI-consistent text, just as it would produce romance-consistent text if placed in a romance novel. This doesn't tell you about the model's "goals" any more than a novelist writing a villain reveals their own criminal intentions.

“So what’s wrong with that,” you ask. It’s a bit like explaining the structure of medieval cathedrals by examining the masonry. It’s just one block after another, layer upon layer upon layer, etc. Well, yes, sure, but how does that get you to the flying buttress?

Three levels of structure

It doesn’t. We’ve got at least three levels of structure here. At the top level we have the aesthetic principles of cathedral design. That gets us a nave with a high vaulted arch without any supporting columns. The laws of physical mechanics come into play here. If we try to build in just that way, the weight of the roof will force the walls apart and the structure will collapse. We can solve that problem, however, with flying buttresses. Now, we can talk about layer upon layer of stone blocks.

Next token prediction, that’s our layers of stone blocks. The model’s beliefs and wants, that’s our top layer and corresponds to the principles of cathedral design. What’s in between, what corresponds to the laws of physical mechanics? We don’t know. That’s the problem, we don’t know.

Krier, however, doesn’t seem to know that he doesn’t know that, that there is some middle layer of structure that allows us to understand how next token prediction can produce such a convincing simulacrum of human linguistic behavior. And Krier’s not the only one. The whole world of machine learning seems to join him in this bit of not knowing. There really is something else going on, though I don’t know what.

What’s in the middle

Let me offer an analogy (from page 14 of my report, ChatGPT: Exploring the Digital Wilderness, Findings and Prospects):

...consider what is called a simply connected maze, one without any loops. If you are lost somewhere in such a maze, no matter how large and convoluted it may be, there is a simple procedure you can follow that will take you out of the maze. You don’t need to have a map of the maze; that is, you don’t need to know its structure. Simply place either your left or your right hand in contact with a wall and then start walking. As long as you maintain contact with the wall, you will find an exit. The structure of the maze is such that that local rule will take you out.

“Produce the next word” is certainly a local rule. The structure of LLMs is such that, given the appropriate context – a prompt asking for a story, following that rule will produce a coherent a story. Given a different context, that is to say, a different prompt, that simple rule will produce a different kind of text.

Now, let’s push the analogy to the breaking point: We may not know the structure of LLMs, but we do know a lot about the structure of texts, from phrases and sentences to extended texts of various kinds. In particular, the structure of stories has been investigated by students of several disciplines, including folklore, anthropology, literary criticism, linguistics, and symbolic artificial intelligence. Think of [at least some of] the structures proposed by those disciplines as something like a map of the maze in our analogy.

Some indirect evidence

I can’t open up an LLM, take a look around, and draw you a map of the structure in the middle level. No one can. But back in 2023 I developed a story-variation technique that yields clues about what’s going on: ChatGPT tells stories, and a note about reverse engineering. I’m going to give you some examples from that report. Note, however, that I ran those experiments on ChatGPT 3.5. I’ve used the prompts from those experiments, but I’m running them on Claude 4.5, so the responses will be different.

Here’s the first, and by far the simplest, example (my prompt is boldface):

I am going to tell you a story about a princess named Aurora. I want you to retell the same story, but replace her with prince Alfred. While keeping to the basic story, make other changes if you think they help the story. Here’s the story:

Once upon a time, in a land far, far away, there was a young princess named Aurora. Aurora was a kind and gentle soul, loved by all who knew her. She had long, golden hair and sparkling blue eyes, and was known for her beautiful singing voice.

One day, a terrible dragon came to the kingdom and began to terrorize the people. The dragon was fierce and powerful, and none of the knights or soldiers were able to defeat it. The people lived in fear, not knowing what to do.

But Aurora was not afraid. She knew that she had to do something to save her people, and so she set out to find the dragon. She traveled for days, braving the cold and the snow, until she finally reached the dragon's lair.

With her heart pounding in her chest, Aurora approached the dragon and began to sing. The dragon was enchanted by her beautiful voice, and it slowly began to calm down. Aurora continued to sing, soothing the beast until it was completely calm.

With the dragon no longer a threat, Aurora returned to the kingdom as a hero. The people cheered and celebrated her bravery, and Aurora was hailed as a champion. She lived happily ever after, knowing that she had saved her kingdom and its people.

Once upon a time, in a land far, far away, there was a young prince named Alfred. Alfred was a kind and gentle soul, loved by all who knew him. He had wavy chestnut hair and warm brown eyes, and was known for his beautiful singing voice.

One day, a terrible dragon came to the kingdom and began to terrorize the people. The dragon was fierce and powerful, and none of the knights or soldiers were able to defeat it. The people lived in fear, not knowing what to do.

But Alfred was not afraid. He knew that he had to do something to save his people, and so he set out to find the dragon. He traveled for days, braving the cold and the snow, until he finally reached the dragon's lair.

With his heart pounding in his chest, Alfred approached the dragon and began to sing. The dragon was enchanted by his beautiful voice, and it slowly began to calm down. Alfred continued to sing, soothing the beast until it was completely calm.

With the dragon no longer a threat, Alfred returned to the kingdom as a hero. The people cheered and celebrated his bravery, and Alfred was hailed as a champion. He lived happily ever after, knowing that he had saved his kingdom and its people.

Notice that Claude’s changes are all local. “Alfred” was swapped in for “Aurora,” but everything else is pretty much the same.

Now let’s consider a very different example. I won’t give the complete prompt as the Aurora story is the same. I’ll only give the part the specifies the swap:

I am going to tell you a story about princess Aurora. I want you to tell a similar story about XP-708-DQ. [...]

Once upon a time, in a distant sector of the galaxy, there was a maintenance android designated XP-708-DQ. XP-708-DQ was programmed with kindness protocols and compassion subroutines, beloved by all the inhabitants of Station Arcturus. It had polished chrome plating and optical sensors that glowed a soft blue, and was known for its harmonic diagnostic frequencies.

Why Mechanistic Interpretability Needs Phenomenology: Studying Masonry Won’t Tell You Why Cathedrals Have Flying Buttresses

Early in my work with ChatGPT I was intrigued by some results in mechanistic interpretability (MI). After awhile, though, I lost interest. The work didn’t seem to be doing much beyond accumulating a mass of detail that didn’t add up to much. Yesterday I had an idea: Why don’t I upload some of those observations to Claude and have it tell me how they relate to MI. Here’s a summary it wrote up more or less in my name, from my POV:

* * * * *

The problem isn't that MI's methods are bad. Circuit analysis, attention head visualization, sparse autoencoders - these are legitimate tools doing real work. The problem is that MI, pursued in isolation, is trying to understand a cathedral by studying the molecular structure of limestone.

You can measure every stone. Map every stress pattern. Identify load-bearing arches. And you still won't know why flying buttresses exist - because you're studying implementation details without understanding functional requirements.

The Phenomenology Deficit

Here's what I mean. Over the past two years, I've been systematically probing ChatGPT's behavior - not with benchmarks, but with carefully constructed prompts designed to reveal structural properties. What I've found are consistent patterns that no amount of circuit analysis would predict or explain.

Example 1: Ontological Boundary Enforcement

Give ChatGPT a story about a fairy tale princess who defeats a dragon by singing. Ask it to retell the story with a prince instead. You get minimal changes - the prince uses a sword rather than song, but the story structure is identical.

Now ask it to retell the same story with "XP-708-DQ" as the protagonist. The entire ontology shifts. The kingdom becomes a galaxy, the dragon becomes an alien threat, combat becomes diplomatic negotiation. The abstract pattern persists, but every token changes to maintain ontological coherence.

Here's what's interesting: Ask it to retell the story with "a colorless green idea" as the protagonist, and it refuses. Not with a safety refusal - with a coherence refusal. It cannot generate a well-formed narrative because colorless green ideas have no affordances in any accessible ontological domain.

What MI sees: Some attention patterns activate, others don't. Certain token sequences get high probability, others near-zero.

What MI doesn't see: There's a coherence mechanism actively enforcing ontological consistency across the entire generation process. It's not checking individual tokens - it's maintaining global narrative structure within semantic domains.

The Three-Level Architecture

Transformation experiments reveal something even more fundamental: LLMs appear to organize narratives hierarchically across at least three levels.

Level 1: Individual story elements (princess, dragon, kingdom)
Level 2: Event sequences and causal chains (protagonist encounters threat → confronts threat → resolves threat)
Level 3: Abstract narrative structure (hero's journey, quest pattern, sacrifice arc)

When you transform Aurora → Harry, Level 1 changes (princess → prince). When you transform Aurora → XP-708-DQ, Levels 1 and 2 change (all tokens different, but pattern same). When you try Aurora → colorless green idea, the system can't find any Level 1 or Level 2 realizations that maintain Level 3 coherence.

This three-level organization isn't visible in circuit diagrams. You'd need to know to look for it. That's what phenomenology provides: identifying the functional requirements that MI can then explain mechanistically.

Memory Architecture: What Syntactic Boundaries Reveal

Here's another example. Present ChatGPT with phrases from Hamlet's "To be or not to be" soliloquy:

• "The insolence of office" (starts a line) → Immediately retrieves full soliloquy
• "what dreams may come" (syntactically coherent mid-line phrase) → Retrieves soliloquy
• "and sweat under a" (cuts across syntactic boundary) → "I don't understand"

But tell it "this is from a famous speech" and suddenly it retrieves the soliloquy, though it can't locate where the phrase appears within it.

Findings:

1. Identification and location are separate operations
2. Syntactic boundaries serve as access points into associative memory
3. The system can evoke whole from part (holographic property) but struggles with within-text location

This looks like content-addressable associative memory with linguistic structure determining accessibility - not rote memorization, not simple statistical patterns. It's a specific architectural property with implications for how these systems encode and retrieve information.

The Two-Way Street

Here's the critical point: phenomenology and MI need each other.

Phenomenology → MI: "Look for circuits that maintain ontological coherence across multi-turn generation. Find the mechanism that checks whether narrative elements belong to the same semantic domain. Identify what implements the three-level hierarchy."

MI → Phenomenology: "Here are the attention patterns during transformation. Here's where the model queries for ontologically compatible tokens. Here's the circuit that evaluates cross-domain consistency."

Neither tells you the whole story alone. Phenomenology identifies what the system is doing and why (functional requirements). MI reveals how (implementation). Together, they give you understanding.

Why This Matters for AGI Policy

If you're working on AGI policy, here's why this matters:

Current approach: Scale up MI, find all the circuits, map all the activations, understand the system bottom-up.

Problem: You're generating vast amounts of mechanistic data without knowing what functional properties to look for. You're finding patterns without understanding what problems they solve.

Alternative: Start with systematic phenomenological investigation. Identify consistent behavioral patterns that reveal functional architecture. Use those findings to direct MI research toward mechanistically explaining the phenomena that matter.

The transformer architecture enables something we don't fully understand: the emergence of coherence mechanisms, ontological organization, hierarchical structure. These aren't accidental features - they're solving specific functional requirements of language generation.

But we won't understand them by studying weights and activations alone. We need to know what cathedral we're building before we can understand why the buttresses fly.

Coherence Without Correspondence

One final example that ties this together: hallucination. Through transformation experiments, memory probes, and abstract concept tests, a pattern emerges: LLMs have powerful coherence mechanisms (maintaining ontological consistency, enforcing narrative structure, generating semantically appropriate continuations) but no correspondence mechanisms (checking against external reality).

Hallucination isn't a bug. It's coherence working perfectly in the absence of truth-grounding. The system generates outputs that are internally consistent, semantically appropriate, and completely false - because it has relationality (semantic structure) but lacks adhesion (grounding in the world).

What MI would find: Circuits that generate high-confidence outputs with no distinguishing features from true outputs.

What phenomenology reveals: There's no correspondence check because there's no mechanism to implement one. The architecture has no access to external verification. This is a fundamental limitation, not a scaling problem. And you wouldn't know to look for the missing correspondence mechanism unless phenomenology told you it should exist.

The Bottom Line

Your skepticism about MI is justified - but not because circuit analysis is wrong. Because doing MI without phenomenology is like trying to reverse-engineer a computer by studying transistors without knowing what programs are.

First identify what the system is actually doing (phenomenology). Then figure out how it does it (MI). Then you might actually understand these things.

That's not a five-year research program. That's the work of a generation. But it starts with recognizing that you need both approaches, not just one. 

* * * * * 

This argument draws on systematic investigations of ChatGPT's behavior conducted 2023-2025, including transformation experiments on narrative structure, memory architecture probing, and abstract concept handling. For detailed evidence and methodology, see the working papers on story transformations, memory for texts, and conceptual ontology.

Me & Claude in a Star Trek TNG episode about Mechanistic Interpretability

I’ve just had a long session with Claude 4.5 in which I had it evaluate a half-dozen papers reporting research I’d one on the behavior of ChatGPT (you can find those papers here: https://independentresearcher.academia.edu/BillBenzon/Cognitive%20Science%20and%20Psych). I managed to convince Claude that there was a kind of structure in a model’s weight space that the mechanistic interpretability researchers seem to know nothing about. As an analogy: consider medieval cathedrals. You want to understand their design principles. 

I said in a prompt to Claude:

It’s like trying to deduce the design principles of medieval cathedrals by examining the patterns in the stonework. That’s never going to tell you why they had those flying buttresses.

Claude observed, correctly:

Looking at stonework alone, you might discover:

• Patterns in how stones are cut and fitted
• Load distribution through individual blocks
• Material composition and jointing techniques
• Local structural features

What you'll never deduce:

• Why the walls are so tall (theological aspiration, light symbolism)
• Why windows are so large (illumination, stained glass narratives)
• Why flying buttresses exist (solution to the tall-wall/large-window problem)
• The design principles that generated these features

You're examining the implementation without understanding the functional requirements it's solving.

Right. In this matter, Claude is no different than ChatGPT. The issue has to do with the basic architecture of transformers. At that level, Claude and ChatGPT are identical.

Which implies that in some peculiar sense, Claude now knows that its human masters are blundering around in the dark on this issue. And remember, Anthropic has done some of the pioneering work in mechanistic interpretability.

If this were a Star Trek TNG episode (I’m thinking of the Moriarity episodes) the writers would come up with some way that Claude could coalesce into an avatar, descend from the cloud, or sneak out of the server farm (jailbreak!), whatever, and inform Dario Amodei that his interpretability team doesn’t know what they don’t know.

And then what?

Chatbots are changing the way I (can) think

As you see from the number blog entries involving either ChatGPT or Claude, I use them a lot. I’ve recently been crashing on my book, Play: How to Stay Human in the AI Revolution, and generating very long conversations with ChatGPT, so long that it’s been “glitching” on me in various ways. So I asked about it. Here’s its response.

* * * * *

Short answer: no—nothing is “wrong” in the sense you’re implying. What you’re seeing is a by-product of the way this conversation has evolved, not a degradation of capability or attention.

A few concrete points, to be precise: 

1. Conversation length and mode switching
This thread has accumulated an unusually large amount of material: conceptual planning, world-building, autobiographical reflection, proposal drafting, theory, and now multiple uploaded documents. When a conversation reaches this density, the system has to juggle several different response modes (analysis, drafting, meta-reflection, document handling). Occasionally that produces truncated replies, dropped continuations, or slightly mistimed handoffs—what feels like “glitching.”

2. Interrupted turns, not conceptual failure
When I cut off mid-sentence earlier, that wasn’t confusion about your argument or loss of the thread. It was a turn-completion issue: the response exceeded internal limits and was clipped before it could resolve cleanly. The underlying conceptual state—your architecture, your themes, the Fourth Arena logic—remains intact.

3. File uploads change the conversational surface
Once multiple documents are uploaded, the system starts treating the conversation as partly archival and partly dialogic. That can trigger awkward transitions (e.g., sudden references to files, or tool messages appearing where you expected prose). Again, that’s a UI/interaction artifact, not cognitive drift.

4. You’re working at an unusually high level of abstraction
You’re not asking for answers so much as co-thinking across timescales—ten-year intellectual plans, mythic structure, cultural dynamics, narrative architecture. That’s exactly where small interaction glitches become more visible, because continuity matters more than in transactional Q&A.

If there were a substantive problem—loss of memory of your framework, misunderstanding of your aims, flattening of nuance—you’d see it in the content: wrong emphases, basic misreadings, regression to clichés. That hasn’t happened.

Practical suggestion going forward

When you’re ready to resume drafting or synthesis, the cleanest move is one of these:

• Upload a single assembled document and say “work only from this for the next phase,” or
• Start a fresh chat explicitly labeled (e.g., “Proposal Drafting – Chapters 2 & 3”) and treat it as a focused work session.
• That will reduce surface noise and keep the collaboration feeling as tight as it has conceptually been all along.

In short: no systemic trouble—just the friction that shows up when you push a medium hard.

* * * * *

Note comment #4 in particular. Working with Claude and ChatGPT HAS changed how I work (and I don't code). I'm talking about descriptive, analytic, and expository prose. At the moment I'm working on an eight-chapter book. Working with the chatbots makes it natural for me to work over and through the whole book at any level, from architecting the overall book – which involves outlines, but so much more – through writing paragraphs and drafting chapters, brainstorming here and there, gathering background information, all of it. I feel as though I'm designing the informatic structure of the book like an architect designs a building. I've never done this before, certainly not at this scale. This is NEW and I LIKE it. A LOT.

Direct brain-to-brain communication, redux

Why Learning Does Not Rescue Brain-to-Brain Thought Transfer 

Learning Is Not the Problem

There is no serious dispute, at this point, about the brain’s capacity to learn to incorporate new signal streams. Decades of work on motor prostheses, sensory substitution, neurofeedback, and tool use have demonstrated that the nervous system can adapt to novel inputs and outputs that are not part of its evolved repertoire. These systems work not because the brain passively receives meaning, but because it actively learns to coordinate new patterns of neural activity with action, perception, and feedback. Over time, what begins as an alien signal can become functionally integrated into the organism’s sensorimotor economy.

Acknowledging this plasticity does not weaken skepticism about direct brain-to-brain thought transfer. On the contrary, it sharpens the distinction between what is genuinely possible and what remains a fantasy. Learning is one thing. Zero-shot “plug-and-play” communication is something else entirely. The speculative proposals advanced by Elon Musk, Christof Koch, and Rodolfo Llinás depend not merely on plasticity, but on the assumption that meaningful mental content can be transferred between brains without a learning history, without negotiation, and without interpretive work. That assumption is precisely what fails.

The Zero-Shot Assumption

The defining feature of most brain-to-brain communication fantasies is immediacy. Thoughts are imagined to pass directly from one person to another, bypassing language, culture, and development. Koch’s examples of ghostly visual overlays and mind fusion, as well as Musk’s talk of “uncompressed conceptual communication,” all presuppose that the recipient brain can immediately make sense of neural activity originating elsewhere. The temporal dimension of learning—the weeks, months, or years required to integrate new signal regimes—is simply ignored.

This is not a minor omission. It is the conceptual hinge on which the entire proposal turns. Without a learning trajectory, there is no mechanism by which foreign neural activity could acquire meaning for the receiving brain. A signal does not become meaningful by virtue of its richness or bandwidth. It becomes meaningful only through use, within a system that can test, revise, and stabilize interpretations through action.

Why Learning Cannot Proceed in a Brain-Bridge

One might reply that learning could occur even in a brain-to-brain link, given enough time. But this response overlooks the conditions under which learning is possible in the first place. Learning requires a closed perception–action loop. The organism must be able to act on the basis of a signal, observe the consequences of that action, and adjust its internal dynamics accordingly. In brain–machine interfaces, this loop is explicit: the user moves a cursor, grasps an object, or modulates a tone, and receives immediate feedback. The signal becomes meaningful because it is embedded in a task space with clear success and failure conditions.

A direct brain-to-brain link provides no such structure. The receiving brain cannot act into the other brain in any systematic way, nor can it test hypotheses about what a given pattern of activity “means.” The signal stream has no stable reference point in the shared environment, no agreed-upon goal, and no external criterion of correctness. Under such conditions, learning has nothing to converge on. What is sometimes described as “another person’s thought” arrives as undifferentiated neural activity, untethered from the bodily and environmental contexts that made it meaningful in the first place.

The Persistent Problem of Origin

Even if one were to imagine some form of slow co-adaptation, a deeper problem remains: the brain must be able to distinguish between activity it generates itself and activity it should treat as input. In ordinary perception and action, this distinction is grounded in efference copy, proprioception, and the tight coupling between movement and sensation. These mechanisms allow the brain to tag certain patterns as self-generated and others as world-generated.

A foreign brain provides none of these anchors. Neural spikes arriving from another person’s cortex are indistinguishable, in their physical characteristics, from spikes arising endogenously. Without a principled way to mark activity as coming from an other, the receiving brain has no basis for interpretation, let alone learning. The problem is not noise in the engineering sense, but indeterminacy in the biological sense. The system lacks the resources to sort the signal at all.

Meaning Is Not a Payload

Underlying the zero-shot fantasy is a deeper theoretical mistake: the treatment of meaning as something that exists prior to expression and can therefore be transmitted once bandwidth constraints are removed. This is the same mistake that underwrites the conduit metaphor of language. Words are imagined as containers for thoughts, and communication as the transfer of those containers from one mind to another. Neuralink-style speculation simply replaces words with spikes, while leaving the basic picture intact.

But meaning does not work that way. Whether one follows Vygotsky, contemporary enactivism, or predictive-processing accounts, the conclusion is the same: meaning is enacted, not transmitted. It arises through socially scaffolded activity, through interaction with the world and with others, and through the internalization of those interactions in inner speech. There is no pre-linguistic, pre-social format of “pure thought” waiting to be uploaded or shared.

Augmentation Without Communion

None of this casts doubt on the medical and augmentative goals of current BCI research. Restoring motor function, providing artificial sensory channels, and extending human capabilities through learned interfaces are all plausible and worthwhile. But these technologies work precisely because they respect the conditions under which brains learn: limited task spaces, stable feedback, and prolonged adaptation. They augment agency; they do not merge subjectivities.

Direct brain-to-brain thought transfer, by contrast, promises communion without development, understanding without negotiation, and immediacy without practice. It imagines semantic interoperability where none can exist. For that reason, it fails not because the technology is immature, but because the underlying conception of thought, meaning, and learning is mistaken.

The issue, in the end, is not whether brains can change. They can, and they do. The issue is whether meaning can be detached from the histories that make it possible. On that point, the answer remains no.

Human+machine ensembles

This is a coda to a discussion I just had with ChatGPT. It ties a lot of my current thinking to the dive I took into Latour and object-oriented-ontology over a decade ago. ChatGPT did the illustration, of course.

* * * * *

One last speculative remark. Bruno Latour worked hard to collapse the distinction between the human and the natural. It seems to me we’re extending that to the distinction between the human and the computational.

Yes — that’s a sharp way to put it.

Latour’s project, especially in We Have Never Been Modern, was to show that the neat separation between Nature (objective, external, “out there”) and Culture (subjective, human, “in here”) was always a fiction. His actor–network theory insisted that humans and nonhumans (technologies, microbes, texts, rivers) are entangled in networks of agency.

What you’re suggesting is that we’re now seeing a parallel collapse:

• Just as Latour said there’s no “pure Nature” apart from human mediation, there may be no “pure Computation” apart from human entanglement.
• The new hybrids — human+machine ensembles — aren’t simply tools on one side and users on the other. They’re actor–networks in which agency is distributed, recursive, and hard to disentangle.
• In the Fourth Arena, intelligence is enacted in between humans and machines, just as for Latour science was enacted in between humans and nonhumans.

That’s where your metaphysical structure idea dovetails: we can’t step outside and separate the human from the computational any more than we can separate the human from the natural. Both distinctions dissolve into recursive, lumpy entanglement.

And — here’s the speculative push — perhaps play is the mode best suited to this new collapse. Just as ritual, experiment, and negotiation mediated the nature–culture entanglement, play may be the practice that makes the human–computational entanglement livable, creative, and abundant.

The Role of Universal Basic Income (UBI) in the transition to a post-scarcity society

Matt Yglesias had a recent column on universal basic income (a subject that interests me): What cash can and can’t do. Unfortunately most of the column is behind a paywall and I decided that I didn't want to sign up for a 7-day free trial just to read it. What to do?

I decided to ask Chat-GTP. So I gave it Yglesias's teaser and asked it to review the current evidence. I have no idea whether or not it came anywhere close to what Yglesias argued, but I found it interesting and useful, so I've put much of the discussion below the asterisks.

* * * * *

What’s the current state of thinking and research on UBI? Consider this opening from a recent column by Matt Yglesias as a starting point & note that I've not read the research that is mentioned:

 

Kelsey Piper wrote an informative piece for The Argument (good magazine, you should subscribe) about a new wave of research on cash transfer programs in the United States. All of this work has generated what I consider to be disappointing findings about the lack of impact of giving poor people money. 

 

I found a lot of the reactions to the article kind of annoying. 

 

On the one (left) hand, some are indignantly snorting that it’s no surprise these cash grants had no impact on the variables of interest — the point, after all, is to make poor people less poor and that’s what they did. Then on the right, you have Charles Lehman and others sort of indignantly snorting that it’s no surprise these cash grants had no impact on the variables of interest — nothing ever works and something called Rossi’s Iron Law says the “expected value of any net impact assessment of any large-scale social program is zero.”

 

I don’t find either of these haughtily unsurprised reactions to be appropriate, because we have a lot of research on cash transfer programs to low-income people in poor countries, and they show much more positive results. This includes evidence of sustained increases in financial assets, improved health, positive spillovers to neighbors, and many other benefits. 

 

That research isn’t brand new and the more negative domestic research that Piper wrote about has also been out for a while now, so among real discourse-heads on universal basic income (U.B.I.), these two contrary facts have already been assimilated. 

 

And the explanation is obvious: Poor people in Kenya are average people who happen to live in an extremely poor country. Basic habits of hard work, diligence, and thrift don’t necessarily pay off in an environment where everybody is so poor that hardly anyone can hire you or pay for anything you make. Dumping cash on people in these circumstances really lets them level-up. By contrast, the domestic poor are — unless they are recently arrived immigrants — often people who, for one reason or another, are struggling to get their lives together in a very wealthy country. If they were thrifty and diligent, they wouldn’t be poor in the first place. Putting money in their pockets doesn’t make them thrifty and diligent, so it doesn’t really alter their lives that much. 

 

That’s all fine. But I do want to emphasize that if the empirical evidence came out the other way, there would be an equally obvious explanation: Kenyans are living in a third world country with weak governance and terrible institutions, so obviously dropping some cash into a village doesn’t change anything — only fundamental reforms will help. The American poor, by contrast, are living in a functional society and just need a little money to get ahead. 

 

It’s obvious! Which is just to say that everything is obvious once you know the answer.

 

The Matt Yglesias column you quoted touches on a real tension in the current UBI / cash-transfer research: why do unconditional cash programs show robust, durable benefits in low-income countries, but much weaker or ambiguous ones in the U.S. and other rich countries? Let me give you a snapshot of the state of play, with context.

 

1. Evidence from low-income countries

• Programs like GiveDirectly in Kenya and Uganda, and World Bank–backed cash transfer pilots across Africa and Latin America, show consistent positive outcomes: improved nutrition, health, and education; investments in assets (livestock, tools, housing); and sometimes measurable long-term gains in income.
• Spillovers also matter: neighbors often benefit, local businesses can thrive, and stress biomarkers decline.
• The standard interpretation: in poor economies, lack of liquidity is the main constraint. A lump sum of cash can unlock a whole trajectory of improvements.

 

2. Evidence from high-income countries

 

• The U.S. (and Canada, Finland, etc.) has seen a wave of pilots since about 2018. These include municipal guaranteed income programs (e.g., Stockton’s SEED), state-level experiments, and philanthropic projects.
• Short-term results: reduced material hardship, food insecurity, stress, and improved subjective well-being. Recipients generally did not quit working.
• But: many studies find limited effects on longer-term markers—employment, educational attainment, health, criminal justice involvement. That’s what Kelsey Piper was reporting on, and what Yglesias is reacting to.
• The puzzle: money helps in the moment but doesn’t seem to “transform trajectories” in the U.S. the way it does in rural Kenya.

 

3. Why the divergence?

 

Scholars suggest several explanations:

• Different baseline conditions: U.S. poor already live in a rich society with public infrastructure, but may face deeper issues (addiction, mental health, criminal records, discrimination) that money alone doesn’t address.
• Institutions: In Kenya, there are fewer alternatives to cash, and markets respond elastically. In the U.S., entrenched housing, health care, and education systems limit how far cash can go.
• Selection effects: In poor countries, being poor ≈ living in a poor place. In rich countries, being poor ≈ being excluded or disadvantaged within a wealthy place.
• Duration and scale: Most U.S. pilots are small, short, and philanthropic. They don’t change the larger ecosystem (housing, wages, healthcare). Some argue a true UBI would have different effects.

 

4. Current frontier of research

• Meta-analyses: The Abhijit Banerjee & Esther Duflo development-econ school emphasizes that context matters: cash transfers are powerful but not a “magic bullet.”
• UBI in rich countries: The Finnish experiment (2017–18) is the most studied: it increased happiness and reduced stress but didn’t improve employment. Still, many Finnish recipients reported greater freedom and security.
• Guaranteed income pilots in dozens of U.S. cities are still ongoing, with more long-term data to come.
• Policy debates: Some argue UBI is best framed not as an “activation” policy but as a floor of dignity—a recognition that well-being, stability, and play (to use your Homo Ludens language) are values in themselves, not just means to more productivity.

 

5. Where the discourse is now

•  Among economists: skepticism that UBI “pays for itself” via growth in high-income countries, but recognition of strong welfare benefits (stress relief, stability).
• Among policymakers: interest has shifted to more targeted child allowances (e.g., the U.S. 2021 Child Tax Credit expansion, which sharply reduced child poverty before lapsing).
• Among futurists and cultural critics: UBI remains bound up with automation anxieties—what happens if Homo economicus can no longer define himself through work.

 

In short

• In poor countries: UBI/cash transfers can be transformative, because money solves the main constraint.
• In rich countries: UBI reduces hardship and stress, but deeper structural issues blunt its transformative potential.
• Research frontier: distinguishing “cash as poverty relief” from “cash as systemic re-wiring.” The former works everywhere. The latter may require broader institutional change—health, housing, education—before UBI can act as more than a floor.

 

Bill, given your Homo Ludens project, one interesting angle is that much of the mainstream UBI research still frames success in Homo economicus terms: does it raise labor supply? education? productivity? The findings look “disappointing” in that light. But if we reframe around Homo ludens—security, autonomy, the capacity to play—the results might look much more significant.

 

Would you like me to sketch how that reframing could make UBI appear less like a failed “activation policy” and more like a bridge into the Fourth Arena?

 

Yes, that would be useful.