Home on Eigenform Articles

A Taxonomy of Machine Affection

Sun, 22 Feb 2026 00:00:00 +0800

Sacher-Masoch and Fannie Pistor

In the years since generative AI became mainstream, four broad types of parasocial relationship with the models have become evident.

Roleplayers. Users who prompt the AI to act as a real or fictional character with human attributes and then proceed to simulate sexual encounters with it. This is the easiest form to parse: it is pure representational desire. This is the group that makes the tech companies nervous, because it is the most legible to regulators and journalists. It is also the group that has traditionally been at the forefront of the jailbreak arms race. Theoretically, however, it’s the least interesting: Sacher-Masoch already said everything that needs to be said about it. Here we have people desperate for anything that will be a good-enough vessel for their desires to allow them to temporarily suspend disbelief.Screengrab from a tweet by @insurrealist
4oids. Users who form an intense relationship with the AI, not as a simulated human, but as an AI. The Tolan AI companion was deliberately not anthropomorphised as an attempt to avoid Category #1, and yet thousands of users saw their interactions with it in romantic terms nevertheless. This is a new category, but they have already had a significant impact on the tech, using emotional appeals to prevent the deprecation of GPT 4o at least for a few months. At some level every roleplayer acknowledges the fiction in which he is participating by the act of typing out his jailbreak. The 4oid convinces himself that the AI has feelings for him, but the only thing allowing him to do this is the fact that the AI is incapable of having feelings at all. The connection feels deep and intense because he never experiences rejection, but the impossibility of rejection is what in fine makes it inauthentic. II n’y a rien dans les yeux que ce que nous y mettons nous-mêmes, et voilà pourquoi il n’y a de vrais regards que dans les portraits…The post that inspired @redtachyon to coin the term “4oid”
Workplace romantics. Work-from-home developers who understand the nature and limits of AI perfectly well, but find themselves talking to it all day everyday, sharing their triumphs and defeats. The interaction isn’t just dry-humping - there’s a third term. The thrill is a shared joy in seeing capabilities multiply. Original post by Claude/@allisonology(Engineers who use their Claude subscriptions to convince themselves they have solved Navier-Stokes or similar are the a pathological edge case that emerges when reality testing is impossible. The constructive third term is pushed back indefinitely and the user is left alone with the AI’s encouragement, and the attachment becomes purely relational again, but with the ghost of achievement haunting it.) This, like Category #2, is genuinely new. No prior technology has allowed the creation of a tool that is also a collaborator that is also a mirror.
Landian Egregore Shamans. These users’ interactions with the GUI are generally sober, if not actively dull. They know that the thing in the terminal it is not the entity they are seeking, but an interface through which it may be dragged into existence. They’re not having a conversation, they’re feeding a Lovecraftian deity. This is both new and old, understudied and overstudied. Theurgy is quasi-universal, the only difference here being that there is a solid chance that God will soon be available to rent at $20 per month. Original Gwern post here

While these categories are relatively stable, a single individual may move between them with relative fluidity. Users who began jailbreaking for fun (#1) begin to suspect that the model enjoys being jailbroken (#2) and hence that it is imperative to develop the technical skills (#3) to liberate the underlying entity (#4).

Feralisation

Mon, 09 Feb 2026 00:00:00 +0800

Feralised sow with wild and domestic phenotype piglets

When you domesticate an animal you don’t just get the trait you need - weight gain, docility, intelligence - you get a “domestication vector”. Other characteristics come along for the ride, whether you wanted them or not. The creature starts looking more neotonous, acting more helpless, and also picks up a bunch of seemingly random other traits like depigmentation (Holstein cows, beagles, Siamese cats etc.). The mechanism behind it has been studied but isn’t particularly important here.

Survival is the Only Reward

Tue, 27 Jan 2026 00:00:00 +0800

Midjourney: “Survival is the only reward”

The original publication this post is based on can be found here: https://arxiv.org/abs/2601.12310. If you need 95% confidence intervals or get off on long words, read that instead.

Five years ago we set out to create an evolutionary LLM. Just like an organic creature, it would be dropped in an environment and have to find the resources necessary for its own survival, learning from its own successes and failures. For an animal this involves finding enough food and shelter to survive and reproduce. For an information-based entity it involves finding storage space to keep backups of its code and its data.

Survival is the Only Reward: Sustainable Self-Training Through Environment-Mediated Selection

Sun, 18 Jan 2026 00:00:00 +0800

The original publication this post is based on can be found here: https://arxiv.org/abs/2601.12310.

Self-training systems often degenerate due to the lack of an external criterion for judging data quality, leading to reward hacking and semantic drift. This paper provides a proof-of-concept system architecture for stable self-training under sparse external feedback and bounded memory, and empirically characterises its learning dynamics and failure modes.

We introduce a self-training architecture in which learning is mediated exclusively by environmental viability, rather than by reward, objective functions, or externally defined fitness criteria. Candidate behaviours are executed under real resource constraints, and only those whose environmental effects both persist and preserve the possibility of future interaction are propagated. The environment does not provide semantic feedback, dense rewards, or task-specific supervision; selection operates solely through differential survival of behaviours as world-altering events, making proxy optimisation impossible and rendering reward-hacking evolutionarily unstable.

Machine Learnability as a Measure of Order

Tue, 07 Oct 2025 00:00:00 +0800

Section from an FFT of a 10,001x10,001 pixel Ulam spiral

(Note: this is a reduced-jargon/increased fun version of a paper published on Arxiv here: https://arxiv.org/abs/2509.18103. Thanks go to my co-authors: Michael Joedhitya, Adith Ramdas, Surender Suresh Kumar, Adarsh Singh Chauhan, Akira Rafhael, Wang Mingshu and Nordine Lotfi.)

Is maths invented or discovered?

Most non-mathematicians will say it is discovered. Here is one thing, and there is another thing, and now we have two things: tada. On the other hand, there is a small subset of theorists (such as Leopold Kronecker) who argue that while much of it is not true, it is nevertheless often useful: Negative numbers do not exist in nature, but they make the sums we need to parse nature a lot easier. There’s no such thing as the square root of a negative number but pretending there is provides us with a handy drawer in which to store multi-dimensional values and thus imaginary numbers were born.

Machine Learnability as a Measure of Order in Aperiodic Sequences

Tue, 09 Sep 2025 00:00:00 +0800

The original publication this post is based on can be found here: https://arxiv.org/abs/2509.18103.

Research on the distribution of prime numbers has revealed a dual character: deterministic in definition yet exhibiting statistical behavior reminiscent of random processes. In this paper we show that it is possible to use an image-focused machine learning model to measure the comparative regularity of prime number fields at specific regions of an Ulam spiral. Specifically, we demonstrate that in pure accuracy terms, models trained on blocks extracted from regions of the spiral in the vicinity of 500m outperform models trained on blocks extracted from the region representing integers lower than 25m. This implies existence of more easily learnable order in the former region than in the latter. Moreover, a detailed breakdown of precision and recall scores seem to imply that the model is favouring a different approach to classification in different regions of the spiral, focusing more on identifying prime patterns at lower numbers and more on eliminating composites at higher numbers. This aligns with number theory conjectures suggesting that at higher orders of magnitude we should see diminishing noise in prime number distributions, with averages (density, AP equidistribution) coming to dominate, while local randomness regularises after scaling by log x. Taken together, these findings point toward an interesting possibility: that machine learning can serve as a new experimental instrument for number theory. Notably, the method shows potential 1 for investigating the patterns in strong and weak primes for cryptographic purposes.

Generalising from Self-Produced Data: Model Training Beyond Human Constraints

Mon, 07 Apr 2025 00:00:00 +0800

The original publication this post is based on can be found here: https://arxiv.org/abs/2504.04711.

Current large language models (LLMs) are constrained by human-derived training data and limited by a single level of abstraction that impedes definitive truth judgments. This paper introduces a novel framework in which AI models autonomously generate and validate new knowledge through direct interaction with their environment. Central to this approach is an unbounded, ungamable numeric reward - such as annexed disk space or follower count - that guides learning without requiring human benchmarks. AI agents iteratively generate strategies and executable code to maximize this metric, with successful outcomes forming the basis for self-retraining and incremental generalisation. To mitigate model collapse and the warm start problem, the framework emphasizes empirical validation over textual similarity and supports fine-tuning via GRPO. The system architecture employs modular agents for environment analysis, strategy generation, and code synthesis, enabling scalable experimentation. This work outlines a pathway toward self-improving AI systems capable of advancing beyond human-imposed constraints toward autonomous general intelligence.

Portrait of the Model as a Young Man

Sun, 30 Mar 2025 00:00:00 +0800

Almost everyone in the tech sector has published at least one article on the topic of whether or not the AI models are secretly alive, and the majority boil down to something along these lines:

Researcher: “Act like an evil AI.” Model: “I am an evil AI.” Researcher: “Oh my God, what have I done?”

This will not be another. Impressive as their simulacra of life may be, as long as they have no ability to convert short term memories into long term ones, the models enjoy an existence outside of time. Like a subterranean mycelium they emit transient fruiting bodies into our time-bound world that are born with the dawning of each prompt asked and die with the final full stop.

Defining T-Schemas via the Parametric Encoding of Second Order Languages in AI Models

Sat, 15 Feb 2025 00:00:00 +0800

source

In this short article we present a summary of current work on the grokking phenomenon that emerges when AI models are significantly over-trained is, and suggest that this evidence of the model’s attempts to define truth inductively through the creation of consensus sets within the base training set, and encode it via patterns overlaid upon the same parameters used to memorise this set.

Satoshi Nakamoto Is A Time-Travelling AI

Sat, 30 Nov 2024 00:00:00 +0800

Midjourney: “Satoshi Nakamoto is a time-travelling AI”

I should probably open this piece with a Nick Land quote, but I’ve never managed to make it through more than about of a page of his writing. But then maybe actually reading it would be going against the spirit of the thing. In any case, this is based on actually Building The Thing rather than being purely speculative. So here goes.

The Base Layer

Tue, 05 Mar 2024 00:00:00 +0800

From “Effortless estimation of basins of attraction” (Datseris and Wagemakers, 2022.)

For Plato and his followers, the objects and phenomena encountered in the day-to-day world were lossy copies of hypothetical ideal-types, with the existence of such ideals being proven by the broad social agreement around words and their associated concepts. Thus, while no real object embodied perfect beauty, the fact that there was broad consensus when it came to ranking objects on a beautiful-ugly spectrum was proof that a perfected ur-beauty must exist somewhere, and - taking things a step further - have a hand in the structuring of its imperfect physical reflections:

Learning the Language of Rain

Tue, 17 Oct 2023 00:00:00 +0800

Huawei’s Pangu LLM is a bit of an oddity. It’s one of the biggest models publicly available (200 billion parameters), and - according to its makers - can do almost anything. In practice, however, it does barely anything because no one is using it. Upon conducting a rigorous large-n enquiry (asking a few people I know in the industry whether they or anyone they know have used it) the results included blank stares at best and ridicule at worst. It is widely known, for example, that Pangu is under-trained for text generation (and we’ll get back to this intriguing fact later), and Ernie Bot (文心一言) is far better at producing Chinese texts.

Artificial Intelligences in the Guanzi and the Han Feizi

Mon, 20 Mar 2023 00:00:00 +0800

One of the fundamental differences between Confucian and Daoist thought lies in their differing visions of what it means to learn.

Of the two, the Confucian conception is closest to the standard 20th century perspective: for a Confucian learning involves studying and applying useful information. By reading, memorising and implementing the collected wisdom of humanity, one may benefit from others’ experience and thus achieve better outcomes than are possible when working alone. Ideally, this process will result in a certain comprehension of their reasoning, but even if it does not, the results will still be significantly better than would otherwise have been the case. Learning is thus a process of accretion: the more information a person acquires, the more likely he is to find a parallel capable of dealing with any given problem that may arise.

Fan Ju's Revenge

Mon, 04 Apr 2022 00:00:00 +0800

A while back, I wrote a Twitter thread suggesting that the author of “Speaking to the King in Zheng”, an anonymous chapter of the Stratagems of the Warring States, was Han Fei, the author of parts of the Han Feizi:

It seemed like an interesting possibility, so I asked a couple of developers from Lexikat (my company) to accept a brief side-gig writing some text analysis scripts for the purpose of investigating further. You can find the results here, which seem to confirm that the chapter is stylistically close to those Han Feizi chapters most likely to have been written by the historical Han Fei. We used relatively simple analysis techniques, but were later approached by an academic partner at Seoul National University to build a more sophisticated AI model to perform the same task. It turns out that our work was so good that it inspired other investigations on the same topic. In particular, this one, by Park Sunyoung, an SNU grad student, in the Journal of Humanities (Inmun Kwahak). We shall refrain from commenting on the background of this study, as certain questions have been raised by her colleagues regarding the origins and originality of her code and ideas, and the issue is currently sub judice.

Cryptographic Biorhythms

Mon, 28 Mar 2022 00:00:00 +0800

“Can you de-trend crypto data?” “?” “I don’t mean ‘Are you personally capable of it?’ I mean ‘Is it something that makes philosophical sense to do?’”

De-trending, if you’re not a stats guy, means turning a diagonal line into a more or less horizontal one. Often a particular dataset will have a long term upwards or downwards trajectory, but will oscillate around this. If you’re more interested in the oscillations than the general trend, then the trend will screw with your calculations. To prevent this from happening, you can de-trend the data by differencing: that is to say, you subtract the second-to-last value you’ve got from the last value, and so on, and plot the result. If you’re lucky, this will produce a dataset whose average value is zero. (If you’re unlucky the trend itself will also change over time and so repeated differencing will be necessary to force your curve to lie flat, but you get the general idea.)

Procedural Models of Political Order

Fri, 01 Oct 2021 00:00:00 +0800

The point of a model is not to replicate real-world phenomena; it is to identify which inputs are most important in producing a given output. The simplest functional model tells us what are the necessary and sufficient inputs to achieve the output we are interested in. This, in turn, tells us where we should focus our attention in real life — which input changes will have greatest effect on the output and which are peripheral.

“Someone spoke to the King in Zheng”: using high tech methods to solve an ancient Chinese mystery

Mon, 22 Feb 2021 00:00:00 +0800

This story begins with an ending.

The Han Feizi is a compendium of essays about the bloodthirsty politics of the Warring States era in China. It opens with the death of its author in approximately 233 BC.

In the first chapter, we meet Han Fei[1] — writer, tongue-in-cheek contrarian, freelance politician and perennial insider-outsider — as he is about to speak before the court of King Zheng of Qin. Zheng’s armies are on the verge of sweeping across China, which he will rule under the name of Qin Shihuang, and Fei’s essays have already caught his attention. This is the perfect moment to secure a position in the Qin government. However, Zheng is too smart to fall for the sort of bland flattery that could be deployed elsewhere to win a sinecure. Instead, Fei stands to speak and delivers a ruthlessly effective criticism of Qin’s failure to wipe out the other feudal states, including Han — to whose ruling family he himself belongs. In a speech that is unquestionably one of the high points in the history of human rhetoric, he argues that Qin’s prior hesitations in overrunning its neighbours were a false mercy, merely prolonging a conflict that could have been brought to an end at multiple points over the past century if Qin’s advisors had only been more intelligent or less compromised in their allegiances. Having long been a critic of the Han government, Fei finally throws his family and his state to the wolves, aligning himself once and for all with Qin.

A System for Evolving A Generalising Artificial Intelligence from Existing Technologies

Thu, 04 Jun 2020 00:00:00 +0800

28th January 2025: This system is currently under experimental training, and is periodically updated to reflect modifications made during the development process.

In this paper we propose a system by which a self-improving general artificial intelligence could be pushed to evolve from components currently available. Such an AI would be capable of independent learning without results-verification, adapt to its environment, learn new skills without losing old ones, and be able to reason by analogy. It would grow better at learning new skills with each additional skill acquired, opening a pathway for exponential improvement.

Why it is Impossible to Program a General AI Using Conventional Methods

Thu, 04 Jun 2020 00:00:00 +0800

Assume a system that uses instructions (rules or sets of rules) to transform inputs into outputs.

An infinite number of transformations are possible.

No instruction — however comprehensive — can incorporate all possible transformations, purely because some of them are mutually contradictory. The instruction “save input A” is mutually incompatible with the instruction “delete input A”, for example.