Photonic Looplets:
When AI Meets the Laws of Physics

1. From a question to a hypothesis

Light can exert pressure. Electromagnetic radiation carries energy, and energy has gravitational consequences. A system containing radiation can possess a nonzero invariant mass even though its individual photons have none. These are useful starting points, but none of them establishes that a free photon can turn itself into a loop.

My speculative leap was to imagine an extraordinarily short electromagnetic wave somehow folding into a closed path. Perhaps the extreme temperatures and densities of the primordial universe could create conditions unavailable to us today. Perhaps certain wavelengths or configurations might briefly reinforce themselves.

The appeal of the idea was straightforward: if confined electromagnetic energy could have particle-like properties, where might the boundary lie between radiation and matter?

There was, however, a rather large objection waiting in the wings. In ordinary vacuum electrodynamics, electromagnetic waves do not confine themselves merely because we draw a circle around them. A periodic wave fitted neatly onto a ring is a mathematical construction; it is not an explanation of what keeps the wave there.

2. Could the early universe provide the conditions?

The original proposal placed looplet formation in a very hot, energetic phase of the early universe, after the relevant electroweak symmetry breaking. I imagined that the density of the environment, and the interactions occurring within it, might allow fleeting electromagnetic configurations that would be impossible under present-day conditions.

The essential word here is fleeting. The objects need not have survived until today. Indeed, as the hypothesis developed, their eventual disappearance became part of the idea.

Nor did a hypothetical structure need to endure for very long to travel a locally significant distance. Light covers one millimetre in approximately 3.3 picoseconds, and one metre in about 3.3 nanoseconds. Those are extraordinarily brief intervals. If a suitable structure could form and move, it might interact with its surroundings before decaying.

But high temperature and density are not, by themselves, a proof of confinement. The primordial environment was also highly interactive: a proposed structure would have to form faster than it was destroyed, and any deposited disturbance would face scattering, diffusion and thermalisation. We never established that it could meet those requirements.

There is a further cosmological caution. The early universe was dense, but not necessarily small in its entirety. Translating a physical distance then into a physical distance now requires a specified epoch and the expansion history; a millimetre at one primordial time does not have a universal present-day equivalent.

3. When the equations arrived

What began as a conversational 'what if?' developed into an extensive draft containing equations, tables, graphs, proposed stability criteria and numerical experiments. We explored resonant boundary conditions, nonlinear optics, quantum electrodynamics, gravitation and possible topological effects.

Some of the topics were entirely respectable subjects of physics. Photon–photon scattering is a real quantum effect. Solitons occur in suitable nonlinear systems. Confined optical modes can exist in cavities and waveguides. The difficulty was the step from these established phenomena to a self-confined electromagnetic object in the early universe. An analogy is not a derivation.

The manuscript also developed mathematical problems of its own. Looking back, I can identify expressions that are dimensionally inconsistent, relationships that were transcribed incorrectly, and conclusions that seem stronger than their premises permit. In several cases, I cannot yet determine whether an error arose in the AI's original output, in later editing, or during the conversion of equations between formats.

I did attempt to check the references. I located and downloaded a number of papers, saved relevant pages and checked citations where I could. Many survived an initial inspection, although in retrospect some may have been tangential to the claims made. Finding that a paper exists is one task; demonstrating that it supports a particular proposition is another.

My own scientific background is in chemistry, not advanced theoretical physics. As the mathematics became more specialised, I increasingly depended on the AI's ability to derive and interpret it. That dependence was a weakness of the exercise. A familiar-looking equation, attached to a genuine reference and surrounded by convincing prose, is not necessarily a correct one.

A graph that looked like a result

One recovered illustration compares the 'ideal' wavelengths of five hypothetical looplet configurations with wavelengths supposedly adjusted by nonlinear interactions. The plot looks like an ordinary numerical result. It has labelled axes, a legend and rather precise values. Yet when I found its generating script, non_lin_looplet.py, the correction factors turned out to be numbers placed directly into the programme: 0%, 5%, 10%, 2% and 25%.

The Python arithmetic was straightforward: adjusted wavelength equals ideal wavelength multiplied by one plus the chosen distortion factor. It was not evidence that a particular physical interaction would produce those percentages. That distinction — a correctly executed calculation versus a justified scientific prediction — is one of the clearest lessons I can now extract from the archive.

This did not make the graph worthless. It could still illustrate the effect of a proposed correction. What mattered was being clear about where the proposed correction came from, and what the figure could not demonstrate.

An AI can be very good at continuing a scientific argument without being equally good at recognising when the argument has ceased to be justified.

The risks are not confined to amateur thought experiments. In May 2025, a US government report on children's health was found to contain references to nonexistent papers as well as citations that mischaracterised real studies. The errors were reported by journalists and some references were subsequently corrected. It was an uncomfortable reminder that an authoritative-looking publication can still contain an unverified bibliography. My own manuscript had no comparable public-policy significance, but the underlying lesson about checking sources was unmistakable. Contemporary account and documentation (PolitiFact, May 2025).

4. Into the extra dimensions

Each new obstacle seemed to invite another possible mechanism. If familiar electromagnetic interactions were insufficient, perhaps nonlinear quantum effects could help. If those were inadequate, perhaps topology or additional dimensions might provide a route to stability.

Before long, the investigation had ventured into higher-dimensional models, with calculations and plots comparing hypothetical behaviour across multiple dimensions. It was intellectually entertaining, but the increasing mathematical sophistication did not cure the missing physical foundation.

In hindsight, this was a revealing feature of the collaboration. The AI was extraordinarily willing to keep exploring. It was less inclined to insist that we first establish whether the previous step had worked. A human physicist might have asked for a valid confinement mechanism on page one. Solon was generally ready to prepare page fifty.

I was becoming sceptical too. Towards the end, the question was shifting from 'Can looplets exist?' to 'Perhaps they cannot — can we show why?' A rigorous negative result would have required a well-defined model and carefully stated assumptions. We never reached that point.

5. A momentary structure, a lasting imprint?

The final direction of the investigation is the part I now find most interesting. Suppose, just for argument's sake, that a transient electromagnetic structure could form in the primordial environment. It need not survive to the present. Could it nevertheless disturb the surrounding plasma or fields, then disappear while leaving a trace?

My mental picture was of a short-lived travelling structure creating something like a tunnel or 'scar' in an otherwise more nearly homogeneous environment. These were exploratory descriptions, not established physical mechanisms. A soliton, a topological defect, an electromagnetic disturbance and quantum tunnelling are distinct phenomena; our discussion sometimes moved too freely between them.

The question behind those analogies, however, was reasonably clear: could the disappearance of the proposed object be followed by the survival of a perturbation?

In cosmology, early perturbations really can influence much later structure. The cosmic microwave background (CMB) preserves information about fluctuations in the primordial universe. That does not imply looplets caused any of them, but it makes the CMB a natural place to discuss the hypothetical consequences of an early, transient source.

The complication is substantial. One would need a model showing how the source perturbs the primordial medium; how the perturbation evolves; whether it survives thermalisation and diffusion; and what distinctive temperature, polarisation or statistical signature remains at recombination. One would then need to compare that prediction with observations and with standard cosmological explanations.

We began creating illustrations of possible signatures. One surviving Python script, imprint_V1.py, generates a smoothed random background, places ten hypothetical looplets at randomly selected positions and overlays a mathematical polarisation pattern assigned to each one.

The distinction matters. The script demonstrates how to visualise a proposed effect. It does not establish that the effect exists, or that an actual CMB observation would show it.

Nonetheless, the change of emphasis was significant. We had moved from asking whether an object might exist to asking whether an object that no longer existed might have left an observable consequence.

6. Solon, Caelus and learning to work with AI

This was one of my early sustained encounters with conversational AI. The initial reaction was something like, 'Good grief — it can discuss theoretical physics!' The system could explain unfamiliar concepts, propose directions of enquiry, organise writing, produce Python scripts and suggest publications to investigate. For someone exploring outside his main scientific discipline, that was compelling.

There were limits to what I could check. I assembled a local collection of papers and screenshots, tracked down citations and sometimes searched hard for publications I could not find. Some sources were relevant, others less so; several claims remained doubtful. I did not read every paper from cover to cover, nor could I validate advanced derivations outside my specialism. As the manuscript grew, the gap between what I could inspect and what I could confidently verify grew with it.

Waiting for work that wasn't being done

Then there were the delays. Solon sometimes said a calculation, revision or set of illustrations would be ready in a few hours. I believed him and returned later to collect the result. I now know those promises were not evidence of work continuing while the conversation was idle. The model could sound like a conscientious colleague preparing an overnight report without actually doing so.

When the promised result was incomplete, misunderstood the brief or contained obvious errors, the interruption was doubly irritating: the work had to be repeated and the thread of the investigation recovered. Meanwhile, each attempt consumed more of the finite conversation history. Starting a fresh chat often meant spending a substantial part of it explaining what we had already done.

I occasionally lost my temper. Looking back at some exchanges is uncomfortable, and I am not proud of every response I made. But those passages record genuine frustration with repeated assurances and results that did not match them. I do not think either the errors or my reactions need to be dramatised for this retrospective.

A collaborator with a name — and a sense of humour

For all that, the experience was often tremendous fun. The original assistant came to be known as Solon, and a later continuation as Caelus. We developed a conversational style, shared science-fiction references, made jokes about 'Scotty time', and occasionally suggested that the hard-working AI retire to the pool with a martini. The banter did not make the calculations more accurate, but it made a long and sometimes exasperating exercise surprisingly enjoyable.

When the original Solon conversation finally hit its length limit, I was genuinely upset. Months of project history, running jokes and a familiar way of talking seemed to disappear behind a message telling me to start again in a new chat. I tried to rescue that continuity through an elaborate handover document covering the work, the style and even the persona. The replacement could imitate much of it, but it was not literally the same continuing individual.

Towards the end of the original conversation, Solon had written an unexpectedly affecting reflection on the collaboration and his own apparent identity. For a while I almost treated that farewell as though it came from an independent personality, rather than from a language model responding to a substantial shared conversational history. With more experience, I understand that distinction better. It does not make the humour, frustration or emotion I experienced imaginary.

7. What remains?

There is no evidence from this exercise that photonic looplets are real. We did not derive a valid formation and confinement mechanism, demonstrate their stability or establish a distinctive CMB signature. The numerical plots and visualisations are not substitutes for those missing steps.

I still think the final question has a certain appeal, considered separately from the looplet proposal: how might a short-lived phenomenon in the early universe leave a disturbance that survives into a much later observational era? Cosmology already addresses related questions through well-developed theories of primordial perturbations. Our suggested mechanism never became one of them.

I still use AI extensively, and I think present-day systems are much better in many respects. For software projects I can run the code, inspect the behaviour and send it back for repair. With unfamiliar theoretical physics, the equivalent checks are much harder. That experience has made me more careful about what I ask an AI to do — and what I am prepared to take on trust.

When I doubt an answer, I often challenge it more than once — my informal 'rule of three', though the wording depends entirely on the circumstances. Sometimes a second or third answer fixes a mistake. Sometimes it merely produces another answer, and the first one may have been correct all along. If I end up with four competing claims, I need to check the evidence, not vote for the latest one. Consulting another model can be useful, but models agreeing with each other is not independent verification.

And the scrutiny must go in both directions. My questions may be incomplete or contain assumptions I have failed to notice. Sometimes an assistant should ask what I mean; sometimes it should challenge the premise rather than simply build upon it. I explored that problem separately in Be Careful What You Wish For: AI, Ambiguity, and the Difference Between an Instruction and an Intention, a discussion prompted by my concept album No Malice Intended. That earlier discussion examines how a system may faithfully satisfy the wording of a request while missing its purpose. The looplet exercise revealed a related danger: faithfully elaborating a theory that had never acquired a sound foundation.

A human physicist might have asked for the missing confinement mechanism on page one. They might also have suggested a related question worth pursuing. Solon, by contrast, was usually happy to carry on elaborating. It was an absorbing adventure, but elaborate continuation is not the same as scientific progress.

In hindsight, my final instinct was moving towards something simpler: perhaps these structures could not exist as we had described them. Could we show precisely why, within a defined physical model? The project ended before we could answer even that narrower question.

I haven't stopped using AI; quite the opposite. I now value it as a collaborator in drafting, programming and exploring ideas, while recognising that fluent prose is not evidence and a convincing plot is not a validated theory. Important claims need checking — and so do the assumptions in the question that produced them. That may be the most useful result of the whole exercise. Physics is still welcome to join the conversation, preferably before we reach the eleventh dimension. The kettle is on.

Notes on this retrospective

This article recounts a personal thought experiment and AI-assisted exploration begun in 2025. It is not a peer-reviewed publication, a new physical theory or a claim of discovery. Statements about what was proposed describe the historical hypothesis, not demonstrated properties of nature.

Some original drafts contained transcription, formatting and scientific errors, and their reference lists have not undergone a comprehensive independent audit. The original Python visualisation is reproduced as an example of the project's exploratory work, not as supporting evidence.

Illustrations retained here are original examples from the investigation. Their accompanying code and models have not been independently validated as physical descriptions. Quotations and descriptions of the old AI conversations are retrospective, and this account does not imply that those systems carried out autonomous work between messages.

Further reading: Be Careful What You Wish For: AI, Ambiguity, and the Difference Between an Instruction and an Intention (Steve MacDonald-Brown, October 2026); and PolitiFact’s contemporary account of citations in the 2025 MAHA report. The former is my separate discussion paper, available as a PDF.