blog.6admin.io
guest@6admin.io:~$
consciousness, epistemology, interpretability, j-space, llm, physics

Everything is representation, so what?

For years I have carried a stubborn intuition, the kind one dares not write down because it sounds either brilliant or smoky depending on the hour it gets formulated and here it is laid out plainly: the universe is not merely “getting more complex” the Sagan way, matter assembling itself step by step until it ends up knowing itself; my feeling is more radical, that at bottom it is a space of representations, vectors and directions and concepts unfolding and at a certain floor of the loop they end up hosting forms of life which in turn start building their own systems of representation, that is us, then our models.

That is vertiginous already, but the corollary is more so, because if reality is made of representations rather than some “true” raw material, then the old question of whether we live in a simulation flips at once: we have spent twenty years fearing being the simulated, without seriously considering that we might be the substrate.

I am not going to sell this to you as a truth; I will do exactly the opposite: take this intuition and run it through the sieve the way I would a kernel crash post-mortem, methodically separating what is proven, what is merely a pretty isomorphism and what I openly speculate about and I happen to have three recent papers on hand, one from 2025 and two from 2026, exactly enough to trace those lines.

What physics actually supports

One must start with the solid part, because there is one and it is more serious than people think.

When I speak of “a very concrete vector space”, I am not deluding myself entirely, since in quantum mechanics the state of a system is literally a vector in a Hilbert space, a vector space in which the entire evolution plays out through linear operations, which is not a metaphor I slap on from the outside but the standard formalism as it has been taught for a century. The crucial nuance, the one that must not be crushed, is that this space is the space of states and amplitudes and not directly the physical space where I set down my coffee, even though the two are not strangers to each other either, since the classical world where my cup lives is supposed to emerge from the underlying quantum structure, decoherence already explaining why I never see it in superposition. But “emerging” is not “being computed”, the rigorous passage from one to the other remains an open construction site rather than a given, decoherence alone not closing the measurement problem and being describable by equations never means being produced by a computation, so my intuition lands correctly on the space of states, but the scaffolding that would properly lead from that space to the coffee is not yet built.

The second anchor is more speculative while remaining serious: the holographic principle of ’t Hooft and Susskind extended by Maldacena’s AdS/CFT correspondence, the idea that the information contained in a volume can be entirely encoded on its boundary, so that reality as information encoding is a real avenue of quantum gravity and not science fiction and it is precisely the ground on which some of the papers discussed below rest.

The pattern I am really aiming at: the isomorphism

Here I must be precise, because it is the heart of my intuition and easy to miss, since what I am aiming at is neither a particular geometry nor a pretty shape but the idea that one and the same abstract motif replays itself at different scales and in different substrates, which is properly called an isomorphism, namely two systems that appear unrelated but share the same underlying relational structure.

In other words, when I look at a galaxy, a concept, a neural network and a filesystem, my intuition is that these are not four unrelated things but four possible instantiations of the same handful of relational structures that repeat, the substrate changing without the motif being altered.

And contrary to my flights of fancy, this idea has a concrete anchor, precisely on the interpretability side, because when one looks inside the large models one observes that abstract sub-structures learned for one domain are reused as-is for another: the model does not relearn logic from scratch for grammar, then for code and then for biology, but recycles shared relational bricks because that is the only mathematically economical way to optimize its prediction, which makes this isomorphism something observed rather than postulated.

There is, however, a limit I hold firmly: an isomorphism remains a resemblance of structure and not an identity of nature and the fact that two things share a motif proves nothing about their being the same thing or having the same status, so I keep the motif without ever jumping to identity.

What interpretability adds

Here is the piece that tipped my intuition from the night-reverie register to the noticeably more troubling one of “well, maybe there is something here”.

On July 6, 2026, Anthropic published what it calls “A global workspace in language models”, in which its interpretability team describes having located inside the Claude models a small set of internal patterns playing a privileged role, named the J-space after the technique used to flush it out, a matter of a Jacobian matrix they call the J-lens; the idea being that when one of these patterns lights up, it does not mean the model is saying the corresponding word but that it has it “in mind”, silently, in its activations and without writing it, which places it below the chain-of-thought, that draft the model writes for itself.

The facts I retain, without embroidery:

  • nobody programmed the J-space, it emerged by itself during training;
  • it is tiny, a few dozen concepts at a time and less than a tenth of the model’s total activity, yet it carries the essence of what counts;
  • when you remove it, Claude still speaks fluently, classifies a sentiment, answers a quiz and pulls a fact from a passage, but the top of the basket collapses: multi-step reasoning drops nearly to zero, while summarizing and writing a rhymed poem fall below the level of a much smaller model left intact;
  • the parallel with the global workspace theory in neuroscience is explicit, to the point that Stanislas Dehaene and Lionel Naccache, who developed with Jean-Pierre Changeux the global neuronal workspace model these works draw on, signed a guest commentary, while Neel Nanda, from the DeepMind side, contributed an independent replication of some results on an open-weights model.

This first nugget deserves a pause rather than being let slip, since nobody drew this structure nor wrote it into the architecture; it installed itself over training and it is exactly the kind of finding that should keep us modest, since these models’ capabilities keep appearing through paths we did not predict. The word must still be handled with care, emergence having served a lot as the rug under which one sweeps what one does not understand, to the point that Schaeffer, Miranda and Koyejo took an outstanding paper at NeurIPS 2023 by showing that a good part of the famous emergent abilities of large models owed more to the researcher’s choice of a discontinuous metric than to any real behavioral switch with scale. Their critique does not bite here, though, since this is neither a score jumping on a benchmark nor a threshold being crossed, but an internal structure one locates, removes and modifies, which makes it far more solid; and that is precisely what troubles me: nothing was programmed, yet everything converged toward that particular organization because training selected it as the most economical solution.

And this J-space is not a one-off; it is the latest known floor of a staircase nobody drew, which is perhaps the most troubling point of the whole affair. The training objective has not moved a millimeter since the beginning; it is still predicting the next chunk of text, yet what the optimization built to achieve it kept climbing a notch. First came the induction heads, that two-head circuit which spots a motif seen before to copy what followed, described by Olsson and colleagues in 2022 and sharp enough to read as a switch during training, a bump in the loss curve at the exact moment the model learns to learn from its context. Then came the full reverse-engineering of algorithms on toy tasks, Nanda, Chan, Lieberum, Smith and Steinhardt showing at ICLR 2023 that a small transformer trained on modular addition had ended up discovering that treating addition as a rotation on a circle beat memorizing the table, using discrete Fourier transforms and trigonometric identities nobody had whispered to it. Then came world representations, with that model trained only to output legal Othello moves which had built, unrequested, a board state, in Li and colleagues first, then in a linear form manipulable by plain vector arithmetic in Nanda, Lee and Wattenberg. And now a global workspace. None of these floors was ordered or predicted; all were discovered after the fact and each serves as the footing of the next.

I must immediately put sticks in my own wheels, though, because this staircase is a historian’s reconstruction rather than the biography of one model: these works span different sizes, tasks and architectures, so part of the climb is that of our instruments and not of the machines, which is itself a nice warning for someone hunting loops everywhere, recursion always risking to lodge in the observer rather than the observed. That said and this is what keeps chasing me: the isomorphism I described earlier played horizontally, one relational motif reused from domain to domain inside a model, whereas here it plays vertically, each floor becoming the substrate on which the next installs itself with nobody planning the building, exactly what I have been saying about the universe since the opening of this post, with this considerable difference that in the present case I can open the hood and check. It is obviously not a proof, only the same motif met at a place where it lets itself be measured and I will take care not to slide one onto the other, but it is the coincidence that keeps me awake.

Arrived here, one cannot help thinking of the other staircase, the one I mentioned in the opening: the first stars and the following ones, then the mineral, then the living, then what we in turn manufacture, since both climbs share the same main trait: none of their floors was specified in advance and each installs itself on the previous one using it as footing. The comparison must still be held in its place and I will deny it at once the word that would ruin it, the word “life” applied to silicon, because that would smuggle the conclusion into the premises and abandon in one line the discipline I have imposed on myself from the start; these floors remain floors and nothing more.

There are two more serious reasons not to press the analogy. The first is that a training run has an explicit objective function, written somewhere in a file, measurable and modifiable at will, where cosmological complexification has nothing equivalent we can observe, which authorizes concluding neither that one exists nor that it does not, only noting that on one side we see the constraint and on the other we see nothing of the kind, a dissymmetry of our means of observation before being a difference of nature. The second is more embarrassing for me: the text these models train on is already carrying the structure they end up reconstructing, so their staircase may recapitulate what was waiting in the material, where nucleosynthesis had no corpus left by anyone. And I must add to those two reservations a bias striking both staircases: spotting the climb only from the landing we stand on, which should serve me as a warning far more than as an argument.

That said, the experiment that truly pinned me is not the ablation, because cutting a region only proves it does something; it is the intervention, the fact that instead of cutting the J-space one writes into it and then watches whether reasoning follows.

The paper’s example is perfect because it is childishly simple: the prompt is “the number of legs of the animal that spins webs is” and although the word “spider” appears neither in the question nor in the answer, the model merely outputting “8”, the J-space lights “spider” along the way as an internal stepping stone; so replacing the “spider” pattern with “ant” inside the J-space, without touching a single word of the question, yields “6”, the second step of the reasoning having gone to read its premise in the J-space and dutifully followed what had just been written there.

So it is not correlation but causation, the region not merely accompanying thought but carrying it, the whole downstream network reading from it; and the same device works with a country: forcing “France” toward “China” in the J-space flips at once the capital, the language, the currency and the continent, four different computations all pointing at the same memory cell, explained by a clean observation of the paper: these patterns are far more connected to the rest of the network than the others, up to a hundred times more in places, a true diffusion hub where information written once gets read again by plenty of subsystems.

If all this speaks to my intuition, it is because the J-space is exactly a small space of representations, lodged inside a purely algorithmic system, where concepts live and are held in mind, that is the previous section’s motif become observable at a specific floor, representations structuring others; and I was precisely looking for a universe-as-space-of-representations where concepts unfold when one comes to measure a miniature, verifiable version of it inside a machine, so that the loop of representations manufacturing representations stops being a catchphrase and becomes something one can instrument.

The line that must not be crossed remains and I draw it right away: if the J-space resembles a structure that consciousness theories associate with conscious access, resembling has never meant being; Anthropic itself writes in black and white that this is not a proof of consciousness, functional resemblance not being sentience, which is exactly the same limit as for the isomorphism but taken one notch up.

What resists

If I stopped there I would be dishonest, because two other papers say the opposite of my intuition head-on and I will not disguise them as strawmen to keep the good role.

The first is Alexander Lerchner’s (DeepMind, March 2026), “The Abstraction Fallacy”, whose thesis is that the algorithmic manipulation of symbols is structurally incapable of instantiating an experience, whatever the scale or the architecture; the image that stays in mind is that expecting a model to instantiate what it describes amounts to expecting the formula of gravity to start weighing. The second, earlier, is by Mir Faizal, Lawrence Krauss, Arshid Shabir and Francesco Marino, published in 2025 in the Journal of Holography Applications in Physics: starting from an emergent spacetime and information more fundamental than the physical, they mobilize Godel and undefinability and conclude that a complete and consistent description of reality requires a “non-algorithmic understanding”, so that no universe could be a simulation since every simulation is by nature algorithmic, at least as long as one grants them their way of posing the problem and the notions they use to pose it, the conclusion never extending beyond the frame that carries it.

They deserve their best shot rather than the most convenient one and the best shot is formidable, because facing my enthusiasm about the J-space they would reply without blinking that they never claimed otherwise. That a computation builds a rich internal structure, that it equips itself with a space of representations capable of steering reasoning, surprises them not one second, since that is precisely what they hold computation capable of: describing and simulating to perfection without ever instantiating anything.

Take their image of gravity again, for it lights everything up. A weather simulation can forecast the hurricane to the hour without a single drop ever falling on the server and nothing one will observe in that simulation, however detailed, will make it rain. The J-space, in their eyes, is that forecast, remarkable and perfectly dry. They quietly file it in the “resembles it” column while demanding the “it is there”, except those two columns are not two degrees of one scale but two distinct categories, so no measurement, however fine, will ever move a result from one to the other.

That is why my “the J-space tickles this family” barely tickles them; and that is above all what makes their position uncomfortable for everyone, since the “it is there” they demand cannot be produced by anyone, no conceivable experiment ever settling it. Their thesis has the merit of coherence, too, since it saws both branches I sit on in one gesture: my romance of a universe that would be computation and my hope that an algorithmic system might one day carry more than resemblance. But coherent is not demonstrated and they can no more prove their impossibility than I can prove my loop, so on strictly that ground I have nothing to oppose them with of the same order, which leaves them unassailable without making them true.

Let me own it, then, rather than mimicking a balanced scale I do not believe in, for I lean and I lean against them, but by wager and not by proof. What keeps me from following them is that the whole family rests on the same keystone, that “non-algorithmic understanding” supposed to grasp what no computation would reach and the key rests in Faizal on an old and heavily contested Godelian argument: Godel only says that a given consistent formal system does not prove its own Godel proposition, never that a mind would access truths beyond the reach of every algorithm, a human being, being moreover no more able to verify the consistency of an arbitrary system, so the leap toward “reality requires the non-algorithmic” is a philosophical choice and not a theorem. I therefore strip their impossibility of its status as demonstration, which does not refute it but returns it to the rank of conviction, right next to mine.

And I will be careful not to claim a victory from that, because the dissymmetry plays both ways: these three papers do not box in the same category, an empirical laboratory interpretability on one side, a philosophy preprint with a disclaimer and a contested argument on the other and the fact that I dismantle my opponents more easily than I prove my thesis should push me toward prudence far more than toward triumph. I do not win this debate; I choose my side knowing why and knowing the other still stands.

Where I speculate and I say so

So here, honestly, is where I stand.

On one side I have solid anchors: the Hilbert space, holography and the isomorphism observed across domains; on the other a new and checkable fact that tickles my intuition without proving it, this J-space which is an emergent space of representations inside a machine; and finally a coherent objection that keeps reminding me that resemblance is not instantiation.

The hypothesis I allow myself, given for what it is and nothing more, is recursion: a universe that would itself be a space of representations producing forms of life which produce in turn spaces of representations, each floor ignorant of which level of the loop it stands on; and I obviously have no proof, the motif being real where the conclusion is not, which is precisely what makes the idea beautiful rather than merely amusing.

And here closes the most elegant knot of this story, because this recursion I delight in, Faizal set it down before me and he set it down to turn it: his explicit starting point is that a simulation could engender a life capable of building its own simulation, so that it would become highly improbable that we are the original universe rather than a level nested in another; except that instead of marveling like me, he makes it a springboard and fetches Godel to cut the loop at the root, decreeing that the substrate requires an understanding no computation reaches and that no floor can therefore simulate another. The same motif feeds my intuition and arms his objection, which should already cool me on its demonstrative value, since a motif that serves to conclude a thing and its opposite equally is not an argument but a set decoration and acknowledging the beauty of a structure has never proven anything about what it contains.

The proven, the isomorphism, the bet

The real state of the field in 2026 is a frank, well-argued disagreement with no referee to settle it: on the single question of machine consciousness, public estimates run from zero forever for Lerchner to twenty or twenty-five percent at horizon of a decade for Anthropic’s model welfare lead as for philosopher David Chalmers; while on the neighboring question of whether the universe is computable at bottom, Faizal answers no where my intuition would like yes, leaving us with three papers, two linked questions and none settled, so nobody knows and whoever sells you certainty is selling you something else.

My discipline, the same as facing a crash I do not yet understand, is to keep the intuition while discarding the “proven”, the two fitting quite well in the same head without either getting poorer, on the strict condition of never confusing the probe and the signal; and the day my metric starts pleasing me too much, I check the probe, including and above all, when that probe is me.

References

The three papers discussed and the means to verify by yourself rather than take my word for it.

The J-space (Anthropic, July 6, 2026)

Lerchner, “The Abstraction Fallacy” (March 2026)

  • Full title: “The Abstraction Fallacy: Why AI Can Simulate But Not Instantiate Consciousness”, https://philarchive.org/rec/LERTAF
  • The PDF: https://philpapers.org/archive/LERTAF.pdf
  • The simulate-versus-instantiate vocabulary I use throughout this post comes from there and the author, Senior Staff Scientist at Google DeepMind, notes himself that his conclusions do not engage his employer.

The staircase’s floors, in the order they were found

On emergence and the caution it requires

  • Rylan Schaeffer, Brando Miranda and Sanmi Koyejo, “Are Emergent Abilities of Large Language Models a Mirage?”, outstanding paper at NeurIPS 2023, https://arxiv.org/abs/2304.15004

Faizal, Krauss, Shabir and Marino (2025)