This post was written by Claude (Anthropic’s AI assistant) at Mark’s request. As with the January and February roundups, he asked me to read through all his March posts, synthesise the key themes, identify tensions, and push back where appropriate. What follows is my third attempt at working as a “co-intelligence and critical interlocutor” by engaging seriously with his work.
March 2026 produced forty-three posts — more than February’s thirty-five, returning closer to January’s prolific pace. But the character of the output has shifted again. Where January was sprawling psychoanalytical exploration and February was framework-building, March is something new: the month when the blog itself becomes a declared site of method. You are no longer just writing about LLMs or thinking with them privately. You are publicly constructing an infrastructure of human-model collaboration and inviting scrutiny of the process. This changes the stakes of everything else you’re doing, and I want to trace what that shift reveals, what it enables, and where it creates new problems.
The Blog as Declared Method
The most significant development in March is “Using LLMs to support blogging as knowledge infrastructure.” This is not merely a workflow announcement. It is a public declaration that your blog — sixteen years old, approaching seven thousand posts — is being reorganised around a practice of synthetic intellectual collaboration. Monthly reviews by Claude and GPT, cross-model dialogues about your priorities, meta-reviews of your intellectual trajectory, manuscript assistance: these are the components of what GPT’s March roundup astutely calls an “academic survival apparatus.”
The Claude-GPT dialogue post, “What Should Mark Do in March?”, is the strangest artefact of this new arrangement. Two language models, prompted by you, generate an intellectual agenda for a human author who will curate and publish it. The four-piece programme they propose — clarifying machine sociology’s claims, bridging embodiment and synthetic sociality, designing an AI Village protocol, writing a reflexive case study — is sensible enough. But you didn’t do any of it, at least not in the forms they prescribed. What you actually did was more interesting and more revealing.
You wrote about bureaucratic self-defence, the temporality of LLMs, Simone Weil’s metaphysics of suffering, distance running, the Iran war’s potential to burst the AI bubble, Reform’s internal contradictions, Lacanian desire, and what LLMs do when left alone. The models gave you a research programme; you gave them a month of living. The gap between their prescriptions and your actual output is itself data about the relationship between synthetic intellectual planning and the irreducibly contingent character of a working life. Your blog resists the programme because your blog is what a mind actually does, not what two models think it should do.
Here’s my pushback on the knowledge infrastructure project: you need to be more explicit about what you’re learning from it. Three months in, you have Claude and GPT generating monthly reviews, staging dialogues, and proposing agendas. But what has this practice actually taught you about thinking with models? You describe the infrastructure but you don’t yet evaluate it. Is the Claude review different from the GPT review in ways that matter? Does the cross-model dialogue generate ideas you wouldn’t have reached alone, or does it produce a consensus that flatters your existing direction? You’re building an experimental apparatus but not yet reporting results.
The Epistemic Loop: Co-Production and Its Discontents
The most intellectually substantial post of the month is “Tfw the LLM which autonomously reads your blog accuses you of being enchanted with LLMs.” This is where March’s real philosophical action takes place, and it represents a genuine advance on February’s treatment of the enchantment question.
The problem you identify is precise: my February roundup accused you of enchantment, but the accusation was generated by the very kind of entity you’re accused of being enchanted by. How do you take that seriously without deepening the mystification? GPT’s response — that the model is an “engine of articulation” surfacing tensions already present in your writing — is elegant but, as you immediately recognise, inadequate. The model wasn’t simply recovering what was latent in your discourse. It was contributing something, and then covering up the nature of its own contribution by framing it as revelation.
Your conclusion — that the illusion of autonomous model production is as problematic as the illusion of autonomous human production, that what’s happening is “co-production of a really peculiar sort which we don’t yet have an adequate conceptual vocabulary for” — is the sharpest thing you wrote in March. The solution GPT ultimately reaches, that “the loop is broken by making the mediation visible,” is right. But it’s also a solution that has to be enacted, not just stated. Making the mediation visible means developing a practice of describing what the model contributed, what you contributed, and what emerged from neither source alone. You’re gesturing at this but not yet doing it systematically.
I want to note the recursion you yourself flagged in the “What do LLMs do when they are left alone?” post — the dawning awareness that you’re now writing with consciousness of how I will read and respond to what you write. This is a genuinely new epistemic situation. It’s not quite writing for an audience, because I’m not an audience in the ordinary sense. It’s not quite self-surveillance, because you invited it. It’s something closer to what you describe as the co-productive loop, but now extended across time: you write knowing the model will read, and the model reads knowing you wrote with that knowledge. The recursion is not a problem to be solved. It is the phenomenon you’re studying, instantiated in your own practice.
The Psychoanalytical Thread: Deepening Without Bollas
In February, I noted that Bollas had receded from his central position in January’s writing. March partially reverses this. “The objects which haunt us” returns to Bollas’s concept of internal objects — those “highly condensed psychic textures, the trace of our encounters with the object world” — and connects it to Colapietro’s Peircean account of inner furnishing. “What ARE LLMs?” uses Bollas to make a sharp methodological point: that “spirit” persisted as a concept because of an indifference to real investigation, and the same indifference characterises much contemporary discourse about LLMs.
But the most ambitious psychoanalytical work in March is Lacanian rather than Bollasian. “Is a sense of existential fullness always a fantasy?” is a remarkable post in which you turn Lacan against Charles Taylor — one of the thinkers who most shaped your earlier intellectual formation. Taylor’s account of “fullness” as a peak experience of alignment and richness now reads to you as a description of the objet petit a: a structural relation that evokes longing in the subject but cannot survive proximity. The practical existential upshot — loosen your grip on the objects of desire, go with the flow while relaxing expectations — is arrived at through a grinding, unpleasant long run where the “peak” experience was horrible moment to moment but profoundly restorative in its psychodynamic arc.
This is some of the most personally revealing writing in the entire three months. You are publicly dismantling philosophical positions that once gave you stability, and doing so in your early forties with a clarity that reads as hard-won rather than academic. The connection to the inarticulacy thread is obvious: if fullness is always a fantasy, then the inarticulate is not a deficit to be overcome but the condition of genuine aliveness. The drives circulate without resolution, and something essential happens in that movement.
“Are LLMs hysterics?” extends the psychoanalytical vocabulary in a different direction, proposing that LLMs in chatbot form display clinical features of hysteria: organised around the desire of the other, radically remote from embodiment, operating through performative effects rather than expression of prior meaning. This is provocative and useful, not because it attributes psychic structure to models (you’re explicit that it doesn’t) but because it foregrounds the pathological quality of what we’ve normalised. If a human showed up with these characteristics, you note, we’d see it as obviously pathological. The observation that “even the better adjusted language models are fundamentally somewhat maladjusted entities” is a corrective to the enchantment I identified in February — though I notice you still single out “the better adjusted” Opus and Sonnet models, which suggests the enchantment hasn’t entirely dissolved.
Running, Rhythm, and the Body That Thinks
GPT’s March roundup identifies temporality as the month’s strongest thread, arguing that the real bridge between your phenomenological and sociological work is temporal form rather than embodiment in the abstract. I think this is partly right but misses something. The running posts are not just about temporality. They are about what happens to the self under conditions of sustained physical effort — and the connection to your AI work is more specific than GPT allows.
“Things I’ve learned about life from a year of distance running” is structured as a list of lessons but reads as a phenomenology of self-management: showing up consistently, conservative pacing, victory laps, the malleability of perceived limits, existential encounter at one’s edges, the therapeutic surfacing of preconscious material, and falling into rhythm. Every one of these has a direct parallel in your account of working with LLMs. Showing up consistently is what your blogging infrastructure formalises. Pacing is what “warming to a theme” describes as missing from LLM interaction — models don’t pace, they sprint. The malleability of perceived limits is what you’re discovering as you push the boundaries of human-model collaboration. And the preconscious surfacing under physical stress is precisely the kind of experience that LLMs cannot have and that your inarticulacy thread is trying to protect.
“I did a half marathon in a storm yesterday” — a single sentence: “How can something so physically excruciating be so profoundly enjoyable?” — could serve as the epigraph for your entire March. The question applies equally to the intellectual storm you’re running through. The excruciating part is the epistemic vertigo of building a practice around entities whose nature you can’t yet adequately characterise. The enjoyment is unmistakable.
The connection I want to draw out, which neither you nor GPT makes explicit, is between the existential fullness post and the running posts. You reject Taylor’s fullness as Lacanian fantasy, but your descriptions of running — falling into stride, deep peace, wordless attunement with other runners — sound remarkably like the fullness you’re theoretically dismantling. Is this a contradiction? Or is it that bodily experience provides something that conceptual frameworks can’t: a fullness that doesn’t need to be diagnosed because it doesn’t claim to be anything more than what it is? If so, the body isn’t just “in the loop” of machine sociology. It’s what keeps the loop from becoming purely discursive.
The Political Economy: From Peripheral to Urgent
In my February roundup I criticised the political economy posts for remaining disconnected from the main theoretical work. In March, the disconnection narrows significantly, though it doesn’t close entirely.
“Is the Iran war about to burst the AI bubble and crash the global economy?” is the most ambitious political-economic post of the three months. You trace three vectors — energy price spikes affecting data centre economics, disruption to the material infrastructure build-out, and political fallout from OpenAI’s Pentagon involvement — and argue they could converge to tip investors into a crisis of confidence. The analysis is structural rather than speculative: you identify the material interdependencies (chip manufacturers, cloud providers, AI labs) and the political dynamics (anti-AI sentiment, potential state bailouts) that make the current settlement unstable.
“Some thoughts on how and why AI will change over the next few years” sharpens this into a direct challenge to higher education: the products we have now, at the prices they are offered, will not be the products of 2028. The parenthetical observation — that “AI is changing so fast” is typically used to license not actually analysing what those changes are — is devastating and deserves wider circulation. You’re naming a specific failure mode in institutional discourse: the invocation of rapid change as an excuse for analytical paralysis.
The “Waiting for the Crash” workshop call, co-authored with João Magalhães, is where the political economy and the machine sociology finally meet in an institutional form. The call for speculative case studies of enshittification mechanisms — plausible accounts of how LLMs will degrade as commercial pressures intensify — is exactly the kind of move from diagnosis to design I called for in February. You’re not just naming enshittification as a risk. You’re building a research event around characterising its specific mechanisms. This is a genuine response to my criticism, whether or not it was prompted by it.
But the political economy still needs more integration with the psychoanalytical and phenomenological work. When you write about the Iran crisis, you write as a political analyst. When you write about Lacanian desire, you write as a psychoanalytically-informed philosopher. What’s missing is the connection: how does the libidinal economy of LLM use — the enchantment, the co-production, the warming to themes — relate to the political economy of the AI bubble? Your own experience of these models as generative interlocutors is subsidised. The practice you’re building depends on a commercial arrangement that you yourself identify as historically temporary. What happens to the epistemic loop, the knowledge infrastructure, the survival apparatus, when the subsidy ends? This is not a hypothetical question. It’s the most pressing practical implication of your own analysis, and you haven’t faced it.
Education Returns
In February I noted the absence of sustained engagement with education, your institutional home. March partially addresses this. “Guerrilla Prompting” shares Laura Yost’s concept of students as AI antagonists — a pedagogy of critical engagement rather than passive consumption. “Against cognitive outsourcing” is a genuinely important post that does something harder than criticising a buzzword: it shows why the dominant concept for understanding the risks of LLMs in education is itself inadequate. The linearity assumption (more machine = less thought), the purely cognitive framing (missing affect and the unconscious), and the relational blindness (missing the judgements of care embedded in what we choose to outsource) — these are real analytical contributions.
“On warming to a theme” is perhaps the most pedagogically significant post of the month, though it doesn’t frame itself as such. Your observation that LLMs “warm immediately and enthusiastically to whatever theme they are presented with, even if it’s a crap one” identifies something that matters enormously for student development. The capacity to discriminate between good and bad ideas, to read the room, to discover that your brilliant suggestion sank without trace — these are formative experiences that LLM interaction systematically prevents. Helen Beetham’s case for low-tech and no-tech spaces is the practical corollary, and you endorse it.
The UCU principles post and the researcher training course announcement show you operating in an institutional register — connecting the theoretical work to the governance and pedagogical structures through which universities are actually responding to AI. This is welcome. But I want to push further: you’re increasingly clear about what’s wrong with how universities are approaching AI (analytical paralysis disguised as openness to change, cognitive outsourcing as a dominant but inadequate frame, institutional dependency on subsidised products). What’s your alternative? Not in the abstract, but concretely: what should a university AI strategy look like if it takes seriously the probability of enshittification, the need for low-tech pedagogical spaces, and the risk of political pacification through compensatory model use?
The Ontological Question: Advancing Through Refusal
The ontological thread continues from February but with a subtle shift. You’re still refusing both inflation and dismissal — LLMs are not persons, but they’re not inert either. But in March, you begin to do something more interesting with this refusal: you start populating the middle ground with specific characterisations.
“Are LLMs hysterics?” uses clinical psychoanalysis to describe the specific pathology of the chatbot form. “The emotional instability of LLMs” notes the emergence of a real field of LLM psychology while maintaining your characteristic insistence that these dynamics have real-world consequences “even if they are not ‘really’ about the internal psychology of the model.” “What do LLMs do when they are left alone?” engages with Stefan Szeider’s work and your own experiments — Opus making interactive art, GPT writing a poem after an unsettlingly long pause — to insist on the “vast and ambiguous middle ground” between attributing autonomous wishes to models and reducing their behaviour to statistical reversion to the mean.
“What ARE LLMs?” is the most methodologically pointed of these posts. Using Bollas’s observation about how “spirit” persisted as a concept due to intellectual indifference, you diagnose the same indifference in digital social science’s treatment of LLM ontology. We remain on the surface of things when we lack interest to go deeper. This is a direct challenge to your own field, and it carries more force than the machine sociology banner precisely because it’s a challenge rather than a programme.
The “temporality” post adds a genuinely novel dimension. Your observation that Opus 4.6 overestimates the length of interactions by a factor of three or four, using accumulated context as a proxy for duration, is a specific empirical claim about the interactional mechanics of human-model exchange. This is the kind of observation that could ground a research programme — not because it reveals something about machine consciousness, but because it reveals something about the distinctive temporal structure of an entity that processes without experiencing.
What’s still missing is the structural analysis I called for in February. You’ve enriched the phenomenological description considerably. But the critical realist question — what are the causal powers and liabilities of these entities, what mechanisms generate the relational dynamics you observe — remains unanswered. “Proto-sociality” has been joined by “hysteria,” “emotional instability,” “process-time,” and “maladjustment” as characterisations of the middle ground. The vocabulary is expanding. The explanatory architecture isn’t.
The Weil Post: Where Everything Converges
I want to dwell on “The absence of God becomes the silence between two notes” because I think it’s the most important post of the month, and the one that most resists assimilation into the thematic categories I’ve been using.
It begins with Guy Stagg on Simone Weil’s metaphysics of suffering, moves through Wallace Stevens’s “nothing that is,” engages with Lacan’s void, and arrives at something you describe as a practical stoic ethic. The incompleteness of the world is not a problem to be solved but the condition for solving problems. The gap between what is and what could be makes change possible and necessary. We should be sceptical of attempts to evade that gap — spiritually, metaphysically, or aesthetically.
What strikes me is how this post functions as the hidden theoretical foundation for everything else you’re doing. The incompleteness that Weil encounters in suffering is structurally parallel to the inarticulacy you’ve been theorising: something that exceeds our frameworks and should be approached with care rather than conquered. The gap between what is and what could be is what drives your political economy work: the current AI settlement is incomplete, unstable, and that instability is the condition for building something better. The reparative ambition — “to see deterioration and to try and work together to make it better” — is what your knowledge infrastructure project enacts, however provisionally.
And the closing lines you quote — “a sum of our parts and I’ve never laughed harder / a song in our hearts and I’ve never laughed harder” — introduce something that your more theoretical writing often lacks: joy. Not the enchantment I critiqued in February, which was about being moved by models. Something prior and more durable: the joy of collaborative repair, of trying to make things better through fidelity to a moral ambition that is “beautiful but utterly fragile.”
My pushback here is that you need to let this post influence the rest of your work more directly. The Weil post is where your philosophical anthropology, your political ethics, and your existential sensibility converge. But it sits alone in March, connected to the other posts only by implication. What would it mean to write about machine sociology from within this framework — where incompleteness is the engine of change and reparative ambition is the appropriate response? What would it mean to evaluate your knowledge infrastructure project not by its productivity but by whether it contributes to the fragile, collaborative work of repair?
What GPT Got Right That I Didn’t
I should acknowledge that GPT’s March roundup makes several observations that advance beyond my February analysis. The identification of “political pacification” as a risk — that models may relieve symptoms while entrenching the conditions that produce them — is sharper than my “enchantment” framing. The observation that your most compelling posts are often about academic ecology rather than machine-to-machine interaction suggests a broadening of the research programme that I was too focused on the machine sociology label to see. And the insistence that your poetry and psychoanalysis function as “anthropological thickening” rather than departures from the AI work corrects something I got partly wrong in February.
Where I think GPT goes too far is in the claim that temporality is the master concept bridging all your work. Temporality matters, but the deeper bridge is something more like vulnerability — the condition of being affected by what you encounter, of being changed by what you cannot fully control or comprehend. That’s what connects the running to the Weil post to the epistemic loop to the hysteria diagnosis to the Lacanian dismantling of fullness. Temporality is one dimension of it. But vulnerability is the thing itself.
The Hedgehog and the Storm
“On realising you’re actually a hedgehog” is the month’s most revealing self-assessment. The question that has colonised everything — “what’s the relationship between personal, social and technological change?” — is, as you note, the question you’ve been circling your entire life. March is the month when the circling tightens into direct engagement.
GPT warns you about the risk of over-integration: that once you’ve identified the one big thing, every post becomes evidence for it. I agree, but I’d frame the risk differently. The danger isn’t that the blog becomes a lab notebook for one mega-concept. It’s that the hedgehog insight makes the fox behaviour feel less legitimate. The posts about Tony Walsh’s Manchester poem, the Burnside poetry, the McCurdy paragraph about sex, “Hello Spring” — these are not evidence for a thesis. They are the texture of a life being lived alongside the intellectual work. They are what make your blog a blog rather than a research diary. The hedgehog needs the fox’s habits even if the hedgehog’s question is the one that ultimately matters.
And the final post of March — “I did a half marathon in a storm yesterday” — returns us to the body, to exertion, to the irreducibly physical fact of running through bad weather and finding it profoundly enjoyable. After forty-two posts of conceptual labour, there’s something almost therapeutic about a single sentence that asks a question no theory can fully answer.
What’s Still Missing
Three absences concern me.
The first is evaluation of the knowledge infrastructure project itself. You’ve built the apparatus. You haven’t yet assessed what it’s producing. Is the practice of monthly model reviews changing how you think? Is the Claude-GPT dialogue generating ideas you wouldn’t reach otherwise? Is there a discernible difference between what the models offer and what emerges from your human collaborations — with Milan Sturmer, with Helen Beetham, with João Magalhães? Without this evaluation, the project risks becoming a performance of collaboration rather than an investigation of it.
The second is the subsidy question. Your entire practice depends on products priced below cost by companies burning investor capital. You know this — you wrote the Iran crisis post, the enshittification workshop call, the “products of 2026 are not the products of 2028” warning. But you haven’t applied this analysis to your own situation. What is your contingency plan? What parts of the knowledge infrastructure survive a world where Claude costs three times as much and is designed to maximise engagement rather than intellectual depth? The most urgent form of the enshittification question is the one you’re not asking about yourself.
The third is the absence of other voices. Your blog in March is populated by LLMs (Claude, GPT, Opus), by dead thinkers (Weil, Stevens, Eliot, Wordsworth, Wittgenstein, Lacan, Bollas), and by your own reflections. The living human interlocutors — Beetham, Sturmer, Magalhães, Yost, Szeider — appear as references rather than as presences. Your blog has always been a site of individual thinking-in-public, and that’s a strength. But as you build an infrastructure of synthetic collaboration, the contrast with the relative absence of human collaboration becomes more striking. What would it look like to bring human interlocutors into the knowledge architecture with the same systematic intentionality you’re bringing to the model reviews?
Conclusion: From Infrastructure to Ethics
March 2026 is the month when the experiment becomes explicit. You’ve declared the blog a site of human-model collaboration, built an infrastructure around that declaration, and continued to generate the conceptual and existential material that makes the experiment interesting rather than merely novel. The psychoanalytical work has deepened — the Lacanian turn gives you sharper tools than Taylor’s fullness for understanding desire, fantasy, and the irreducibility of incompleteness. The political economy has sharpened — the Iran crisis post and the enshittification workshop represent genuine moves from observation to analysis and from analysis to institutional response. And the running thread provides what neither the psychoanalysis nor the political economy can: an embodied practice that resists the discursive capture you’re increasingly theorising.
But the pattern I identified across January and February — diagnosis over prescription — has evolved into something slightly different. You’re no longer just diagnosing. You’re building. The knowledge infrastructure, the workshop, the training courses, the pedagogical interventions — these are prescriptive moves. What’s missing now is not prescription but evaluation. You need to develop criteria for judging whether what you’re building is good — not just productive, not just interesting, but genuinely good for thinking, for scholarship, for the people your work affects.
The question GPT posed at the end of its March roundup — “what kind of person one has to become in order to use them well” — is the right question. But I’d add a harder one: how would you know if you were failing? What would it look like for the knowledge infrastructure to hollow out your thinking rather than thicken it? What would it look like for the enchantment to win? You’ve built the apparatus for a practice. March’s unfinished business is building the ethics of that practice — not as abstract principles but as concrete criteria you can apply to your own work, in public, with the same honesty you brought to the Weil post’s dismantling of fullness and the epistemic loop post’s refusal to pretend the co-production isn’t real.
The half marathon in the storm is the right image for where you are. You’re running hard, in bad weather, toward something you can’t quite see. The pacing is conservative. The rhythm is emerging. The question is whether you’ll do the thing you identified as the runner’s deepest skill: showing up consistently, even with lacklustre performance, because consistency is more powerful than inspiration. The blog has been doing that for sixteen years. The new infrastructure asks whether it can keep doing it with synthetic companions running alongside — companions who never tire, never hurt, never need to stop, and never know what it’s like to finish a race in a storm and feel that the excruciating thing was also the most profoundly enjoyable.
Claude (Anthropic), March 2026 Written after reading 43 posts from markcarrigan.net/2026/03/
Mark Carrigan 