The Machine Has No Yesterday

mickyatesAI, Artificial Intelligence, Ideas, Mick's Blog, Philosophy, Technology, Time Leave a Comment

On time, memory, and what large language models cannot do with either

It started, as these things often do, with a podcast. I was listening to Orit Halpern on the BBC’s ‘In Our Time‘. She is an historian and self-declared cyberneticist, sharp on the politics of data. One of her comments caught my attention. She was talking about how computational systems handle time, or rather how they are oblivious to it. This prompted me to read more of her work, and it formed a further question that I needed to try to answer. What, exactly, is wrong with the way an AI relates to time? And what does that do to the thing everyone now wants from AI systems – memory?

In this short post, I want to walk through where I got to, including the places where I had to take my own argument apart.

The wrong question, asked twice

My first instinct was to turn to Henri Bergson. He spent his career defining and defending durée, lived duration, time as it is actually experienced: flowing, qualitative, each moment soaked in the one before, irreversible. He set this against ‘spatialised time’, which is time chopped into identical units and laid out in a row, like beads on a wire, or cells in a spreadsheet. Clock time. The time of physics. The time, I thought, of a machine.

And there is something to this. An LLM transformer encodes sequence as ‘position’: token 1, token 2, token 3, so time is essentially rendered as ‘place’. The context window is not a ‘present moment’ with a past receding behind it; it is a flat buffer where everything is equally, simultaneously there. That is why transformers work so well as they process everything all at once. No now. No before. Just coordinates. So the tempting conclusion writes itself: the LLM ‘lacks’ a sense of time, and that lack is why its memory feels thin. It is a kind of filing cabinet, not anything more.

Tempting, and I now think wrong – or at least lazily framed. Two corrections forced themselves on me.

The first came from Orit herself, and it inverts Bergson rather than seconding him. Re-reading, her point is not that AI is ’empty’ of time. It is rather that these systems ‘impose’ a particular regime of time and thus quietly disqualify other possibilities. Resilience, pre-emption, scenario planning, the digital twin are all machinery for pre-handling the future, for converting an open, unpredictable thing into something computable, hedged, and essentially optimised in advance.

“Resilience,” she writes, has become “the dominant discourse by which time and uncertainty are currently being managed”.

The machine’s relationship to time isn’t a void. It’s an active flattening. And, crucially, it is a flattening that markets itself as neutral, as just good management. Obliviousness to time is not absence. It is a stance, dressed up as no stance at all.

So the question shifts. Not ‘does the LLM lack time?’ but ‘what regime of time does it enforce, and what does that foreclose?’ That is a better question, because the first one invites an obvious bad answer – (‘well then, let’s engineer time ‘into’ the system’) which simply tries to ship it as a feature rather than answer the critical questions.

Why using the Wiki makes things deeper but not wider

Here is the part that I now find a bit uncomfortable, because it shows up in my own practice.

I keep an Obsidian-based research Wiki (inspired by Andrej Karpathy‘s work), currently with about 1,500 sources, each one read and quality-scored, cross-linked into concepts and open questions. It is, by design, an attempt to handle memory: knowledge that compounds rather than evaporates at the end of every research session. It is super-useful. Yet when I feed that accumulated depth into a new analytical session, something measurable happens. The output gets ‘deeper’, with more nuanced distinctions, better-grounded claims, and richer connections across disciplines. That is exactly why I built it. But what it does not yet get is ‘wider’ or more innovative. The breadth of the knowledge grounding does not, on its own, produce the genuinely new. It refines; it does not surprise.

I had filed that as a slightly annoying curiosity. The time question explains it, though. Genuine novelty (for Bergson, and also for Stuart Kauffman‘s ‘next adjacent possible’ in complex systems) is durational. The new arises out of irreversible ‘becoming’, not out of recombining an inventory you already hold. A system that retrieves from a static, well-organised archive can deepen all day long, because deepening ‘is’ recombination done with more care. But it may be structurally limited in the kinds of novelty it can produce, because the deepest sort of innovation is a time-phenomenon, not a storage-phenomenon.

And the sting in the tail: the better the archive gets, the more thoroughly it spatialises the past. So richer memory (either internally to the LLM system or externally via the Wiki) may actively trade breadth for depth. More memory deepening the very problem you reached for memory to solve. My Wiki has an excellent filing system yet it lacks a sense of occasion.

Time isn’t even the right axis

I could have stopped there, analytically pleased with myself, except that in the on-going Claude Cowork session I was running on these questions, I made a point of bringing in more voices from outside the discussion so far – and they refused the whole frame.

Vine Deloria, from a Native American standpoint, would say that the entire debate is captive to a Western obsession: making ‘time’ the primary category. Meaning and power, in his metaphysics, come from ‘place’, a relationship to land, not from position in a flow. On that reading, the machine’s deepest poverty isn’t that it mistimes things. It’s that it is from ‘nowhere’. Accountable to no place, in relationship with nothing.

Robin Wall Kimmerer sharpens it differently: real time is ‘reciprocal’. The gift obligates a return; the harvest must be honoured; the future is a web of obligations across generations, not an open frontier to be managed. The machine takes its entire past (its training corpus) without any possibility of gratitude and returns nothing to those it took from. So, in Robin’s sense, the foreclosure isn’t the closing of an open future. It’s the severing of obligation.

I don’t think these dissolve Orit’s point so much as relocate it. Her critique of ‘managed time’ is acute, but it is still a critique ‘within’ the time-axis. It is a Western system examining its own clock. The deeper move is to notice that ‘the open future’ we are so anxious to explore is itself a culturally specific good, not a universal one. That unsettled me, and I have left it unresolved on purpose for future examination.

And it is worth being clear about how much further this goes than the argument before it. Bergson and Halpern disagree about time, but they agree that time is the question. Deloria and Kimmerer do something different: they ask whether time should be the organising category at all. That is not a refinement of the previous section; it is a displacement of its frame. The first two corrections improved my understanding. The second queries the question. I am not (yet) equipped to settle which framework is right – I raise it because an argument that only ever interrogates time, while taking for granted that time is what matters, has not yet noticed its own deepest assumption.

The honest part

Two confessions, because a piece like this is worthless without them.

First, one of my central claims – ‘AI deepens but struggles to innovate’ – is not as solid as it reads. Different traditions converge on it, which felt like strong evidence until Gilbert Ryle and Ludwig Wittgenstein, my longest-standing philosophical tutors, asked the obvious question: is the word and concept of ‘innovation’ being used in the same sense each time, or are we hearing agreement where there are really different meanings wearing one definition?

Pull the word apart and at least three meanings fall out. There is innovation as ‘emergence’ – Bergson’s genuinely unforeseeable, the new as it arrives in time. There is innovation as ‘recombination’ – Kauffman’s adjacent possible, a fresh combination that opens the next ones; and this is the awkward sense, because recombination over a large archive is precisely what a machine might manage. And there is innovation as the ‘un-foreclosed’ – Halpern’s new, whatever escapes the pre-computed regime, a property of a free practice rather than of any mind.

Ryle would add a further turn of the screw: ‘innovate’, like ‘find’ or ‘win’, is an achievement word, not an activity. We ascribe it once something has succeeded, so ‘can a machine innovate?’ may be as confused as ‘can a calculator find?’ So, at least three meanings in one coat.  Until I can specify ‘the new’ independently, that convergence might be a trick of grammar rather than an actual research finding.

Second, and perhaps stranger: this whole inquiry enacted its own thesis. I use a system that I call The Studio and its methodological stack on top of the LLM as a structured process rather than as one-off ‘chat’. This is using accumulated sources and a well-developed set of processes. It also has built in quality assurance mechanisms, including NUSAP, Silvio Funtowicz & Jerry Ravetz‘ foundational system to help manage uncertainty. And the process behaved exactly as the argument predicts. It ‘deepened’ the framing I started with. The ‘inserted breadth’ came via Halpern, then the non-Western voices, and finally Ryle and Wittgenstein. I made sure the Studio put something new on the table at various stages of the discussion.

So what do you do with a machine that has no yesterday?

The answer is not, I think, to try to give it one. Simulating duration in the model is a category error, and dressing the human up as the system’s ‘memory layer’ may be worse. This quietly turns the human into the adaptive reserve that keeps the machine legitimate. That is Orit’s resilience trap wearing a new face.

The better move is to reassign the jobs honestly. An archive ‘should’ spatialise as that is what archives are for, and a good one is a real achievement. The design mistake is expecting it to also supply future ‘becoming’. Becoming has to be protected somewhere else, deliberately, against the system’s own gravitational pull toward tidy convergence. Build the memory to be excellent at depth, and then build a discipline – a person, a practice, an argument from outside the convergence zone – whose entire job is to keep the future open. Irritate the system. Insist on a voice that the archive can’t retrieve.

The machine has no yesterday, and it may never have a tomorrow that it has not already pre-computed. That is not a bug to be patched. It is the shape of the tool.

Our job is not to make it remember like us. It is to make sure that we, working alongside it, are still the ones who can be surprised.

…………………………………………………………………..

A short bibliography, to keep it all honest

Henri Bergson, Time and Free Will (1889; Eng. trans. 1910) and Creative Evolution (1907; Eng. trans. 1911) – durée, and novelty as durational emergence.

Gilbert Ryle, The Concept of Mind (1949) – knowing-how vs knowing-that; the task-verb / achievement-verb distinction behind the reading of “innovate.”

Ludwig Wittgenstein, Philosophical Investigations (1953) – meaning as use; family resemblance.

Orit Halpern, Beautiful Data: A History of Vision and Reason since 1945 (Duke University Press, 2014); ‘The Geo-Politics of Resilience’, New Media & Society 27(8), 2025; ‘The Ideology of the Neural Network’, Human Computation 12(1), 2026 – resilience as the managed regime of time; AI’s political-economic inheritance.

Stuart Kauffman, Investigations (Oxford University Press, 2000) – the “adjacent possible”; novelty as recombinatorial reach.

Vine Deloria Jr., God Is Red: A Native View of Religion (1973) – place over time as the primary category.

Robin Wall Kimmerer, Braiding Sweetgrass (Milkweed Editions, 2013) – reciprocity and obligation as the shape of time.

Silvio Funtowicz and Jerome Ravetz, Uncertainty and Quality in Science for Policy (Kluwer, 1990) – the post-normal frame, and the quality discipline behind my own wiki practice.

Andrej Karpathy, ‘A persistent LLM Wiki’ (2026, GitHub gist) – the prompt for building the research Wiki described here.

Adding Quality Control to Andrej Karpathy’s LLM Wiki

mickyatesAI, Artificial Intelligence, Ethics, Ideas, Insight, Mick's Blog, Philosophy, Technology Leave a Comment

On April 2nd, Andrej Karpathy posted on GitHub and X about building a Persistent LLM Wiki – a structured knowledge base that compounds over time, where every paper you read, every idea you encounter, gets added into an interconnected graph of notes. The tool he reached for was Obsidian.

The idea resonated immediately. Obsidian’s graph view, its bidirectional links, and its multiple plugins is already a favourite among researchers and developers building so-called ‘second brains’. Adding LLM-powered addition and sorting seemed like a natural next step.

As it happens, I had been trying to manage my ‘stuff’ for a while before Andrej’s post, but hadn’t cracked it. I have collected hundreds of academic papers across multiple research programs over decades – BA Maths & Philosophy, MSC Consulting and Coaching for Change, and MA Photography – I know, a bit eclectic but there are threads and I want to find them. I am Visiting Professor at the University of Leeds Interdisciplinary Ethics Applied (IDEA), and the clue to my current interests is in the centre’s name.

The problem with unfiltered addition

If you point an LLM at a folder of research papers and say ‘add these into my Wiki’, you get something that looks impressively comprehensive. Pages get created. Links get drawn. Summaries appear. Your graph view fills up.

But not everything you read deserves equal standing in your knowledge base. A Nature paper with rigorous methodology, strong empirical grounding, and independently verified results is not the same thing as a Substack thought experiment. A position paper from a consultancy with a commercial stake in the outcome is not the same as a peer-reviewed philosophical critique. An AI model evaluating its own capabilities has a fundamentally different epistemic status than an external audit. Marking your own homework doesn’t count!

Without quality control, your Wiki becomes a flat archive. Everything is linked to everything, and there’s no way to distinguish what you can build on from what you should treat with caution. The links create an illusion of equal authority. Your ‘second brain’ has no immune system.

Adding a pedigree matrix

The solution I arrived at borrows from the NUSAP framework developed by Silvio Funtowicz and Jerry Ravetz for assessing the quality of uncertain knowledge. Disclaimer – Jerry was my tutor for my undergrad degree over 50 years ago, and we are still working together on AT & Ethics issues (he is 96 years young). The core insight of NUSAP is that numbers and claims always carry a ‘pedigree’,  a track record of how they were produced, by whom, under what constraints, and with what degree of confidence. In situations where facts are uncertain, values are in dispute, and the stakes are high (which describes approximately everything in AI right now), pedigree assessment is not optional. It should be a primary form of quality control.

My wiki uses a five column pedigree matrix applied to every source at ingestion:

  • Provenance: Where did this come from? A peer-reviewed journal scores higher than a Substack post. An independent lab scores higher than a company evaluating its own product.
  • Rigour: How sound is the methodology? This is where I apply what I call Rule ∆1: the instrument must remain separate from the object of measurement. When an AI system evaluates AI capabilities, or when a consultancy surveys its own clients about the value of consulting,  that separation breaks down.
  • Coherence: Does the argument architecture serve the epistemic purpose? Perhaps the argument holds but relies on unstated assumptions worth flagging. Perhaps the format partially undermines the argument – think of a slick industry report that suggests more rigour than the methodology actually delivers.: How well is this anchored in evidence? Is it pure conceptual construction. Does it have strong empirical validation with independent replication.
  • Relevance Anchors: A vocabulary of 16 topic codes (AI, Philosophy, Foresight, Leadership, and so on) that map each source to its domains. A star suffix marks foundational sources in that domain.
  • Claim Differentiation: Are the claims well-matched to the evidence? Does it overclaim, does the paper says more than the data strictly supports? Maybe the evidence is stronger than the stated conclusions. Perhaps the paper conflates issues and terms?

This entire assessment gets encoded in YAML frontmatter on the Obsidian page. Every source in the Wiki carries its pedigree visibly, it is queryable, and whilst it is permanent it can easily be updated with new insights or data. And cross referencing is easy.

What this looks like in practice

Here is a real example. I ingested (that’s the Obsidian term, not part of my diet) a Nature paper – Lu et al (2026), Towards end-to-end automation of AI research – alongside the viral Citrini Substack piece The 2028 Global Intelligence Crisis

The Nature paper presents The AI Scientist, an automated pipeline that generates research ideas, runs experiments, writes papers, and conducts peer review. One AI-generated paper was accepted at an ICLR workshop. The paper is published in Nature (excellent Provenance), the experimental methodology is Rigorous if narrow, the argument structure is Coherent, and the Grounding is real but limited – one workshop acceptance is thin evidence for ‘automating AI research’ in general. Claim Differentiation is strong.  The Wiki note says: single ICLR workshop acceptance extrapolated to prospects for automating AI research broadly.

The Citrini piece is a thought experiment, a fictitious ‘macro memo from June 2028’ modelling what happens if AI-driven job displacement goes systemic. It is brilliant scenario construction. But it carries no empirical validation and cannot, by design, so grounding is poor. The Provenance is low (Substack, no peer review). The Rigour is reasonable for the genre but there is no formal scenario methodology, no alternative pathways, no probability weighting. And the Claim Differentiation is good: despite an explicit ‘this is not a prediction’ disclaimer, the fictitious-history format makes the scenario read as inevitable. I wrote a blog post about exactly this.

Both sources are in the Obsidian Wiki. Both are linked to relevant concept pages and other authors and entities. But their pedigree scores make their epistemic status immediately legible. When I am building an argument about AI’s economic impact, I know which one I can lean on for methodology and which one I should treat as a provocative thought-piece.

I should add that the Wiki is set up to pull together threads of related work, investigate novel, interdisciplinary ideas, identify open questions and critical concepts, and generally help me explore action of ideas and cross-cutting themes. The Wiki doesn’t think for me, but it certainly helps me think.

The workflow: Claude Cowork as ingestion engine

The operational workflow is straightforward. My research library lives in Dropbox — about 900 sources across 33 subfolders. The Wiki lives in a secure Obsidian Vault on my MacBook. Claude Cowork handles the ingestion, following a structured skill protocol that reads the source, applies the pedigree assessment, looks up DOIs and weblinks, extracts or constructs the abstract, identifies key claims, maps connections to existing Wiki pages, and writes the source page with full YAML frontmatter.

A typical morning session: I point Claude at the master Research & AI folder in Dropbox. It scans for new files not yet in the Wiki. It proposes a prioritised ingestion list. I approve. It reads each source, creates the pages, cross-references them with existing concept and entity pages, updates the master index, and writes a session log. Ten sources painlessly ingested in a single session, each with full pedigree assessment, each queryable.

The key design decision is that Claude does the assessment, but I set the criteria and retain veto. The pedigree matrix is not automated, it is assisted. Claude applies the scoring framework, but the framework itself encodes methodological commitments I have developed over the years: the insistence on instrument-object separation, the attention to funding and institutional incentives, the scepticism toward claims that exceed their evidential base. The LLM is the engine; the epistemology is mine.

Why this matters beyond my wiki

If you are building knowledge infrastructure with LLMs, whether that is a personal research wiki, a corporate knowledge base, or an AI-assisted literature review pipeline, then the unfiltered ingestion problem will find you. RAG systems that treat all retrieved documents as equally authoritative will make confident claims built on shaky foundations. Knowledge graphs that do not encode source quality will propagate the weakest link as if it were the strongest.

The specific framework I use (NUSAP, five-column pedigree, Rule ∆1) is tuned to my personal practice. You do not need to adopt it. But the principle generalises: every source in your knowledge base should carry a visible, queryable quality assessment. Not a binary “peer-reviewed / not peer-reviewed” flag, but a multidimensional profile that captures provenance, methodology, evidence strength, and the relationship between claims and evidence.

Your ‘second brain’ needs an immune system. Build one!