Профиль

The weird thing about Badiou is that his definition of Events in terms of self-belonging can be mathematically internalised by moving to Hyperset theory, which, ironically enough, is like an abstract scaffold for looping dynamically unfolding structures, i.e., a natural way of thinking time.

I may not like his commitment to set-theoretic model theory, but his use of it is extremely different from most mainstream analytic phil, in virtue of what is basically a subtractive mathematical reframing of Heidegger's ontological difference. The ontic only properly turns up in Logic of Worlds.

I'm not sure about that to be honest. I need to revisit an unfinished piece on Derrida. I think you're wrong about Badiou though. He is if nothing else resolutely committed to a mathematical analysis of the non-determinism inherent in the unfoldiing of thought via independence/forcing.

The link wasn't meant to be there. The app converted a typo I later corrected into a link without me noticing and now it's stuck there!

But using inferential connections as control signals provides leverage to then shape more powerful control structures by using natural language, and that's where we get into the land of constitutions, personas, and more or less ad hoc attempts at linguistically shaping control.

I say semi-agents because it's not obvious that the sorts of control structures created simply by, e.g., verified reinforcement on math problems do much beyond canalising the non-deterministic (or perhaps, unpredictable) behaviours imbued by irreducibly noisy training data.

Probably also worth noting that this is exactly what chain-of-thought was, in some sense. The control structure added on top of the base model is precisely what enabled RLVF strategies to then reinforce inferential structure over distributional association, breeding semi-agents.

My claim is this: the problem of impressing an explicit control structure upon a neural network and the problem of designing a control structure that can automate backpropagation in a compositional manner that isolates and trains regions of the network are essentially the same.

This is to say that frame semantics (and indeed, the implicit KRR structures in OOP class systems) are compositional in precisely the right way, we just never figured out how to automate them at scale. What does this representational problem have to do with control then?

There's a nice critique of the LLM paradigm implicit in approaches from later waves of GOFAI knowledge representation and reasoning strategies that were designed to respond to the frame problem, e.g., frame semantics, even when they were not scalable due to too much hardcoding.

So we don't just grow representational structure in the network, we also grow control structures, but these are grown in a more or less maximally holistic way, which is why they are so goddamn computationally and data resource intensive.

I increasingly think that most AI alignment worries are corollaries of architectural facts about the way DNN training happens to be done: that backpropagation is structurally frame-problem like entails that there can be no explicit control structure imprinted on the network.

The graduated community were wonderful, and for all I am ragging on Derrideans, I miss the company of Simon Fuller, who is sadly no longer with us, and the motley crew of PhD students who we would spend long nights talking with well before I even hit my MA. Not to mention the UG crew.

Other members of the department on the continental side worth a few stories of their own that alas I can’t spend my night telling: Keith Ansell-Pearson, Peter Pollner, Christine Battersby, and Darren Ambrose (who taught me aesthetics on sabbatical cover, and much besides).

What might have been. But just as with important insights learned from Greg, David, and others, it has taken me more than 15 years to really appreciate the deep structure underlying that observation, so maybe I’ll finally write the thesis (or at least the paper) I should have written back then.

There was a moment during one week of Stephen’s course on the Science of Logic when during the discussion I observed that most philosopher (Aristotle, Husserl, Kripke) understand essence through identity and necessity and Hegel does the exact opposite. Stephen said this was a PhD thesis question.

Which brings me to Stephen, the living avatar of Hegel on Earth. He probably had a bigger impact on me than any other single member of staff besides Greg, teaching me at both BA and MA levels. My greatest PG regret is that I did not pick him as a PhD supervisor. He’s an intellectual force of nature.

But my favourite anecdote about Warwick is that, after being introduced to Brandom by a sufficiently open minded McDowellian ethics TA, I later discovered at least 9 copies of it on the shelves of various department members, but only Stephen Houlgate had read more than a couple chapters.

I also learned a lot from Stephen Butterfill in philosophy of language and logic (particularly the problem of vagueness), though my views and my exam answers in the latter did not necessarily reflect an honest engagement with what I took to be wrong about the standard semantic approaches.

This isn’t to say I learned nothing from the other side. I learned a lot from Bill Brewer’s peculiarly Berklean approach to philosophy of perception and the history of empiricism. He once complimented me on being Humean, but I made the mistake of mentioning Derrida and his eyes visibly glazed over.