Barry Norton 🌈 ♾ 📚

Due to a peculiar medical condition…

There are goldfish calling out this guy’s horseshit (and the half of the country who voted for him again)

I’ve been inspired by so many papers I only dreamt existed.

Now I don’t even need to dream.

The latter is just a case of where money is made off the back of your work, not whether it is. The former (if even honestly considered) doesn’t protect your work in the rest of the world.

The UK is dancing around two legislative options:

1) make the models themselves infringing if used in the UK;

2) making at least this US precedent into a statutory exception in the UK so that people train their models in the UK (rather than risk changes of opinion in appealed civil US cases)…

The thing is, the use of creative works in pre-training is neither theft nor infringement according to US case law. Once someone buys a copy of your work — and it can be second hand, not giving you a penny — they’re free to use it in training a generative model there, as that is ‘transformative’…

While I'm ranting: Linux is a kernel, not an operating system.

Android has a Linux kernel.

Not happy with 'pervert' being pejorative (with regard to smart glasses being used in unacceptable ways).

This is not compatible with reclamation in kink and queer spaces. Find better.

Teaching, marking, admin and proposal writing. Not what my imagined ideal had been.

As someone who left academia, I was disappointed by how little time it afforded just to think and write.

Good Christ.

In other news, I've had an Even Realities tab open for weeks — glasses with heads-up displays, but no cameras or microphones. I'd like them, but I think the well's been well and truly poisoned.

Even where messy humans are part of the orchestration, process modelling (my PhD) and semantic process mining (a really interesting period of my postdoc work) exist — this 'context graph' stuff is another wheel being reinvented with a new hammer.

I think you answered your own question while I was replying (with very similar thinking).

Yupyupyup. One that at least made an inspectable deterministic plan. Post hoc introspection by LLMs has been shown to be nothing to do with execution, and a priori 'reasoning' doesn't lead to determinism.

He was obsessed — like, obsessed — with William Burroughs and cut-up too. So much so that he's angry with me for going with Magic 8-Ball (it's just more accessible to most).

(OK, John Cage and David Bowie too)

Mid-90s Barry also treasured a copy of UK Wired magazine (the first iteration, before it folded) with Eno on the cover. Deep inside me, he's thrilled. (Sorry for the fanboying).

I'm as likely to trust OpenClaw to coordinate my tools as I am to walk out into traffic blind-folded and with earplugs with a magic 8-ball for a guide dog.

(I hope that's not a problematic analogy — I need some coffee)

RAG doesn't solve this. Not nearly. And through GraphRAG it was exciting to see knowledge graphs, and thereby ontologies, become cool again. But they don't fix this.

It's a huge advance, don't get me wrong. And I work on this, professionally. Ng's paper in 2009, and subsequent deep learning, then Word2Vec in 2013 fundamentally challenged what we had earlier been doing. But language models' frequent preference for parametric over contextual knowledge — sand.

Bruce Sterling. I'm discussing 1990s AI with Bruce f***ing Sterling!

Again, that's what DAML-S was about.

All of these shopping and travel agent examples are last century AI tasks re-heated. To the point that the ontologies community were literally laughing at loud on seeing them, like old friends.

Just in case I sound combative (it's actually being defensive of this work each side of the turn of the century), let me just say that if I could go back to the mid-90s and tell myself about this, it's a discussion with the co-author of The Difference Engine, not LLMs, that would blow his mind.

Multi-agent systems always included on-going interactions, and planning was always iterative.

I don't see how agents are anything but AI Planning, plus the flawed approach to 'latent' semantics and reasoning (i.e. language-based via embeddings and self-attention).

Language models, trained at scale using transformers, are amazing. Amazingly good at natural language. Plus their parametric 'knowledge', 'reasoning' and 'agentic' behaviours were a real surprise — but all are fundamentally fragile and prone to hallucination. All applications are thus built on sand.

The whole notion of MCP to expose tools to agents: that was DAML (DARPA AGENT mark-up language), then later OWL-S aka Semantic Web Services — again part of my PhD thesis.

Embeddings, then transformers, start from the flawed notion that you can do all of this via natural language.

Before AJAX came MS OLE (later COM). I wrote my bachelor's dissertation on it. Using AI technologies to dynamically orchestrate (SOAP/WSDL plus AJAX) using these kinds of interfaces (aka multi-AGENT systems) — my PhD thesis, conceived 1999. Same, but REST (JSON) — postdoc. Nothing new under the sun.

I wish people would come to understand transphobia as systemic, structural, and pervasive so we could stop treating transphobia as explicitly limited to people saying "I hate trans people" and accept the way our society conditions us all to dismiss, diminish, and dehumanize trans people.

Every time I listen to a podcast in Denmark I get an advert with vacation locations whose names are butchered — and I've never been 100% sure that that's the joke.

I hadn't even noticed this.

"Invisible whips directing invisible labor"

Someone wrote tumescent, having always been disappointed to have been born after emancipation.