iiinova — where your digital self keeps evolving

Obviously the landscape has changed dramatically since March -- but I think this document still holds up well given its scope.

000

We define 'tampering' as 'adapting an ML model or circumventing the safeguards or guardrails of that model for the purpose of using it for harmful activity not intended by the model creator'.

100

Back in March we finished a piece of work I led on tampering with generative AI models. The output for this is a summary of the literature aimed at policymakers seeking to build mitigations. This document is now available online here: www.turing.ac.uk/sites/defaul...

turing.ac.uk

Also learning from this process how much more fiction stays with me, despite the fact I read both fiction and non fiction equally (or the latter slightly more)

000

This was very difficult to put together. The inclusions are much easier to defend than some exclusions: - One, No One, One Hundred Thousand by Pirandello - The Anatomy of Fascism by Robert Paxton - Calvino - Orwell

God I really love making lists myninebooks.com/g/5ea8401c7899

A picture from myninebooks showing the following selection:

1. Catch-22 by Joseph Heller
2. Cat's Cradle by Kurt Vonnegut
3. The Remains of the Day by Kazuo Ishiguro
4. The Name of the Rose by Umberto Eco
5. Fictions by Jorge Luis Borges
6. Human Acts by Han Kang
7. The Rings of Saturn by Winfried Georg Sebald
8. The Story of the Lost Child (book 4 of the Neapolitan Quartet) by Elena Ferrante
9. Inventing the Renaissance by Ada Palmer

100

Of course, you can call this thinking or understanding if you want. But in my view you would be redefining the word beyond all recognition.

001

LLMs are astonishingly effective, but the notion they sample from sophisticated distributions over text is a better fit than "thinking" or "understanding" because it also fits their failure modes.

100

So the answer is that capability on things which require thinking and understanding in humans is not prima facie evidence of thinking or understanding. We don't think chess engines think or understand because we have better theories of how they work which.

100

On Evaluating Cognitive Capabilities in Machines (and Other "Alien" Intelligences)

But there is a deeper problem where the field as a whole is systematically weak in its methods, leading to this being poorly explained: aiguide.substack.com/p/on-evaluat...

aiguide.substack.com

(Apologies for the length of this post, which means it gets cut off in the email version.

100

If you accept that LLMs do no think or understand in the ways that humans do—and how we do these is not well understood—then whether they do "think" or "understand" depends mostly on how broadly you are prepared to define these terms.

100

Phil Swatton @philswatton.bsky.social · 19/09/2026

I think people are right to say there is a third option, but saying so starts from objecting to the combination of "by any any reasonable stabdard" and "very different from us".

100

How come? Is it on a similar line to Rorty's anti realism or something different?

040

Phil Swatton @philswatton.bsky.social · 13/09/2026

Great primer for anyone looking to understand the Saxony-Anhalt result and how to properly contextualise it, along with why the BSW is important to the prospects of an AFD government there

Phil Swatton @philswatton.bsky.social · 13/09/2026

Some reflections on @yuanfenyang.bsky.social's brilliant Rita Hinden memorial lecture hosted by @renewaljournal.bsky.social on ending 40 years of neoliberalism. dysfunctionalprogramming.substack.com/p/embracing-...

dysfunctionalprogramming.substack.com

Embracing Uncertainty

Reflections on the 2nd Rita Hinden Memorial Lecture by Yuan Yang

144

Phil Swatton @philswatton.bsky.social · 01/09/2026

And the widespread practice of 'jailbreaking' LLMs guarantees that safeguards won't be sufficient to prevent this. Worrying about AGI is usually worrying about the wrong thing!

010

Phil Swatton @philswatton.bsky.social · 01/09/2026

For non-AI folks, abliteration is a technique for removing safeguards from a model. The reason I think the OpenAI cybersecurity stuff is concerning is that open-weight models lag closed-source models in capability by ~6 months, meaning that soon, many attackers could get an uplift in capability

163

Phil Swatton @philswatton.bsky.social · 25/08/2026

Great quote from a 100% correct piece. We know all of our models of the world are wrong, but we still have to do research and take action in the world. How do we balance those concerns?

011

I'd agree with 'liberal democratic' and the reasons why a consensus was built around it. I think the case for it on the right is the basic case of peace treaty where we all play by the same rules to avoid authoritarian or violent alternatives, rather than that I imagine it's a neutral choice.

100

I'm not sure what point you are making with this list. I'm aware what you had in mind and signalled some common ground in the second part of my reply. If you think I'm nitpicking, you can choose not to reply. If you think language is important and signals ideas, you can address the point.

100

I think my point is that "prevent bad outcomes" is hopelessly vague and open to abuse (you can imagine an authoritarian banning opposition parties using the same language), not that parties shouldn't be acting as filters on candidates or enforcers of norms (which I advocate for in the thread)

210

But this kind of norms enforcement isn't about stopping the public from choosing 'bad outcomes' - if the centre-right wins an election, I usually think that's a bad outcome given my political beliefs! But it isn't a violation or undermining of democratic norms, which is the thing we care about here

100

Of course - what you do once the mistake has already been made is another q. As is the q of if the Republican Party still thinks its long-term interests are served by allowing Dems to take power again, having violated the law many times over. Avoiding punishment is an incentive not to cede power.

Agree this is *the* critical problem of C21 politics - in 'How Democracies Die' (not about the radical right per se), Levitsky & Ziblatt argue for strong norm enforcement by political elites vs antidemocratic outsiders as a solution to this kind of problem, and against attempts to co-opt them.

100

I do agree though that it is important to protect vs authoritarian actors by ensuring strong consequences as in South Korea, or as there should have been for Trump over Jan 6th. Less about outcomes, more about strong enforcement of norms or rules of the game

Granting that this format isn't good for nuance, I dislike the language of "protect the public from being able to choose bad outcomes". Strikes me as straightforwardly an argument for dictatorship and hostile to the pluralism the republican tradition values (who decides what counts as bad?)

200

Yeah I think this is the basic hope. Don't over-assume the permancy of the turn to the radical right, try in the meantime to keep them out of power until (moderate, pro-drmocratic) conservatives are able to reasssert themselves as the mainstream of the right

0141

Conservative Parties and the Birth of Democracy - Wikipedia

Agreed but would go further - in any party system where the antidemocratic right is the dominant right faction, democracy is trouble. Suspect elite and civil service enforcement of democratic norms is the important thing On my TBR is this which I think is relevant: en.wikipedia.org/wiki/Conserv...

en.wikipedia.org

2252

Hetan Shah @hetanshah.bsky.social · 08/08/2026

I hope so too, but I fear that how bad things got will be quickly forgotten if/when Dems get back in power and the basic alternation between left and right governments continues. The health of democracy depends on having both a left and a right whose mainstream is committed to democratic norms

4594

Reposted by Phil Swatton

Oh this is a great piece from @add-hawk.bsky.social making the case that play and games are the meaning of life, drawing on The Grasshopper by Bernard Suits www.ft.com/content/7b95...

Imagine, says Suits, utopia - some future paradise where technology has solved all our practical problems. We have perfect medicine, unlimited energy and boundless resources. What would we do with our time?
We would play games, says Suits, or we would be bored out of our minds. And if playing games is all we do in utopia, then games must be the meaning of life.

95719

Phil Swatton @philswatton.bsky.social · 24/07/2026

I did a brief lit review on AI text detectors for work back in Dec '25, so hopefully not too dated. Evaluations often fail to provide realistic settings and lead to over-optimistic estimates of performance - of course can't say whether this is true of Pangram: www.turing.ac.uk/sites/defaul...

turing.ac.uk

130

Phil Swatton @philswatton.bsky.social · 08/07/2026

It isn't a net gain if I have to spend more time understanding bloated LLM generated code when reviewing PRs!

020

Phil Swatton @philswatton.bsky.social · 08/07/2026

Two brief thoughts on why I think it's still worth keeping LLM use out of introductory programming for stats/data science: - You need to be a good programmer to get good code out of an LLM - You need to be a good programmer to evaluate the code you get out of an LLM

26510

Phil Swatton @philswatton.bsky.social · 01/06/2026

Very proud of my wife Samira for passing her viva today. Blessed to be married to such an intelligent and accomplished scholar

030

Phil Swatton @philswatton.bsky.social · 29/05/2026

This leaflet format is the British political equivalent of carcinisation

121

Reposted by Phil Swatton

maybe: annie rauwerda @annierau.bsky.social · 29/05/2026

"why use google?" webpage from 1999

8069951813

Touché - though his home sec at the time was Callaghan (of the old right), and Wilson himself while starting as a Bevanite was v much a political chameleon

010

Yes - and given that loose meaning, I'm not sure it makes sense to identify Labour's _historical_ comfort zone with the _soft left's_ comfort zone, because that latter group isn't generally comfortable with immigration restrictionism?

000

We agree on the history - I think I associate that history, perhaps erroneously, with the old right, and had generally assumed Labour's soft left is pro-immigration

210