Phil Swatton

Pleased that our paper on the ability of LLM-driven persona agents to simulate human personality dimensions has now made it to publication in Royal Society Open Science: doi.org/10.1098/rsos...

In short, these agents do in fact simulate personality dimensions but these are distinct from those of humans.

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...

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'.

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

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

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".

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.

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".

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...

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".

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.

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

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

Another one of my reports written at the Turing is now available online, which is an informal literature review on AI-generated text detection. A brief thread as I've seen some discussion on the topic ⬇️ Link: www.turing.ac.uk/sites/defaul... 1/5

5. All detection solutions are highly vulnerable to simple perturbations and edits, and this is not reflected in most benchmarks 6. Among proprietary solutions, Originality appears promising but lacks sufficient evidence and likely suffers from the same limitations as other approaches 4/5

Another one of my reports written at the Turing is now available online, which is an informal literature review on AI-generated text detection. A brief thread as I've seen some discussion on the topic ⬇️ Link: www.turing.ac.uk/sites/defaul... 1/5

The review was based on aggregating results from existing benchmarking efforts. We identify some core results: 1. SoTA detection depends on in-domain transformer encoder classifiers 2. Detectors perform best on the training data + models they were trained on 2/5

Great obituary of Habermas by @mattpolprof.bsky.social: jacobin.com/2026/03/habe... I only recently learned of the outline of Habermas' thought after reading Matt's recent review of his career: www.liberalcurrents.com/also-a-revie...

Great post. I think the biggest examples of this kind of motivated causal attribution on my side of the aisle is the notion that Russian disinfo changed the outcomes of the 2016 EU Ref/2016 US Presidential election. Disinfo is v. concerning for a variety of reasons, but rarely changes vote choice.