So I had this idea yesterday about how LLMs might be a kind of strange attractor and how the trajectory of a response to a prompt is influenced by those attractors within the model. With enough prompts, the relative position of attractors can be mapped, and this information can be used to see if a model has changed. If the prompts are getting different responses but the attractors haven’t moved, then its something in the context (e.g. system prompt) upstream.
I pinged the people I know who might be interested, and 24 hours have gone by and not a single response. I may have to go do this on my own, which is really disappointing.
At least ACM is responsive on book things.
Latent Reading of Finnegans Wake by Nina Beguš, Tessa Haining, Gasper Begus
- Finnegans Wake has repeatedly served as a limit case for theories of communication, computation, and literary representation. From Claude Shannon’s account of the Wake as an extreme form of “semantic” compression to Jacques Derrida’s description of Joyce’s text as a “hypermnesiac machine,” the novel has seemed to anticipate the media systems later used to explain it. This essay returns to that feedback loop in the age of artificial neural networks and proposes latent reading: a critical literary practice in which a model trained on a literary work becomes an additional object of literary interpretation. We train and analyze two models on the Wake: FinneganLM, a GPT-2-based language model trained from scratch on the novel’s text, and FinneGAN, a speech-generating GAN trained on audio derived from the novel. We distinguish external latent reading, which analyzes narratability, representation, and navigability in both models, from internal latent reading, which applies AI interpretability techniques to learned representations. Employing both interpretability and literary interpretation, we show how latent reading extends the practices of close and distant reading, and opens an entirely new area of literary research, by exploring what kind of machine a book can make.
- The notion that the State Department had been working to silence Americans is perhaps unfamiliar to many people, but it’s a key part of a larger web of conspiratorial ideas that have been promoted for a decade on far-right blogs, podcasts, and social media. The basic theory is that, under the guise of combating disinformation, a sprawling constellation of government agencies, academics, civil society groups, and Big Tech platforms—or what believers refer to as the “censorship-industrial complex”—has been working to suppress conservative and populist speech online. The fight against “disinformation,” as Benz put it on a 2025 podcast, “is censorship in disguise.” R/FIMI, along with a predecessor office with the same mission, the Global Engagement Center (GEC), had been among the right’s long-standing targets; now it was one of its latest casualties.
Tasks
- Bills – done
- Dishes – done
- Chores – done
- 11:00 Fidelity – done
- Download next chapters to review. Done. No chapter 5?
- Read a paper
- Finish going over Panopticon paper
- Talk to Stacey about list – sent email
- Groceries
SBIRs
- Timesheet – done
- Some firedrill that HAD TO BE DONE NOW
viztales 