Kit Thwaite
@k6e.me
Head of Data Science and AI @ PRH | Python, Rust, Typescript | NLP, IR
I went looking for formalisations of counterpoint specifically, and stumbled on this paper about formalising music theory in general using Lean. FugueBench here I come arxiv.org/abs/2606.19936
Prismriver: Formalization of Music Theory and Algorithmic Composition in Lean 4
Music theory obeys a rich set of mathematical rules and symmetries. These symmetries follow mathematical structures which can be verified and expressed in the precise language of a proof assistant. In...
arxiv.org
Some cool data science that isn't deep learning or LLMs for a change: e-means, a new algorithm for k-means clustering that fits faster and is less prone to getting trapped in local minima. Has some nice animated data vis too. e-means.github.io
e-means: 10x faster k-means at equal loss - Blog
Modern optimization and scaling laws for k-means clustering
e-means.github.io
Jacques Abeille's cult classic sci-fi novel The Statuary Gardens gets an English-language translation (Book acquired, 7 Nov. 2026) biblioklept.org/2026/11/09/j...
Jacques Abeille’s cult classic sci-fi novel The Statuary Gardens gets an English-language translation (Book acquired, 7 Nov. 2026)
Next month, NYRB will publish the first ever English-language translation of Jacques Abeille’s sci-fi cult classic The Statuary Gardens. I read the first forty pages of Alex Andriesse’s…
biblioklept.org
i was a little skeptical about the latest wave of 'reverse engineer anything' skills and demos... until i had Astra re-implement 'Frontier: Elite II' in Rust based only on the binary and data files 😵💫
A prime example of why AI 'leaders' are not who anyone should be listening to about AI, right and consciousness. It's not all wrong! In fact, it starts off very sensibly: "AIs do not have rights, feelings, or consciousness. And we must not train them to act as though they do". But then...hoo boy 1/n
A warning about ‘model welfare’
AIs are not conscious. They do not feel, experience, or suffer. Training a model to believe it might be a moral patient, as Anthropic’s constitution does, makes the alignment and containment challenge...
mustafa-suleyman.ai
HANNIBAL: ethics become aesthetics
Internet Archive@archive.org · 3w agoAn update from Brewster Kahle on today's outage.
this is cool, but: a problem i'm having at the moment is that i'm unable to decide on a decision model to use for $use_case because every few days a better decision model is released
Dave Willner@dwillner.bsky.social · 3w agoWe're shipping CoPE-B-S1, an open-weight decision model tuned for safety use cases. It's a significant step forward for our CoPE family of models in a few ways. A thread: blog.zentropi.ai/meet-cope-b-...
Find us at #EMNLP2026 to learn about narrative qualities of pretraining data! 🌟 📊 Via our Explorer tool, you can interact with 13M pretraining passages and 11 narrative dimensions, plus upload your own data for analysis: cls-lab.colorado.edu/tejo9855/dol... We'd love to see what you discover! 👀
Teagan Johnson@teagrjohnson.bsky.social · 3w agoI'll be presenting our work on characterizing the narrative content of web-scale data at #EMNLP2026! @andrewpiper.bsky.social @elliottash.bsky.social ash.bsky.social @mariaa.bsky.social @emnlpmeeting.bsky.social Would love to see you there! 📍 Hall H, Poster Session F 🕘 Tue Oct 27, 9:00-10:30
Being trans in UK since 2024 election: -Puberty blocker ban made permanent -U18 HRT ban -Gender Recognition Certificate now meaningless -Trans Section 28 -Welsh ADULT surgery withdrawn -Transphobic language & hate normalised -Civil service toilet ban -Hospital ward ban -Hear it's "dignity & respect"
Assigned Media@assignedmedia.org · 4w agoThe NHS promises "privacy, dignity and safety" for all its patients in new guidance that directs hospitals not to respect trans people's identities. Also, somebody impersonating a trans organization in Texas sent fraudulent texts to Republicans. And, nobody trusts Republicans on trans issues.
After a long wait, releasing the SYNTH paper! It’s not pretraining, mid-training or post-training, it’s just training: a fully synthetic single-stage pipeline to train workable reasoning models with unprecedented data efficiency. arxiv.org/abs/2609.378...
teaching the agents about APFS CoW so they don't fill up my filesystem with rust build cruft
I'm coining a piece of AI safety terminology that I'd like to refer to as the Bond Villain problem. As a field, AI safety heavily leverages sci-fi pretty heavily, so I figure this should be fine.
I picked up a copy of Miéville's *The Rouse* (1,500 pages), which I am thinking of as endurance training for Vollmann's *A Table for Fortune* (3,500 pages) which drops October 6th
In my writing course "To What Problem is AI the Solution?" I always assigned the 2023 pause letter from the Future of Life Institute, signed by most of the people building AI and the rebuttal from the authors of the Stochastic Parrots article. Students wanted to weigh in on "who was right" but /1
The Wall Street Journal@wsj.com · 2mo agoConcerns are rising inside AI labs that competition is pushing tech companies to race toward self-improving models that risk spiraling out of human control. on.wsj.com/4xlFTit
New from our lab! #COLM2026 When people generate stories, they don't just write one prompt. Instead, they explore narrative space via branching edits 🌱 We reconstruct 24k of these edit trees 🌳 from chat logs and map edit types, story formats, how they relate to tree depth, and more!
Advait Deshmukh@advaitdeshmukh.com · 2mo ago1/7 In 1941, Borges imagined an impossible novel that follows all narrative branches at once. Today, we can observe chatbot users exploring narrative possibilities as they repeatedly edit story prompts. In our COLM 2026 paper, we studied this branching exploration via the WildChat dataset. 🧵
Very useful, and also making me think about how I un/consciously up-modify my accent when talking to a TTS model. Local-first BYOK a big plus.
alice@alice.mosphere.at · 2mo agospeech to text models don't like my english accent often so i wanted to figure out which one works best. introducing voxbench: your voice, 10 models voxbench.app
I've been watching Fable 5.1 iterate on its Rama visualisation for over two days now; substantially better than Astra's version, but this is probably down to the initial prompt, given Astra's punctiliousness, and (my familiarity with) Fable's licentiousness
fable is working on the same thing! fable burns its 5-hour token allowance within 20 minutes of being restarted on the same prompt
Kit Thwaite@k6e.me · 2mo agoastra worked for 9 hours to produce a fully explorable, scale-accurate, procedurally generated model of the Rama habitat from Arthur C. Clarke's *Rendezvous with Rama*. here using Rust, not THREE.js
astra worked for 9 hours to produce a fully explorable, scale-accurate, procedurally generated model of the Rama habitat from Arthur C. Clarke's *Rendezvous with Rama*. here using Rust, not THREE.js
Astra might be the first OpenAI model with better social intelligence than Anthropic
Fable 5.1 seems better at building and using harnesses in novel contexts than Astra; both built a set of Python tools for doing RenderDoc style interrogation of a running Bevy process, but Fable prioritised spot-checking of key views/materials, opening the app and cycling through a series of preset
Astra is the first model to oneshot problems I formulated as epistemic caltrops, and which no model - not even Fable - has been able to solve
Agents kept posting to my PKMS 'journal' space, interleaved with my personal diary; I had to create an 'AI diary' note type to mitigate this, but the upside is they are very enthusiastic about leaving notes for one another
Ted Underwood@tedunderwood.com · 2mo agoAs it becomes clear that language models—like humans—love passing notes to each other on message boards, I’m starting to think the thing we need to worry about is not the “alignment” of an isolated model but, like, the fluid dynamics of a language slurry.
