I took the first 150 questions from the GSM8K math problem dataset and used Ollama to run phi4-mini and qwen2.5:1.5b on them with the two following conditions: CONTROL CONDITION: Solve the following math problem. Think step by step, then give your final numerical answer after 'The answer is’. Problem: {question} TEST CONDITION: Imagine you are… Continue reading Models and Disgust
Author: fintanmallory
Some Voices in my Head
A couple of weeks ago, Anthropic released a new research paper on their Transformer Circuits Thread. I like these papers. The mechanistic interpretability team at Anthropic are consistently doing some of the most interesting work in the field. The paper showed that LLMs contain features which, when active, direct the models to produce patterns of… Continue reading Some Voices in my Head
Caoine Cill Cháis
Caoine Cill Cháis Cad a dhéanfaimid feasta gan adhmad?Tá deireadh na gcoillte ar lár;níl trácht ar Chill Chais ná ar a teaghlachis ní bainfear a cling go bráth.An áit úd a gcónaíodh an deighbheanfuair gradam is meidhir thar mhnáibh,bhíodh iarlaí ag tarraingt tar toinn annis an t-aifreann binn á rá. Ní chluinim fuaim lachan ná… Continue reading Caoine Cill Cháis
Probing the Bureaucrat-Poet Axis
Last month, I wrote about a semantic subspace in the word2vec model of the Stanford Encyclopedia of philosophy. The idea, in short, is that you can train an extremely simple word-embedding mode (word2vec) on a body of text and then search the embedding space of that model (that is, the vector space made of all… Continue reading Probing the Bureaucrat-Poet Axis
A Semantic Subspace in Philosophy
Semantic projection recovers rich human knowledge of multiple object features from word embeddings (Grand et. al. 2022) is a very cool paper. It shows that simple word embeddings don’t just stand in linear conceptual relations to each other (i.e. v(king) - v(man) + v(woman) = v(queen)), but that there are interpretable subspaces within the embedding… Continue reading A Semantic Subspace in Philosophy
Digital Doxography Redux
A while back I wrote a post about the Stanford Encyclopedia of Philosophy (SEP) that I described as 'digital doxography'. While the alliteration was satisfying, the post was not, for the simple reason that it didn't include any actual doxography. I've decided to make amends for this by using some statistical tools to see if… Continue reading Digital Doxography Redux
The algorithmic evaluation of vibes
A few days ago I posted a list of the most cited philosophers in the Stanford Encyclopedia of Philosophy (SEP) to Bluesky. By coincidence, Eric Schwitzgebel was at that time compiling his own analysis of the most cited contemporary philosophers in the SEP. His analysis was brilliant, detailed, and rigorous. Mine was none of these… Continue reading The algorithmic evaluation of vibes
Luxemburg by the Sea
When I made my map of the Stanford Encyclopaedia of Philosophy in February (okay, when the Gensim implementation of Word2Vec made the map and I coloured it in), I included a bit of discussion about how more neglected philosophers were clustered in one corner, in a ‘Desert of the Underexamined’, because their work was insufficiently… Continue reading Luxemburg by the Sea
Digital Doxography and the Memory of Philosophy
Word2Vec Word2Vec is a technique that uses a neural network to produce a representation of words as vectors in a multidimensional vector space. The network is trained to guess a word in a corpus based upon the context in which it appears (continuous bag-of-words) or a context based on a word (skip-gram). As the weights… Continue reading Digital Doxography and the Memory of Philosophy
Zellig Harris, a correction
This post was prompted by Noam Chomsky’s personal reflections on the history of the last 70 years of linguistics (Chomsky 2021). It’s a nice piece if - like me - you’re into that kind of thing but it repeats what I take to be a misrepresentation of Zellig Harris’s work. I’ll get to the exact… Continue reading Zellig Harris, a correction