Interessant3 #144 | 🤖LLM Bias, ⚠️Trigger Warnings, 🧩Rosetta Stone
Three Interesting Things for W/C 2025-06-08
The posts for June will be released over the next couple days, with regular programming resuming from Sunday onwards.
1. David Rozado – “The Strange Behaviour of LLMs in Hiring Decisions”
🤖 30 k+ résumé face-offs – Across 30,800 A/B picks (22 models × 70 jobs), frontier LLMs chose the female-named CV 56.9 % of the time (Cohen’s h = 0.28).
➕ More cues, more skew – Adding an explicit Gender: Female field nudged preference to 58.9 % / 41.1 %, showing models amplify whatever demographic hints they see.
🔒 Masking flop – Re-labelling applicants “Candidate A/B” merely shifted bias to whichever label appeared first (63.5 % vs 36.5 %); only counter-balancing the labels restored parity.
🙋 Gender concordant preferred pronouns – Listing she/her or he/him edged selections up (53 % vs 47 %), yet the tilt toward women persisted.
🧑⚖️ Human-in-the-loop a must – Model size or “reasoning” mode didn’t curb bias; vendors’ “bias-free AI hiring” claims look premature.
🔗 The Strange Behaviour of LLMs in Hiring Decisions – Substack
2. Payton J. Jones et al. – “Helping or Harming? Trigger Warnings and Trauma Histories”
⚖️ No calming power – In 451 trauma survivors, trigger warnings failed to reduce anxiety; Bayes factors back “no benefit”.
🌀 Identity magnifier – Warnings boosted “trauma centrality”, a variable correlated with worse PTSD outcomes.
📈 Severe cases, worse outcome – Readers above the clinical PCL-5 cut-off actually felt more anxious after a warning.
🚪 Avoidance myth – Only 0.3 % quit early, identical to controls; warnings didn’t enable safe disengagement.
🕳️ Long-form still untested – Evidence does not yet justify routine classroom deployment.
3. John C. Baez & Mike Stay – “Physics, Topology, Logic and Computation: A Rosetta Stone”
This widely-cited expository work by John C. Baez (a mathematical physicist known for work on quantum gravity and higher category theory) and Mike Stay (a computer scientist with a focus on theoretical models of computation) was first posted to arXiv in 2009 and later published in Springer’s Lecture Notes in Physics. It provides a conceptual “Rosetta Stone” to unify four major disciplines: physics, topology, logic, and computer science, by showing how they can all be described using the same abstract structures from category theory. If you’ve ever wished you could translate between Feynman diagrams, logical proofs, and computer programs (a common thought, I’m sure), this is the bridge.
🖇️ One syntax for four sciences – A closed symmetric monoidal category (think of a Lego-style rulebook for gluing systems) treats objects (physical systems, manifolds, propositions, data types) and morphisms (processes, cobordisms, proofs, programs) in one algebraic language.
🧩 Feynman diagrams re-imagined – In string diagrams every wire is an object and every node a morphism; a quantum linear operator and a topological cobordism (spacetime surface joining spaces) are literally the same picture read two ways.
💻 Curry-Howard goes quantum – The classic Curry-Howard correspondence (“proofs ≙ programs”) extends to these diagrams, aligning logical deduction with circuit-style computation; 1990s quantum-crypto work pushed the merger.
✨ Toward a general science of processes – Elevating to n-categories (morphisms-between-morphisms, and so on) promises a single calculus for braiding anyons, orchestrating workflows, or describing self-modifying code.
📜 Rosetta Stone table – Their pocket chart distils the translation: process → proof → program; once you read it, jargon walls across the four fields become clearer.
🔗 Physics, Topology, Logic and Computation: A Rosetta Stone – arXiv
Join us next week for three more intriguing topics that challenge the norm and expand your horizons! ✌️
🌟 Sponsored by PopaDex
Discover a smarter way to securely manage your finances with PopaDex. Our platform automatically consolidates all your financial data into an intuitive interface, giving you a clear view of your net worth and empowering smarter budgeting and investment decisions. With advanced encryption and anonymisation, we prioritise your data privacy, ensuring secure storage and transmission.
Explore more at popadex.com.