LYDIA
i care about the great present, good futures, & differential technological progress! i research ML in san francisco and read mathphil at oxford.
Current projects:
RLing LMs
'Threading the needle'
'Table stakes'
Yes, follow me at https://curius.app/lydia-nottingham
personal favorites
- Vitae per person (VPP): a new metric of civilizational resource
- How we collected 10,000 hours of neuro-language data in our basement
- peripatetic: 7; ludic: 5; corollary: 8
- Beware the Delmore Effect
Community
- Pay for results, not failures
- Open-source your life in a closed group
- Inaugural cohorts as 'nothing above'
- 90/30 ML Reading Group: Retrospective
mathematics
macrostrategy
- The Good-Faith Actor Principle
- On Redistribution
- Vitae per person (VPP): a new metric of civilizational resource
- Full-stack futurism, apparently
- 'Which Future?' (Nielsen): best insights, extracted
scratchpad
- Mutual Information in Probabilistic Graphical Models
- Back when weights were outrageously interpretable
- The cost of getting good
- on bootstrapping morality
- why don't we explicitly train models to be good at generalization?
- the trouble with making predictions as an agent
bio
Research
- Testing models' self-forecasting abilities
- Lit Review: Stated vs. Revealed Preferences in LMs
- I fine-tuned Llama-3.1-70B on my blog posts
- How to get your LM to give strict, short, decisive outputs
Machine learning
- The Bitter Lesson Eats 'System Prompt Optimization with Meta-Learning'
- How we collected 10,000 hours of neuro-language data in our basement
- Defending the Data Processing Inequality
- The Bitter Lesson: Six Years On
- Let's automate causal inference!
how should we understand alignment?
- We need to make ourselves people the models can come to with problems
- Model AI decision-making through Markov chains, not utility functions
- Three Degrees of Generalization
- Backdoors are a Lipschitz clamping problem; Grokking is context-dependent
- Alignment is a predictability problem; whack-a-mole is a losing game
- Update re 'spontaneous instantiation'
Programming
collective action
Stack
Artificial intelligence
- AIS student, you can be an IC
- open-source / open-weight distinction
- Field Notes from EAG NYC
- GPT-5 Livestream Notes
- Announcing Manival
- (Not) Buying a Chinese AgiBot
- 'GiveWell for AI Safety': Lessons learned in a week
what should one person do?
- Transforming AI Safety into a 'numbers go up' game
- Complementary Inputs to Utopia
- Astronomical Waste & Conscientious Objection
- Against commentary as a full-time occupation
- A tiny amount of force applied in just the right place
- Pick your contrarianism at the right level of abstraction
- A rapid response to Celeste re E2G vs. direct work
economics of ai
- Why falling labor share ≠ falling employment
- Should AI companies become profit-maximizing firms?
- Endogenous Automation Will Hit You
shows
microstrategy
- Introduce yourself with a call to action
- Do you love yourself or your field?
- Be selectively decorrelated
- Against 'One-Way Streets'
- No taxicab fallacy for knowledge workers
- Beware the Delmore Effect
- I live in Interest-Space
- How to read properly
- list of introspection / life clarification activities
- Provide yourself compelling evidence about yourself
be more public
"table stakes"
Personal strategy
- End-of-December Review + Lydia-2026
- Oxford Equilibrium: A Two-Slope Optimal Timing Problem
- Lydia-2027
- academia & rationality (pt. 1)
- academia & rationality (pt. 2)
- ForeverHaven
- prioritization