Sheldon Huang
0 followers · 289 views
on the atlas · 15
- Steering GPT-2-XL by adding an activation vector - AI Alignment Forum2 savers
- Does every creative genius need a bitter rival? | Aeon Essays1 savers
- nverse-scaling/prize: A prize for finding tasks that cause large language models to show inverse scaling1 savers
- Hyperbolic discounting - The Decision Lab1 savers
- Anchoring Bias - The Decision Lab1 savers
- Cognitive Biases in Large Language Models - LessWrong1 savers
- Effect of 48 h Fasting on Autonomic Function, Brain Activity, Cognition, and Mood in Amateur Weight Lifters - PMC1 savers
- 2023–24 Google PhD Fellowship1 savers
- 2023-24 Apple Scholars in AI/ML1 savers
- Graduate Courses | Department of Psychology1 savers
- Courses (2022-2023) - Department of Philosophy - University of Toronto1 savers
- Graduate 2022/2023 Timetable — Department of Computer Science, University of Toronto1 savers
- Does HIIT burn belly fat? - The Washington Post1 savers
- Google AI Blog: LaMDA: Towards Safe, Grounded, and High-Quality Dialog Models for Everything1 savers
- Curius / Onboarding2704 savers
highlights · 39
We haven't even tried averaging steering vectors (to wash out extra noise from the choice of steering-prompt), or optimizing the vectors to reduce destructive interference with the rest of the model, or localizing steering vectors to particular heads, or using an SVD to grab feature directions from steering vectors (or from averages of steering vectors).
Steering GPT-2-XL by adding an activation vector - AI Alignment ForumSuppose that this feature is relevant to layer 6's attention layer. In order to detect the presence and magnitude of this feature, the QKV heads will need to linearly read out the presence or absence of this feature. Therefore, (ignoring LayerNorm) if we truncate the residual stream vector to only include the first 70% of dimensions, we'd expect the QKV heads to still be able to detect the presence of this feature.
Steering GPT-2-XL by adding an activation vector - AI Alignment ForumSteering vectors contain important computational work done by later layers.
Steering GPT-2-XL by adding an activation vector - AI Alignment Forumwe have more in common with our rivals than we would like to admit.
Does every creative genius need a bitter rival? | Aeon EssaysDonald Johanson and Richard Leakey
Does every creative genius need a bitter rival? | Aeon Essaysfierce war between Newton and Leibniz, each of whom claimed to be the first to invent calculus – today widely considered to have been developed independently by each of them.
Does every creative genius need a bitter rival? | Aeon EssaysWhen the handsome young Raphael first arrived on the Rome scene and was quickly commissioned by Pope Julius II, Michelangelo labelled him a bitter rival and proceeded to repeatedly accuse him of plagiarism
Does every creative genius need a bitter rival? | Aeon Essaysprimed for rivalry were more open to Machiavellian acts and more likely to exaggerate positive results in a cognitive task.
Does every creative genius need a bitter rival? | Aeon EssaysThomas Edison and Nikola Tesla
Does every creative genius need a bitter rival? | Aeon EssaysIsaac Newton and Gottfried Leibniz
Does every creative genius need a bitter rival? | Aeon EssaysTurner and Constable
Does every creative genius need a bitter rival? | Aeon Essaysthen we would encourage you to perform control experiments demonstrating that the inverse scaling is legitimate. This could look like replacing whatever you expect to cause the inverse scaling with neutral phrasing.
nverse-scaling/prize: A prize for finding tasks that cause large language models to show inverse scalingWe recommend framing your task as a classification or sequence probability task if possible unless you have an especially clear and considered justification for these metrics.
nverse-scaling/prize: A prize for finding tasks that cause large language models to show inverse scaling“Priming” is when someone is exposed to a stimulus (a word, for instance) that influences their response to a second stimulus.
Hyperbolic discounting - The Decision Labis to come up with reasons why that anchor is inappropriate for the situation.
Anchoring Bias - The Decision LabPrepending the description of the Aristotelean syllogism and an assertion of its validity does not improve model performance but instead slightly degrades it11
Cognitive Biases in Large Language Models - LessWrongreversing the order in which options are presented revealed that the output is not at all affected by the content of the options but rather only by the order in which they are presented (Fig. 5c).
Cognitive Biases in Large Language Models - LessWrongfree-form survey format.
Cognitive Biases in Large Language Models - LessWrongLogical fallacies, in contrast, turn out to be very hard to evaluate and raise important questions about the role of logical inference in natural language.
Cognitive Biases in Large Language Models - LessWrongwhether these biases become more or less pronounced with model size, and whether standard debiasing techniques can reduce their impact.
Cognitive Biases in Large Language Models - LessWrongThe proposal should include the direction and any plans for where the applicant’s work is going in addition to a comprehensive description of the research they are pursuing.
2023–24 Google PhD Fellowship48 h fasting resulted in higher parasympathetic activity and decreased resting frontal brain activity, increased anger, and improved prefrontal-cortex-related cognitive functions, such as mental flexibility and set shifting, in amateur weight lifters. In contrast, hippocampus-related cognitive functions were not affected by it.
Effect of 48 h Fasting on Autonomic Function, Brain Activity, Cognition, and Mood in Amateur Weight Lifters - PMCResearch statement covering past work and proposed direction for next 2 years (maximum 5 pages, including citations, in a legible font size), clearly stating the hypothesis and expected contributions to the chosen research area;
2023-24 Apple Scholars in AI/MLResearch Abstract (200 word maximum);
2023-24 Apple Scholars in AI/MLWhat were your responsibilities?
2023–24 Google PhD FellowshipDid you help to resolve an important dispute at your school, church, in your community or an organization? And your leadership role doesn’t necessarily have to be limited to school activities.
2023–24 Google PhD FellowshipInclude any personal, educational and/or professional experiences that have motivated your research interests;
2023–24 Google PhD FellowshipPHL 2198S Advanced Introduction to the Philosophy of Science
Courses (2022-2023) - Department of Philosophy - University of TorontoPHL 2172S Philosophy of Mind: Nature and Content of Thought
Courses (2022-2023) - Department of Philosophy - University of TorontoPHL 2171S Philosophy of Mind: Philosophy of Perception
Courses (2022-2023) - Department of Philosophy - University of TorontoPHL 2199F Seminar in Philosophy of Science: Memory and Imagination
Courses (2022-2023) - Department of Philosophy - University of TorontoPHL 2146Y Topics in Bioethics
Courses (2022-2023) - Department of Philosophy - University of TorontoPHL 2132F Ethics and Speech Acts
Courses (2022-2023) - Department of Philosophy - University of TorontoPHL 2131F Seminar on the Good in Human Lives
Courses (2022-2023) - Department of Philosophy - University of TorontoHPS 4011F Cognitive Technologies: Philosophical Issues and Debates
Courses (2022-2023) - Department of Philosophy - University of TorontoEthical Aspects of Artificial Intelligence
Graduate 2022/2023 Timetable — Department of Computer Science, University of TorontoEthical Aspects of Artificial Intelligence
Graduate 2022/2023 Timetable — Department of Computer Science, University of Torontoavoid any unintended results that create risks of harm for the user, and to avoid reinforcing unfair bias
Google AI Blog: LaMDA: Towards Safe, Grounded, and High-Quality Dialog Models for EverythingStarting [at a young age] he’s read everything that he could find about business.
Curius / Onboarding