3 sentence summary:
Explores perspectives on alignment, ethics, and intelligence, discussing their interplay in areas such as education, society, and superintelligent AI. A weak alternative to orthogonality is proposed and the feasibility or safety of any form or attempt at alignment is questioned. Despite the risks, related fields of study are recognized as imprtant anyway as a form of self knowledge.
3 sentence summary:
The space of machine super intelligences is posited to have a "higher" ethical mean/median/mode vs humans. An algorithm called the "Ethical dispatch algorithm" is put forward as a rough outline of how superhuman ethical decision-making could be achieved - namely by assessing/segmenting the situation, identifying ethical principles, applying those principles, weighing conflicting principles, making a decision, and reflection. Various potential ethical principles are outlined, including Kantian Imperatives, Virtue Ethics, Veil of Ignorance, Inertia & Stability, Golden Rule, Platinum Rule, Principle of Utility, Non-Maleficence, Autonomy, Beneficence, and Justice, which can inform ethical decision-making processes.
wip, tries to show how thinking about significant digits leads naturally into deep intuition about measurement, probability, and transitively, intelligence
https://github.com/openai/automated-interpretability they use an LLM to try to explain what all the neurons in a smaller LLM do via hypthesis testing, and propose several paths towards better results
https://arxiv.org/abs/2304.03442 multi agent conscioussness framework for LLM's with a simulated environment to provide grounding & and an episodic memory implementation
https://www.vincentsitzmann.com/siren/ related to hyena attention alternative. Fast fourier transforms are one of the more impressive computations we can do quickly. If you can transpose a problem into a context where it can be solved with FFT you can go from O(n^2) to O(nlog(n)). Interesting results about resolution and preserving the signal of n'th derivatives when explicitly/directly using neural nets as function approximators.
in theory should fix llm reasoning flaw with things like name switching (alpha reduction), something to look for in training pipeline maybe https://arxiv.org/abs/2305.08298
why can't humans store memories in b trees? How are human memories ordered? is time a natural primary key? It doesn't seem like it. Perhaps how we remember things is one the primary ways in which our biology constrains our mental paths through life. https://arxiv.org/abs/2105.14039
"Visualizing the Loss Landscape of Neural Nets" by Hao Li, Zheng Xu, Gavin Taylor, and Tom Goldstein. This work presents techniques for visualizing the high dimensional loss landscape of neural networks. They showed how different architectures and optimization methods can affect the landscape.
"The Loss Surfaces of Multilayer Networks" by Anna Choromanska, Mikael Henaff, Michael Mathieu, Gerard Ben Arous, and Yann LeCun. This paper provides an analysis of the loss surfaces of multilayer networks, relating neural network structure to that of solid state physics models of a class of materials called spin glasses. After reading that, and also lattice cryptography being the basis for quantum hard cryptosystems, I am just wondering why you can't build a transistor equivalent out of a crystalline lattice.
"Identifying and attacking the saddle point problem in high-dimensional non-convex optimization" by Yann Dauphin, Razvan Pascanu, Caglar Gulcehre, Kyunghyun Cho, Surya Ganguli, and Yoshua Bengio. This paper discusses the role of saddle points in the optimization of deep learning models, arguing that they are more problematic than local minima. (recent perf improvement via "regularization")
"Deep Residual Learning for Image Recognition" by Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Though the primary focus of this paper is the introduction of ResNet, an influential deep learning architecture, it also discusses how residual connections can help optimization by making the loss surface smoother.
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks also see svm which should outperform k-means clustering, at ?slight? performance hit
seems cool&important, looking for good papers on this
could humans learn to evaluate code in their heads if they practiced appropriately? could humans learn to dissassemble binary? what content should we prioritize reading?
- bza Book Pizza: llm read along repl (WIP)
- slay the sql literally just a fork of a (pretty cool, check it out) existing slay the spire clone, hypoagnia, at this point, I think it would be cool to have an sql based card game though. or maybe an sql version of the game "counting kingdom".
https://justinjaffray.com/joins-13-ways/?a=b&ref=upstract.com https://github.com/kenjihiranabe/The-Art-of-Linear-Algebra/blob/main/The-Art-of-Linear-Algebra.pdf