Trying To Create A Theory of Mind: Commentary On Apollo Research's Paper On In-Context Scheming

It is not enough to know what text an LLM would generate, but why the LLM would generate that text. That is because software development is moving towards a world where the LLM’s text is being used in more and more situations. LLMs are not only generating useful content, but can also act as agents when given access to external tools. If we can understand why an LLM generates text in a certain way, then we can predict what text it may generate next, which would allow us to develop policies that can take advantage of said predictions. This “theory of mind” could help us control LLMs, limiting risks that might arise from them.

Navigating Human Disempowerment: Ibn Khaldun's Insights On The Nature of Groups

This essay was written for BlueDot Impact’s AI Governance course. I was part of the April 2024 cohort. The title of this essay was generated by GPT-4o and edited by Tariq Ali.

Statistical Approaches Are Actually Pretty Good: A "Mea Culpa" Regarding One Of My Predictions About Text Generation

For some time, I have written blog posts on dev.to about text generation, and even went so far as to collect them into an online ebook. In one of the chapters of said e-book, written in 2016, I made predictions about the future of text generation. One of those predictions was “Statistical Approaches Will Be Part Of A ‘Toolkit’ In Text Generation”.

Deploying GitHub Repos onto repl.it: Easier Than Expected

repl.it is a very useful tool for distributing code snippets for other people to use.

Comparing the Pseudocodes of Cuckoo Search and Evolution Strategies: My First Foray into Understanding Metaheuristics

EDITOR’S NOTE: As a fair warning, half-way through writing this blog post, I realized this whole endevour of comparing the pseudocodes of two different metaheuristics might not have been such a great idea (and may have indeed been a clever act of procrastination). I still posted it online because it may still be useful for other people though, and because I’m somewhat interested in metaheuristics. I also kept my “conclusion” footnote where I came to my realization about exactly how much time I wasted.