Claude as a Research Partner
Google Deepmind
I ran across this competition on Hackernews I think, and I found it fascinating because it is getting at the core of some ideas I've been having. Knowing that the research is active is very exciting for me.
Some nostalgia
I failed too many CS classes in college (shout out Carleton) to complete the degree in 4 years, so I switched to Linguistics as my major at the beginning of my Junior year. This was one of the best decisions I think I made for my sanity, and intellectual development at the time. I had just taken my first Linguistics class during the Spring Semester of my Sophomore year. I was introduced to some of the most interesting ideas I had ever read about in that short survey that was the introduction course.
What really ended up interesting me in the end in Linguistics was Phonology and Morphology, and then their combination in the subfield of Morphophonology. Syntax was fun because being able to break sentences down into their core components felt very rote to me, I was able to apply some natural algorithm my brain had developed with English being my second language, but was my primary language from the age of 7 onwards. I was not aware of those algorithms until I started to just break things down and finally understand all those things that they were trying to explain to me in descriptive grammar lessons. But descriptive linguistics is where it really clicked for me. I couldn't follow a bunch of grammar rules to determine if a sentence was grammatical. Hell most of my personal writing is an abomination of run-on after run-on after run-on, but I don't particularly care because that is my style of speaking, and that is the way I like others to experience my train of thought while I am writing.
Claude as a research assistant
Well I found this Kaggle competition really enticing the other night, and I started to explore a phenomenon one of my classmates presented about. The thing that I found most interesting about the subject was the difference in performance of the descriptive application of phonological rules to derive the correct compounding of words.
As I was working with Claude to think through what this might end up looking like as a benchmark that can be handed off to Kaggle, I was getting such an exhilarating experience.
When I engineer with colleagues, or just in general when I have conversations with people about something they are passionate about or highly knowledgeable about I ALWAYS am asking questions. For me the Socratic Method is the fundamental way that I learn. If I see something, I become critical of it, I use prior knowledge to question any assumptions in what were made for something to be presented to me. I just want to know more, you cannot feed me enough knowledge.
So as I was working with Claude every single thing I knew enough to be dangerous about, enough that I had been exposed to the concept in my past education, I started quizzing it on specific terms: why is it relevant to this problem? What happens if we do this? How could we potentially avoid hinting a model too strongly to using data in its training set?
These were probing questions to my initial thesis that Claude would have to defend over and over again. This was the first time Claude was critical of my thinking, and I was critical of its thinking going back and forth.
The ability for me to directly learn and have access to all the world's knowledge in this method is something I've taken for granted because most of my LLM interactions have been for code generation, a problem I honestly find significantly less interesting, I want to know how these LLMs truly work, and that I think I have a novel idea of how to measure something innate in human knowledge and in our universal grammar mechanisms.
Would it not be amazing to have an AGI that is able to describe Universal Grammar to us, and make predictions based on what the UG can imply!?
When I was a kid the thing I wanted the most was to be able to talk and communicate with my computer by talking to it. I learned a lot about computers as a result, and then I fell in love with Linguistics.
Now I want to build the best version of these things, why? Because science is really what it comes down to for me. Anyways having a lot of fun.