Anshul Kundaje#1650
I wrote about AI in academia. "PhD-level thinking", LLM bias, grunt work, alignment, AGI, data center water use, AI politics -- something for everyone.
Many users praised the academic analysis of AI bias, alignment, and politics as an enjoyable and insightful read, while some felt it overlooks PhDs without funding.
Link:
https://open.substack.com/pub/theinfinitesimal/p/thoughts-on-ai-in-academia?r=43f9ax&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true
@kareem_carr ... even when they are given the solutions.
The *are* capable of exploring a parameter space though (e.g. I recently asked Opus to implement a toy example of a transformer running on a biological task and, after initial failure, it spent several hours converging on a ...
Based on my very recent experience as a PhD student, I could never really get away with just saying, “Here’s the answer.”
The output of an investigation was much more about the robustness around the answer: being able to defend my choices, show the analyses I’d done to validate the solution both directly and indirectly, present the options I explored that didn’t work, and demonstrate that I understood the landscape of possible approaches, not just the one I picked.
Is that something these AI models are actually able to do?
I understand the article to be saying that they can’t build cumulatively and seem to start from scratch each time. But my concern is slightly different. Even the outputs themselves seem shallow to me, like a plant with one main taproot but none of the nested secondary roots you need to create mature work.
@kareem_carr This is a good question. I think it's true that the models do a poor job of evaluating the landscape of possible approaches. They seem to land on solutions fairly arbitrarily (even though they implement them well) and aren't good at discriminating among multiple paths ...
@SashaGusevPosts this was really good. :) You might want to have a look at my piece linked here...I think you'll like it too!
@SashaGusevPosts was an enjoying read , also kind of confirms that models kinda look for shortcuts or they settle or something like a local minima
@kareem_carr ... simulation set-up and model that worked). So I think an open question is whether that parameter search skill can eventually translate into an ability to discriminate between higher-level research directions.
@SashaGusevPosts something for everyone except the phd who can't get funding
@SashaGusevPosts Interesting 🙏
@profgoose Thanks, will read!
Anshul Kundaje#1650I wrote about AI in academia. "PhD-level thinking", LLM bias, grunt work, alignment, AGI, data center water use, AI politics -- something for everyone.