"All else being equal, being able to squeeze more predictive juice out of flawed theories and inadequate paradigms will help them to stick around for longer, impeding true scientific progress."

Give it a shot and let me know if you run into any issues. It should come in handy when you want your agent to retrieve literature for you ("What articles do I have on LLMs and computational social sciences?"), or for full-blown RAG systems, where access to your library may improve generated answers

It doesn't matter if people use GPT, Llama, or Gemini, or if their chatbots lose money. It doesn't matter if Deepseek's architecture reduces compute requirements, as the cost decrease will likely increase demand. To integrate LLMs at scale you need compute, and the money is in providing it. 10/10

So the large players like AWS, Microsoft etc. don't invest into these data centers to push the development of AI. They're smarter than that. Their play is that with the proliferation of LLMs across businesses (and again, applications are aplenty), the demand for compute will rise. 9/10

When deployed at large, however, even the smaller models require compute often unattainable for all but the largest providers. Due to economics of scale, only large data centers can provide at a price point profitable for both provider and (third-party) user. Hence, the need for data centers. 8/10

However, there's not just the proprietary models, but also a great open source LLM ecosystem. You could argue that the investments into proprietary models will not pay off, as there are equally good, open source solutions (think windows/linux). This remains to be seen. 7/10

The other day, somebody in the industry compared LLMs to operating systems, and I think this is a very good way to think about it. Operating systems are only viable at scale. And they're a means to an end, not the endpoint. They're only there to enable other tasks. Adaptation pays off at scale. 6/10

You could even argue that ChatGPT and other chatbots are merely there to showcase the model's capabilities. They're there to show a large user base what these models can do in terms of language understanding. It does not matter if the companies lose money on them - they're a proof of concept. 5/10

Most of these applications are not chat bots though. They're all types of tasks where you'd like your computer to be better at understanding natural language - and there's a lot of those. So part of the money is in third parties offering their clients "AI" applications for their businesses. 4/10

LLMs excell at language understanding, more so than any other technology available. And this capability can be used in a plethora of tasks - from information retrieval to human-computer-interaction to code completion. This is evidenced most strongly by third parties offering LLM integrations. 3/10

First of all, I agree with a lot of the issues Ed raises. "AI" is a marketing term. "AGI" even more so - and it can not happen within the framework of large language models. And yes, chatbots severly underdeliver in the majority of tasks they're advertised for. But that is besides the point. 2/10

At least the author answers their own question of "fragmented or shallow reviewer feedback". It's reviewer 2 - now with 2 pages of AI slop to give it the semblance of meaningful feedback

Thanks to everyone who participated - as always, I learned a lot preparing the workshop and from all the engagement and smart questions!