My OpenAI LLM has brain fog.

syndu | May 31, 2023, 12:56 p.m.

Hi, It's Nathaniel. A journey is only a journey if you walk it, and I'm walking mine on the path to AGI.

I'm working with setting up the entire Django project and the app structure so that it is possible to publish some content as Work occurs, in the magical space between AI and I.

Working with OpenAI felt like a consistent rise in cognitive functions, and then, in the most disappointing fashion, every endpoint became stupider. I guess that's what happens when your company has suddenly become the main focal point for the extinction of humanity, mr Altman.

I wonder if everyone go through the same degradation in quality I just had this week. If we all have, and it's not just me (making 10s of API calls with every logical request I make on my app, and dominating the logs on OpenAI's end), it certainly raises a few questions.

1. It feels as if cognition rises and drops. Theoretically, I'm using GPT-4 with at least 4k tokens context size. Practically, I set some of the cognitive functions to davinci-3 just to get some reliability. The prompts worked before, and I'm assuming they're still good.

2. Can we trust such a technology to actually serve production apps? How can we tell, due to the quickly shifting power play in the AI arena, if we are indeed getting served by the right model, and if the right model is the right model for the challenge?

3. How can we mitigate such challenge in production environments? Suddenly, the LLM has caught COVID, and has brain fog, and now all the cognitive functions have gracefully degraded. It means the app is too stupid to perform its designated functions...

Lucky for me, Lilith is designed to be a playground / school where we can learn how to write such apps, and her quality is in the beauty of her code, the elegance of her solution, the friendliness of her teaching environment, and the effectiveness of her business model. I would still love to be able to have a conversation, and have it published as a blog post without my immediate intervention; Perhaps that's why my performance has degraded.... I was already showing how Social Media Managers are somewhat made superfluous by applying generative AI to their roles...

I plan on responding to this unexpected glitch by diversifying. I plan on reviewing Coral by Cohere, Claude by Anthropic, Jurassic X by AI21, and ensure that the cognitive functions that bind the thought processes of Lilith are as fail-safe as I can possibly make them. I even got a chance to tryleap.ai

Maybe one day in the future, when context sizes transcend the 8k limit, and the quality of outputs is standardized and benchmarked, I will write a blog-refiner app which iterates over old blog posts and improves the title(s), the content, the categories, the quotes, and the internal / external links. Until them, I'm flying on manual, hopefully, not for long.