Is the theory important for leveraging "AI technologies"? How much should one understand those technologies, for example how the GPT models, mixture of experts, diffusion networks… work? Some clarifications: Some parts are useful (which? Let me know in the thread!)
LLMS
-
Mixtral 8x7b and Gemini Pro in the Hugging Face Arena
By
–
You can now run Mixtral 8x7b and Gemini Pro in the Hugging Face ChatBot Arena. Interestingly, Mixtral 8x7b is currently ahead of Gemini Pro in the Chatbot Arena leaderboard and tied with GPT-3.5 Turbo – but the differences are small.
-
AI Breakfast for Monday, December 18th
By
–
AI Breakfast for Monday, December 18th Running Mixtral 8x7b, OpenAI's official Prompting Guide, and how Google is being sued for essentially killing their own business model with AI:
-
Cohere RAG Chatbot with Quickstart Connectors Guide
By
–
Cohere Chat with RAG comes with quickstart connectors – around 100 pre-built connector implementations, including Slack, Elastic, GitHub, Google Drive, and more. In this LLM University chapter, learn about RAG with quickstart connectors. https://
txt.cohere.com/rag-chatbot-qu
ickstart/
… -
GPT Model Hallucinations and Token-Level Confusion Explained
By
–
It’s a hallucination, likely because at the token level, gpt-3.5-turbo and gpt-4.5-turbo would be represented similarly and the model already knows it has gpt-4 access in some cases. Plus we called the new model gpt-4 turbo which has added to the confusion.
-
ChatGPT Update and New Custom Actions
By
–
There was an update on ChatGPT itself. Not 100% sure it is available everywhere but I assume so. You should notice it after opening custom actions page.
-
GPT-5 Will Surpass Doctors in Medical Expertise Soon
By
–
One think we shouldn’t forget is how GPT 5 and more will be in a year or so much better than any doctor.
-
Predicting Uncertainty in Discrete Token Probability Distributions
By
–
A prediction problem in a discrete and finite domain is easy because we know how to represent uncertainty in the prediction.
If there are 30,000 possible tokens, we output a 30k-dimensional vector of values between 0 and 1 that sum to 1 (thank you softmax).
We can't output a -
GPT’s Apparent Intelligence Improvement Questioned as Marketing
By
–
Don’t know man… GPT seems really much smarter than before. I guess it’s all a PR stunt.
-
AI Model Training: Logit Bias vs. RLHF-Flattened Logprobs
By
–
Not seeing why that's a problem w/ <= 5 classes since you can logit bias. The real issue is whether there's enough information left in the RLHF-flattened logprobs