I'm ready to believe fine-tunes can do useful things with new facts from their fine-tuning data… but I need to first interact with a working demo to get there!
LLMS
-
Selective Model Weight Updates for Production AI Systems
By
–
Definitely a promising direction, but it doesn't look to me like that technique is useful for production scenarios yet Being able to selectively update model weights to tell a model "the new CEO of Fiat is X" would indeed be incredibly useful if it was fully reliable
-
Gorilla: Fine-tuning LLM capabilities for improved function calling
By
–
This Gorilla? That looks like fine-tuning for improved capabilities like function calling, not for baking in extra knowledge https://
gorilla.cs.berkeley.edu -
Fine-tuning LLMs with RAG for personalized email generation
By
–
Thanks, that's a great example "Panza produces a fine-tuned LLM that matches your writing style, pairing it with a Retrieval-Augmented Generation (RAG) component which helps it produce relevant emails." So fine-tuning for writing style, RAG for knowledge
-
PhD Research on Improving LLMs Zero-Shot Task Adaptation
By
–
Happening today! PhD researcher Nihal V. Nayak will explain his research into how to improve large language models’ zero-shot task adaptation on users’ specialized, private data faster than ever. Sign up here: https://
buff.ly/3JFtOhR -
AI language models may predict multiple tokens at once
By
–
The future of AI language models may lie in predicting beyond the next word: Multi-Token Prediction Studies suggest that the human brain predicts multiple words at once when understanding language, utilizing both semantic and syntactic information for broader predictions – now
-
Long Context Evaluation: Jamba vs Mixtral Performance Comparison
By
–
Because we focused on long context, we took the L-eval versions of the datasets and also turned them to 3-shot format to challenge with longer context handling. Mixtral's performance was measured in the same setting as Jamba.
-
Input vs Output Token Pricing for LLMs
By
–
Are those prices for input or output tokens? I think they're for output tokens, but usually I find the input tokens price is more relevant – especially since most models with long context have a much lower cap on output tokens (eg Claude 3 allows 200k input but only 4k output)
-
RAG vs Fine-tuning: Understanding AI Model Enhancement Techniques
By
–
"Continue will use similarity search to automatically find important sections when answering your question" sounds like an implementation of RAG, not fine-tuning
-
Gemini Pro 1.5 Long-Context Performance and Pricing Tradeoffs
By
–
The one downside I've seen from super long-context models like that is performance – my experiments with Gemini Pro 1.5 really long prompts has been that they can take 30-90s or more to return a response Also price: input is $7/1m tokens!