KAN we do it? Responsible AI, Alignment, Adversarial Tuning, and improved models
In the News
🗞 This Week in News
Responsible AI Report from Microsoft - they share their ongoing efforts in Responsible AI, which include the creation of 30 AI tools. This quote in particular is important: “There is no finish line for responsible AI. And while this report doesn’t have all the answers, we are committed to sharing our learnings early and often and engaging in a robust dialogue around responsible AI practices.”
🥁 Interesting Products & Features
If you have used another LLM to evaluate the output of another LLM, this one is for you: Prometheus-Eval 2, an open-source SOTA evaluator language model
Hermes 2 Pro is an upgraded, retrained version of Nous Hermes 2, consisting of an updated and cleaned version of the OpenHermes 2.5 Dataset, as well as a newly introduced Function Calling and JSON Mode dataset. The model demonstrates excellent general task and conversation capabilities but also excels at Function Calling and JSON Structured Outputs. GitHub.
Hugging Face Robotics Library “LeRobot” - provides models, datasets, and tools for real-world robotics in PyTorch. The goal is to lower the barrier to entry to robotics so that everyone can contribute and benefit from sharing datasets and pretrained models.
📄 Interesting Papers
RAG and RAU: A Survey on Retrieval-Augmented Language Model in Natural Language Processing - a summary of SOTA RAG and retrieval-augmented understanding. GitHub. Authors from Stanford.
DOCCI: Descriptions of Connected and Contrasting Images - The DOCCI dataset consists of comprehensive descriptions on 15k images specifically taken with the objective of evaluating T2I and I2T models. The descriptions contain objects and attributes, spatial relationships, text rendering, world knowledge and view/scene in natural language. Authors from Google, UNC, Princeton.
Boosting Segment Anything Model with Adversarial Tuning - This paper introduces “ASAM”: improvements to the SAM with adversarial tuning demonstrates significant improvements across a diverse range of segmentation tasks without additional data or architectural changes. Authors from vivo mobile communications.
LoRA Land: 310 Fine-tuned LLMs that Rival GPT-4, A Technical Report - authors measure the quality of LLMs fine-tuned with quantized low rank adapters across 10 base models and 31 tasks for a total of 310 models. They found that 4-bit LoRA fine-tuned models outperform base models by 34 points and GPT-4 by 10 points on average. Authors from Predibase.
FLAME: Factuality-Aware Alignment for Large Language Models - This paper explores how to make the LLM alignment process more factual. They identify that training an LLM on new knowledge or unfamiliar texts through supervised fine-tuning can actually encourage hallucination and reward functions used in standard reinforcement learning can also encourage hallucination, because it guides the LLM to provide more helpful responses on a diverse set of instructions, often preferring longer and more detailed responses. Authors from Meta AI, University of Waterloo, and Carnegie Mellon.
KAN: Kolmogorov-Arnold Networks - This paper introduces an alternative to the multilayer perceptron (MLP). While MLPs have fixed activation functions on nodes ("neurons"), KANs have learnable activation functions on edges ("weights"). KANs have no linear weights at all -- every weight parameter is replaced by a univariate function parametrized as a spline. The authors show that this simple change makes KANs outperform MLPs in terms of accuracy and interpretability. KANs can be intuitively visualized and can easily interact with human users. Authors from MIT.
🧠 Sources of Inspiration
Best practices for creating software diagrams - a Hacker News thread
How Perplexity builds product - an interview with head of product at Perplexity. A great read for anyone interested in building products (especially with emerging tech).
Modern Advances in Prompt Engineering - a substack on all things prompt engineering.