LLMS - AI Dynamics

Llama 2 Paper Analysis: Supervised Finetuning vs RLHF Evaluation

By

–

31 July 2023 14h51

Finally got a chance to spend some time with the 77-page Llama 2 paper (
https://
arxiv.org/abs/2307.09288). Here are some takeaways at a glance. Appreciate that someone finally did a comprehensive supervised finetuning vs RLHF evaluation! (Lower right)

→ View original post on X — @rasbt

31 July 2023

TGI Inference Project Positioning Debate

By

@julien_c

–

31 July 2023 14h18

don't really agree here:
1/
TGI is/was super early in its history (very few outside contributors, and we discussed with all of them) 2/
it's smaller than other similar inference projects (vllm, …) so unclear it had even "taken the lead" Happy to discuss further though!

→ View original post on X — @julien_c

31 July 2023

InstructGPT and ChatGPT represent significant advancement beyond previous models

By

@rasbt

–

31 July 2023 12h59

I’d say that InstructGPT, ChatGPt, and everything after (which were all released after this paper) were quite a leap forward

→ View original post on X — @rasbt

31 July 2023

GPT Engineer: Automatic AI-Powered Codebase Generation

By

@miketamir

–

30 July 2023 21h01

GitHub – AntonOsika/gpt-engineer: Specify what you want it to build, the AI asks for clarification, and then builds it.
https://bit.ly/3pnaWgS It generates an entire codebase based on a prompt. #AI #MachineLearning #DeepLearning #LLMs #DataScience

→ View original post on X — @miketamir

30 July 2023

AI-Generated Meeting Minutes Raise Questions About Summarizing Vague Discussions

By

@stephanemallard

–

30 July 2023 19h18

Génial de pouvoir faire des comptes rendus de réunions automatisés avec l'IA. Problème : dans les grands groupes, comment l'IA fera-t-elle pour résumer du bullshit sans actions concrètes ? https://
searchenginejournal.com/openai-publish
es-tutorial-for-ai-generated-meeting-minutes/492789/
…

→ Voir le post original sur X — @stephanemallard

30 July 2023

ChatGPT Struggles Generalizing Math to Uncommon Bases

By

@mit_csail

–

30 July 2023 18h00

ChatGPT can do math, but what if it's in base-9? MIT researchers show that while ChatGPT-like models can perform many tasks, they struggle to generalize to uncommon task variants, indicating a lack of underlying abilities: https://
bit.ly/3NZOz9s @zhaofeng_wu

→ View original post on X — @mit_csail

30 July 2023

LLaMA Training and PEFT for Neural Networks Zero to Hero

By

@jeande_d

–

30 July 2023 16h12

LLaMA training & PEFT would be a very appreciated entry in NN Zero to Hero :-). And plays well with recent entries(nanoGPT, State of GPT).

→ View original post on X — @jeande_d

30 July 2023

Neural Networks and Gzip versus Large Language Models Deep Dive

By

@rasbt

–

30 July 2023 14h21

Following up on some discussions of the "NN + Gzip versus LLMs" methods a few weeks back here's my little deep dive with a reimplementation and additional experiments: https://
magazine.sebastianraschka.com/p/large-langua
ge-models-and-nearest/
… This includes fixing the tie-breaking and more!

→ View original post on X — @rasbt

30 July 2023

LLM Hallucination: Document Analysis vs Factual Accuracy

By

@nathanbenaich

–

30 July 2023 13h05

it is quite amazing how LLMs like claude v2 can in one turn give you solid responses about a document i uploaded and in the next turn when i ask about a specific person, the response is entirely made up. –> further to @bxchen recent piece, LLMs are useful when constrained to a

→ View original post on X — @nathanbenaich

30 July 2023