TLDR Index
TLDR AI

TLDR AI 2024-02-22

6 sections20 entrieslinks blockcomplete parse

Parse notes

  • Subscription/account footer truncated before parsefooter_truncated

01GENERAL

02HEADLINES & LAUNCHES

GOOGLE DEEPMIND OPEN SOURCES GEMMA BASED ON GEMINI

Google has released the weights for its Gemma 2B and 7B parameter models, available via HuggingFace. The models are decoder-only Transformers trained on 2T and 6T tokens. They substantially outperform Llama 2 on a wide range of benchmarks and come in base and instruction-tuned versions.

blog.google3 min read

HELM INSTRUCT BENCHMARK

The language modeling group at Stanford released a widely used benchmark called Holistic Evaluation of Language Models (HELM). They also released a version for instruction following called HELM-Instruct. It is multidimensional, open-ended, and absolute.

crfm.stanford.edu21 min read

03RESEARCH & INNOVATION

25+ LORAS FOR MISTRAL THAT OUTPERFORM GPT-4

Predibase released 25 narrow task-specific LoRAs that each outperform GPT-4 on specific tasks. These task-specific LoRAs are built onto the Mistral 7B model. This is another piece of evidence to support the ability of local models to perform exceptionally well on exact tasks.

predibase.com8 min read

COMBINING UNSUPERVISED TRAINING AND SUPERVISED FINE-TUNING

This repo challenges the effectiveness of combining supervised fine-tuning and reinforcement learning in training. Using a more advanced model like GPT-4 for initial training can outperform the more complex two-step method.

github.comGitHub repository

04ENGINEERING & RESOURCES

MULTIMODAL CONVERSATIONS DATASET

Meta has released the MMCSG dataset, which has thousands of two-sided conversations recorded via the Aria glasses, to encourage research in various social audio and visual tasks.

ai.meta.com6 min read

MULTILORA INFERENCE SERVER

Lorax is an inference server that supports hot-swapping LoRAs onto a single base model. This means a dramatic decrease in RAM while supporting a wide range of model tunes.

github.comGitHub repository

EVALUATING LLMS IN GAMES

This project introduces GTBench, a platform for testing LLMs like GPT-4 in various game scenarios to discover their strengths and weaknesses in strategic reasoning. While these models show promise in games based on probability, they struggle in deterministic games and vary in performance based on their design and training.

github.comGitHub repository

05MISCELLANEOUS

MULTIVIEW DIFFUSION++

A diffusion model that can generate multiple views of an object and then stitch them together into a compelling 3D version.

mvdiffusion-plusplus.github.io5 min read

THIS IS WHY THE IDEA THAT AI WILL JUST AUGMENT JOBS, NEVER REPLACE THEM, IS A LIE

AI is going to automate parts of the workforce. The reaction so far has been two-sided – will this efficiency mean the same humans can do more or will fewer humans do the same tasks? This article explores comparisons between the impact of tech on manufacturing and agriculture and what we’re seeing today with the modern knowledge worker.

donaldclarkplanb.blogspot.com5 min read

MULTIMODAL DIALOGUE

ChatterBox is a vision-language model that excels in multimodal dialogues, specifically in the newly established task of multimodal multi-round referring and grounding.

github.comGitHub repository