TLDR Index
TLDR AI

TLDR AI 2023-05-16

6 sections15 entriesinline urlcomplete parse

Parse notes

  • Subscription/account footer truncated before parsefooter_truncated

01GENERAL

02HEADLINES & LAUNCHES

TOGETHER COMPUTE $20M SEED ROUND

Together compute is, among other things, building a cloud platform for efficient machine learning. With strong academic connections, and highly talented technologists, they have a good shot of building really useful tools for the ecosystem. One of their notable works to date is the release of the red pajama data set, which is an open replication of the closed data set used to train Llama.

together.xyz2 min read

03RESEARCH & INNOVATION

MEGABYTE - MILLION BYTE SEQUENCES

Transformers are not actually end to end. There is a separate training process for the tokenizer which is strange and often leads to odd performance in general. However, if we try to naively train on bytes we quickly run out of context length given the increased length of sequences. Additionally for truly multimodal problems training directly on bytes removes complex patching and tokenization schemes. This work allows models to train directly on bytes with sequence lengths up to a million bytes. Maybe this will even help nucleus sampling for rare words too!

arxiv.org28 min read

VIDEO QUESTION ANSWERING

Visual question answering has seen massive adoption over the past several months, and rapid progress. Largely enabled by pre-trained models such as Blip and clip. Applying these models to video is challenging due to the dramatically increased computational requirements. This work introduces a cascade of different models that dramatically improves the robustness of these systems by operating at different timescales. They use Blip2 and find strong performance.

arxiv.org45 min read

04ENGINEERING & RESOURCES

MICROSOFT GUIDANCE

Programs are how we talk to computers, prompts are how we talk to AI, and AI is now talking to computers. This means that we need a better way to talk with AI. Guidance is essentially a prompting language and toolkit developed by Microsoft that allows you to set guidelines and instructions in a way that encourages reproducible and robust prompts.

github.comGitHub repository

ENHANCING 3D SCENE RECONSTRUCTION AND REALISM FROM LASER DATA

NFL is a new method that takes LiDAR data (a way of measuring distances using lasers) and creates realistic 3D scenes from new viewpoints. This technique works better than others and can help improve tasks like mapping and understanding the environment around us.

research.nvidia.com5 min read

05MISCELLANEOUS

UNCENSORED MODELS

Eric Hartford's blog post explores the implications of uncensored AI models, suggesting that AI censorship might be leading us down a risky path. He argues for the development of systems to better manage the outputs rather than suppressing them. He mentions that more research and dialogue are needed around the ethics of AI use, and proposes the idea of a multi-stakeholder governance model for AI systems.

erichartford.com10 min read