The Daily Index — Tuesday, September 12, 2023 · TLDR Index

HEADLINES & LAUNCHES

TLDR AI

TLDR AI

META IS WORKING ON A GPT-4 COMPETITOR

Meta has been snapping up AI training chips and building out data centers in order to create a more powerful new chatbot it hopes will be as sophisticated as OpenAI’s GPT-4.

theverge.com · 1 min

TLDR AI

SPEED UP SPECULATIVE DECODING WITHOUT A DRAFT MODEL

Language model inference is generally slow because these models are heavily memory-bound. So people have introduced speculative decoding using a smaller draft model to “run ahead” and offer suggestions to the large model. This works well, but it’s complicated to implement, and finding a good draft model is tough. Instead, you can create a Medusa model that acts as its own draft by splitting off activations from various internal stages. This speeds up Llama 2x without compromising performance.

together.ai · 8 min

TLDR AI

A BETTER UNDERSTANDING AND REASONING ABOUT PICTURES

The authors of this study have created a way to test how well vision-language models can also think logically about what they see.

arxiv.org · 22 min

TLDR AI

MAKING SELF-DRIVING CARS SMARTER WITH LANGUAGE PROMPTS

This project introduces NuPrompt, a new set of instructions in everyday language that helps self-driving cars understand and track objects around them better. It works really well!

arxiv.org · 15 min

TLDR AI

CAN LLMS REASON?

The debate surrounding large language models concerns whether they possess human-like reasoning abilities or rely on memorization and pattern-matching, with studies showing mixed results. Evaluating LLMs for true abstract reasoning is challenging due to their tendency to excel in tasks related to their training data, necessitating further research for clarification.

aiguide.substack.com · 10 min

TLDR AI

PATTERNS FOR BUILDING LLM-BASED SYSTEMS & PRODUCTS

This article shares seven practical patterns for integrating large language models into systems: measuring performance, adding recent knowledge, fine-tuning for tasks, caching to save time/cost, ensuring quality output, managing errors, and harnessing user feedback.

eugeneyan.com · 66 min

TLDR AI

ANOTHER GROUP OF WRITERS IS SUING OPENAI OVER COPYRIGHT CLAIMS

OpenAI faces a lawsuit from writers alleging illegal use of their copyrighted works to train ChatGPT. They argue that ChatGPT's capabilities to summarize and analyze their content indicate direct training on their works, constituting copyright infringement. This follows similar suits from other authors against OpenAI.

theverge.com · 2 min

TLDR AI

OPENAI CONFIRMS THAT AI WRITING DETECTORS DON’T WORK

OpenAI has confirmed that AI writing detectors are ineffective.

arstechnica.com · 1 min