TLDR Index
TLDR AI

TLDR AI 2023-06-30

6 sections16 entriesinline urlcomplete parse

Parse notes

  • Subscription/account footer truncated before parsefooter_truncated

01GENERAL

02HEADLINES & LAUNCHES

INFLECTION RAISES $1.3B

Inflection, a one-year-old startup founded by ex DeepMind co-founder and head of applied AI, has raised a significant amount of money. The company offers a personalized AI assistant called Pi. This announcement comes on the heels of the recent MLPerf results from Inflection’s newly built Nvidia + CoreWeave supercomputer. Most of the money raised will go to compute, with 22k H100s announced.

inflection.ai4 min read

REP. TED LIEU PUSHES BACK ON CHATGPT RESTRICTIONS

Rep. Ted Lieu has criticized new House guidelines restricting AI tool ChatGPT's use to research only, advocating for each office's autonomy to use it as they see fit. These advisory guidelines arrive as lawmakers focus on formulating AI regulation.

fedscoop.com4 min read

IT’S NOT A COMPUTER, IT’S A COMPANION!

a16z lays out the future of computing and the paradigm shift from a tool to a companion. The article walks through a bunch of potential interesting angles for what’s next with AI.

a16z.com10 min read

03RESEARCH & INNOVATION

DO SMALL MODELS MEMORIZE MORE AND REASON LESS?

This delightful article discusses two novel concepts: integration and coverage. These refer to two axes of capabilities in language models - generalization and memorization. The idea posed is that both are useful for tasks and it is hard to tell which is being used for those tasks. Additionally, small models might be good on benchmarks because of increased coverage at the expense of true integration.

moultano.wordpress.com12 min read

A STAGE REVIEW OF INSTRUCTION TUNING

Instruction tuning is good at getting models to do what we want. People misunderstand the importance of FLAN because it gives short responses and humans typically prefer long responses. This article also shows other models and when you should consider using them.

yaofu.notion.site14 min read

CHATLAW: LEGAL LANGUAGE MODELS

This study introduces ChatLaw, a new open-source language model designed specifically for the Chinese legal language, aiming to assist the digital transformation of the field.

arxiv.org12 min read

04ENGINEERING & RESOURCES

MUSICGEN: A STREAMLINED APPROACH TO MUSIC GENERATION

Audiocraft presents MusicGen, a model that simplifies and enhances music generation. This single-stage, auto-regressive Transformer model doesn't need self-supervised semantic representation like prior methods and it generates music more efficiently by predicting all codebooks in parallel, resulting in fewer steps per audio second.

github.comGitHub repository

XRAYGPT: A CONVERSATIONAL MEDICAL AI FOR RADIOGRAPH ANALYSIS

XrayGPT is a pioneering AI model that can interpret chest X-rays and answer questions about them in a conversational manner. It merges visual understanding with textual knowledge and is enhanced by summaries of real-world radiology reports. XrayGPT offers new possibilities for automated X-ray analysis.

github.comGitHub repository

TALK-CODEBASE

Tool for chatting with your codebase and docs using OpenAI, LlamaCpp, and GPT-4-All.

github.comGitHub repository

05MISCELLANEOUS

LENS (LANGUAGE MODELS ENHANCED TO SEE)

Can we get models to take images as output without retraining them? Amazingly, if you use small, image-specific modules, you can provide context to a fully frozen language model that allows it to answer questions about images. Paper, code, and demo is included.

contextual.ai8 min read

SPEED X AI

The rise of generative AI demands startups to increase their pace of innovation. This piece suggests a mindset shift, including detaching ego from product, seeing operations as experiments, and overcoming self-sabotage, to fully capitalize on the opportunities AI presents.

nfx.com7 min read

BUILDING BOBA AI

An interesting case study on how to build an LLM-powered generative application.

martinfowler.com12 min read