TLDR Index
TLDR AI

TLDR AI 2023-03-29

6 sections16 entriesinline urlcomplete parse

Parse notes

  • Subscription/account footer truncated before parsefooter_truncated

01GENERAL

02HEADLINES & LAUNCHES

GOOGLE PARTNERS WITH AI STARTUP REPLIT TO TAKE ON MICROSOFT’S GITHUB

Google is teaming up with Replit to combine its AI language models with Replit's Ghostwriter software, which helps programmers write code. This collaboration aims to compete with Microsoft's GitHub Copilot, which uses OpenAI's language-generating AI. Ghostwriter will use Google's AI to improve code suggestions, program completion, and answering developer questions. Replit plans to enhance its software by offering end-to-end support for code writing, testing, deployment, app review, and collaboration. The partnership will also see Replit expand its use of Google's cloud services and distribution of its software.

archive.ph2 min read

OPENFLAMINGO

Last year Deepmind published research on their Flamingo model which is a powerful multi-modal text and vision model. This model could answer questions about images while still being a powerful text model. Laion has now released a framework for training Flamingo models and a 9B parameter checkpoint trained on 5M samples of the Multimodal C4 dataset and 10M samples from LAION-2B image dataset. The model is not usable for commercial applications as it uses the 7B Llama checkpoint from Meta. It also doesn’t yet match Deepmind’s Flamingo on benchmark performance.

laion.ai4 min read

PERPLEXITY AI SERIES A

One of the startup leaders in AI enabled search has launched an iPhone app and raised additional funding. They’re aiming to be the leader in knowledge discovery. We’ll see which of these services find differentiation, but the 2 million users in just a few months is impressive growth.

perplexity.ai1 min read

03RESEARCH & INNOVATION

CHATGPT IS THE NEW MTURK

NLP needs data, this data is usually labeled by crowdsourced workers on various platforms. It turns out, maybe unsurprisingly, that gpt-3.5-turbo is extremely good at following instructions for labeling and often outperforms human annotators on relevance, stance, topic, and other important measures. Even more importantly, it also costs much less (about 20x) than hiring human annotators.

arxiv.org18 min read

EVA-CLIP: IMPROVED TRAINING TECHNIQUES FOR CLIP AT SCALE

The paper introduces EVA-CLIP, a set of models that significantly improve the efficiency and effectiveness of CLIP training. EVA-CLIP achieves superior performance compared to previous models with the same number of parameters but smaller training costs, and the complete suite is released for open research.

arxiv.org7 min read

UNRAVELING IMAGEN: ZERO-SHOT CLASSIFICATION SHOWDOWN

Imagen, a text-to-image diffusion model, rivals CLIP in zero-shot classification, outperforming in shape/texture bias tests and suggesting generative pre-training as a valuable approach in vision-language tasks.

arxiv.org12 min read

04ENGINEERING & RESOURCES

SYNTHESIS

This paper discusses how text-to-image diffusion models, like DreamBooth, can generate realistic images from simple text inputs, but can also be misused to produce fake news or disturbing content targeting individuals. The paper proposes a defense system called Anti-DreamBooth that adds subtle noise perturbations to user images to disrupt the generation quality of any DreamBooth model trained on them, and evaluates its effectiveness against a wide range of algorithms on various text-to-image model versions.

github.comGitHub repository

AIX

AIx is a CLI tool to interact with LLM APIs.

github.comGitHub repository

05MISCELLANEOUS