TLDR AI 2023-11-06

xAI trained an autoregressive language model at 34B params. It is quite performant and will power an AI system built into the X platform.

OpenAI Dev Day is today and it is rumored that we can expect enhancements to OpenAI’s developer tools, new pricing plans for ChatGPT, and a peek at Gizmo V8, a new and improved version of the ChatGPT iOS app.

Google is extending its Vulnerability Rewards Program (VRP) to include generative AI, encouraging research into AI safety and security. Simultaneously, it's expanding its open-source work to ensure better AI supply chain safety. It is also leveraging the Secure AI Framework (SAIF) and the Google Open Source Security Team (GOSST) to protect AI supply chains from threats like model tampering and data poisoning.

Paper digest generated summaries of many papers from NeurIPS 2023. Interestingly, much of this work is almost a year old at this point and has already been readily adopted by the community.

If you fine-tune three task-specific models from the Mistral base, one for generation, one for critique, and a final one for regeneration then the entire system shows dramatically improved generation performance. Synthetic data is used to tune these models.

Sketching a rough outline of what you want a robot to accomplish is a novel form of communication and turns out to work surprisingly well for standard pick and place tasks.

Video2Music’s AI framework generates emotionally congruent music for videos, leveraging a unique dataset and transformer model validated by user studies.

Qdrant uses a lightning-fast vector database to store embeddings.

Giskard is a Python library that automatically detects vulnerabilities of AI models, from tabular models to LLM, including: performance biases, data leakage, spurious correlation, hallucination, toxicity, security issues and many more. Get your model to production quickly and safely.

Install Giskard (via GitHub) or try it in Colab

This is Deepspeed's new software for competing with vLLM and text-generation interfaces for the fast serving of language models. It comes with many state-of-the-art acceleration features and preliminary results suggest it's up to 2.4x faster.

Poisoning the data well and other data pipeline attacks are a huge challenge for the cyber security community and often a blindspot for most AI organizations. This post outlines (with code) what the attacks are to help with future mitigation.

Cruise, GM's driverless car company, frequently relies on remote operators for its vehicles, challenging the company's claims of autonomy. The revelation raises concerns about the true capabilities of self-driving cars and prompts calls for transparency and investigation into the safety and autonomy of such vehicles.

A great deep dive into some novel vulnerabilities that come from generative AI plugins. In this case, the attack resembles SQL injection but came via an infected Google Doc. While Google has fixed the issue, the security researcher was able to exfiltrate other users’ prompts.

A collaborative, LLM-infused cloud environment designed for data analysis and storage of engineering materials. With built-in AI Agents, Dystr automates workloads, freeing engineers to focus on core tasks. Join the ranks of engineers at Google, Valve, Meta, & SpaceX. Start with free credits!

Layer cuts your research time by 40% with AI-powered summarization, information extraction, and more.

Google's program allows early career postdoc-level researchers to join a team with resources and continue their research agenda.

The latest update on ChatGPT Plus introduces fresh beta functionalities, including the ability to work with PDFs. Users can upload and interact with documents alongside receiving multimodal assistance.