TLDR AI 2024-10-02

Pika has released version 1.5 with more realistic movement, big screen shots, and Pikaffects.

There were lots of great announcements at OpenAI's Dev Day 2024, like vision model fine-tuning, real time API, prompt caching, and model distillation.

Governor Gavin Newsom has vetoed SB 1047, a bill aimed at regulating AI development and imposing safety protocols on high-cost models. Newsom cited concerns that the bill's broad applicability to all large, computation-heavy models was not the best approach for effective AI regulation. Despite the veto, Newsom remains committed to developing AI safeguards, having recently signed multiple related bills and consulting experts for guidance.

Apple continues building its in-house capabilities with a strong multimodal foundation model.

Meta has a great new paper about using a mixture of judges models to properly perform multi task RLHF in post training. It dramatically improves final performance of the models on a number of benchmarks.

This survey examines the honesty of LLMs, a key factor in aligning AI with human values. It reviews challenges such as models confidently presenting wrong answers and difficulties in distinguishing known from unknown knowledge.

Ready to save up to $2,000 on every vacation? Dollar Flight Club has your back—try it for just $1 if you sign up in the next 12 hours. That means; Hawaii from $174, Italy from $291, or France from $295 round trip. Join over 1 million members saving on travel this year. Sign up now!

PerCo (SD) is a new perceptual image compression method based on Stable Diffusion v2.1 aimed at ultra-low bit ranges.

This project highlights the need to evaluate large language models (LLMs) on their combined abilities, not just individual skills. While most models are trained on specialized datasets targeting specific capabilities, real-world tasks often require a blend of expertise across multiple areas, referred to as cross-capabilities.

LexEval is a benchmark designed to evaluate large language models (LLMs) in the legal domain. Given the high demands for accuracy, reliability, and fairness in legal applications, LexEval aims to help assess the strengths and weaknesses of LLMs in legal contexts.

Nvidia did a really great ablation on different ways to include images into a language model. It found that the LlaVa concatenation approach performs the best.

AI is rapidly expanding access to traditionally expensive and inaccessible services like education, healthcare, and personal styling. Generative AI, like ChatGPT, provides affordable, personalized assistance, acting as tutors, healthcare advisors, and stylists, reducing the reliance on costly human professionals. This shift democratizes access, making high-end services available to a wider audience at a fraction of the cost.

OpenAI CEO Sam Altman has proposed a global initiative to build data centers and chip factories to support advanced AI development. Altman initially sought trillions but is now targeting hundreds of billions. The plan involves collaboration with global tech giants and governments but faces regulatory and logistical challenges. Despite initial skepticism, discussions continue with potential expansions in the US, Europe, and Asia to bolster computing power for AI.

While LLMs can boost productivity, they lack critical thinking and self-awareness.

Inbox Zero is an open source AI personal assistant for email.

OpenAI plans to restructure into a for-profit benefit corporation to attract investors - its non-profit will retain a minority stake.