TLDR AI 2024-06-05

Sam Altman’s Investment Empire πŸš€, Jailbreaking of LLMs πŸ”¨, Tree Diffusion 🌲

Headlines & Launches

A Right To Warn About Advanced AI (3 minute read)

A group of current and former AI employees is calling for advanced AI companies to commit to principles that ensure transparency and protection for employees who raise risk-related concerns. They highlight the need for companies to avoid enforcing non-disparagement agreements, facilitate anonymous reporting processes, support open criticism, and prevent retaliation against whistleblowers.

The Opaque Investment Empire Making OpenAI's Sam Altman Rich (15 minute read)

Sam Altman is one of Silicon Valley's most prolific and aggressive individual investors. He manages an investment empire with holdings worth at least $2.8 billion as of early this year. Much of the portfolio isn't widely known. This article takes readers through what's known about Altman's investments.

Will Scaling Solve Robotics? (15 minute read)

Over 900 people attended last year's Conference on Robot Learning, which featured 11 workshops and almost 200 accepted papers. One of the largest debates at the event was whether training a large neural network on a very large data set was a feasible way to solve robotics. This post presents the different sides of the argument to deepen people's understanding of the debate. Scaling has worked for other similar fields. However, it is impractical as there isn't much robotics data available and there's no clear way to get it. Even if scaling works as well as in other fields, it likely still won't solve robotics.
Research & Innovation

MMLU Pro (26 minute read)

MMLU is a common benchmark for reasoning tasks. It is often considered to be both the gold standard and something models have overfit. MMLU Pro is a new, harder, and cleaner benchmark for measuring language model reasoning.

Tree Diffusion: Diffusion Models For Code (18 minute read)

Fantastic diffusion paper that diffuses code for images. It can directly make edits as part of the diffusion process. It is slow, but can be combined easily with search to dramatically improve reasoning ability.

Jailbreaking of Large Language Models (16 minute read)

Researchers have introduced improved methods for optimization-based jailbreaking of large language models, building on the Greedy Coordinate Gradient (GCG) attack.
Engineering & Resources

Image Compression (GitHub Repo)

Control-GIC is a new framework for generative image compression that allows fine-grained bitrate adjustment while maintaining high-quality results.

Omost Image Synthesis (GitHub Repo)

From the creator of ControlNet, Omost is a way to gain control over your image generation. It first rewrites prompts as a set of descriptive code. Then it uses that to render the final image. Importantly, you can edit the code before or after generation to slightly change the model output.

Improved Video Super Resolution (4 minute read)

Researchers have developed a training-free video interpolation method for generative video diffusion models. This new approach, compatible with various models, enhances frame rates without the need for extensive training or large datasets.

LLM inference speed of light (13 minute read)

Using theoretical speed of light modeling as grounding is really important for problems where the amount of computation and memory access is known a priori as it helps validate the quality of implementations and predict the impact of architectural changes.

What I learned from looking at 900 most popular open source AI tools (13 minute read)

This review of open source AI repositories aims to give readers a big-picture view of the seemingly overwhelming AI ecosystem.

Plentiful, high-paying jobs in the age of AI (20 minute read)

It's possible that many of the jobs that humans do today will continue to be done by humans indefinitely, no matter how much better AIs are at those jobs, due to comparative advantage.
Quick Links

Even The Raspberry Pi Is Getting In On AI (1 minute read)

Raspberry Pi is launching an AI chip integrated with its camera software that will allow AI applications like chatbots to run natively on the microcomputer.

Facia (Product)

Prevent fraud and spoofing attacks with advanced facial recognition technology.

Using AI To Decode Doc Vocalizations (2 minute read)

University of Michigan researchers, in collaboration with Mexico's INAOE, have developed AI tools that analyze dog barks to determine playfulness or aggression, as well as identify the dog's breed, age, and sex.
The most important AI, ML, and data science news in a free daily email.
Join 500,000 readers for