TLDR AI 2024-05-20

While some are worried about the use of AI in cyberattacks, Orca Security is using it to build better defenses. This blog post shows how AI-powered security enables companies to:

• Navigate complex cloud environments by generating specific remediation instructions and code to mitigate cloud risks.
• Stay ahead of attackers with AI-powered search that helps organizations understand which assets are vulnerable within minutes of a vulnerability being found.
• Tackle industry skills shortages by simplifying tasks, lowering skill thresholds, and suggesting optimal configurations.

Learn more about AI in cybersecurity 👇

→ Read the full article on the Orca blog

→ See how you can simplify investigations and accelerate remediation with Orca's built-in generative AI

OpenAI dissolved its superalignment team after the departures of co-leads Ilya Sutskever and Jan Leike and several other researchers. This team, created to address long-term AI risks, will have its work absorbed by other research efforts.

Unitree has released the first few details on its G1 Humanoid Agent, a $16,000 robotic helper. The robot is equipped with 3D LiDAR sensors, a depth camera, three-fingered grippers, and a 9,000-mAh battery pack. It can recover from brutal kicks and punches, as seen in the short demo video in the article. At around 35 kg, the robot can fold itself down to 690x450x300-mm proportions for compact carry.

xAI has released the weights and architecture for Grok-1. Grok-1 is a Mixture-of-Experts model with 314 billion parameters. The released model is a raw base model checkpoint from the Grok-1 pretraining phase, which means that it is not fine-tuned for any specific application. It was trained from scratch by xAI using a custom training stack.

LoRA is a popular way to fine-tune models to inject style or knowledge. This paper explores the trade off between power and forgetting when using LoRAs. The research finds that LoRAs are able to learn less than full fine tuning, but maintain more of the original “out of distribution” performance.

Meta has announced a natively multi-modal model called Chameleon that operates on images and text simultaneously like GPT-4o. It outperforms many alternative models. The Meta team has since far surpassed this work on internal models.

Google's most recent family of models has an updated technical report. There is little information about the models or the data used, but there is significant information about the evaluation and safety measures taken, which is an interesting window into large-scale alignment.

Using GPT-Fast, you can run the most recent Google VLM in real time. This opens up a number of powerful downstream tasks as the model is extremely easy to fine-tune for specific tasks.

Experts is the easiest way to create and deploy OpenAI's Assistants and link them together as Tools to create a Panel of Experts system with expanded memory and attention to detail.

Google has released a model explorer that makes it easy to inspect the computation graph of your models. It can be useful for debugging and performance engineering.

Garple is a website that sells short, easy-to-say .com domains that would make good startup names. All the code was written by an AI coding agent.

Gamma raised $12M from Accel to reimagine presentations using AI. So far, 60M+ Gammas (AI-generated slides) have been created across 18M+ users.

Google kicked off its I/O developer conference with a two-hour presentation containing a rapid-fire stream of announcements. This article summarizes these announcements into an easy-to-read list. The announcements include Firebase Genkit, an addition to the Firebase platform aimed at making it easier for developers to build AI-powered applications in JavaScript/TypeScript; LearnLM, a family of generative AI models fine-tuned for learning; AI-generated quizzes on YouTube; a new discovery feature for apps in Google Play; scam call detection; Ask Photos, a feature that lets users search across their Google Photos collection using natural language queries; and much more.

Whether you're new to AI or a seasoned expert, this free webinar by MindStudio will show you how to automate the tedious parts of (almost) any job. Register for the No Code AI Building Webinar

Find ideal business customers using AI and Bayesian models for optimal delivery and conversion rates.

Julius AI has launched the first-ever AI code interpreter for the programming language R, potentially revolutionizing how academic researchers perform statistical analysis.

An experiment in running the small MoonDream VLM directly on the web with WebGPU inference.