TLDR AI 2025-05-26

Some wear grit, results — and a NASDAQ ticker.

Meet Jeremy Barnett, CEO of RAD Intel, the pre-IPO AI company backed by Adobe, Fidelity, and tech insiders from Google, Meta, Amazon, and YouTube.

RAD is up 1600%, delivering 3.5x ROI for brands like Hasbro and MGM — and just secured its ticker: $RADI.

That momentum is powered by RAD's AI, which predicts what wins before it launches — already driving six-figure campaigns across beauty, retail, and entertainment.

Sales contracts have doubled in 2025, with new activations lined up for Sephora, Home Depot, and Anthropologie.

Shares are $0.60 until May 29. Get in now — before the price moves.

Nvidia plans to launch a new artificial intelligence chipset for China at a significantly lower price than its recently restricted H20 model. The company has scheduled mass production to begin as early as June. The new chips will be priced between $6,500 and $8,000, well below the $10,000 to $12,000 the H20 sold for. The lower price reflects its weaker specifications and simpler manufacturing requirements.

o3 Operator, OpenAI's CUA-powered browser agent, has replaced its previous GPT-4o-based model.

Netflix's FM-Intent is a hierarchical multi-task learning model that improves recommendation accuracy by modeling user session intent from implicit signals.

A lot of modern techniques have been developed since the original 'Attention Is All You Need' paper. This article looks at some of the most important ones developed over the years and tries to implement the basic ideas as succinctly as possible using the Pytorch framework.

Hugging Face has extended Tiny Agent design to Python, using the Model Context Protocol (MCP) to streamline tool integration for LLMs.

GUI-explorer is a training-free agent that autonomously explores and extracts knowledge from mobile app interfaces using unsupervised methods. It improves task success rates without retraining.

Large language models should never output anything but tool calls and their arguments. Exclusively working through tools allows models to externalize large parts of their intelligence to more efficient domain-specific programs. Some developers are already switching to a tool-use paradigm, but this is still limited to very short contexts and to only parts of the model output. This post provides examples of how the tool-use paradigm benefits different domains, like text editing, video understanding, and 3D generation.

SpatialScore is a new multimodal benchmark suite for evaluating 3D spatial reasoning in large models. It integrates 28K samples from 12 datasets.

ChatGPT usage grew amazingly fast - much faster than PCs, the web, or smartphones. However, it was just a website - people didn't need to go buy a new device or wait for infrastructure to be deployed. Despite its advantages, ChatGPT's daily active users to weekly active users ratio is pretty bad. This may fix itself over time as models mature and more use cases emerge. The chatbot interface might only be working for some people and some use cases, so it might be better to wrap features and capabilities inside other things.

Oracle will purchase about 400,000 GB200 chips for the Stargate data center in Abilene, Texas.

Based on Lao Tzu's Tao Te Ching, “The Way of Code” is a meditation on humility and creativity while vibe coding featuring interactive AI-generated visualizations.

A quick look at Gemma 3n, a new member of the Gemma family with free weights that was released during Google I/O.

Claude 3.7 Sonnet easily detects when it's being evaluated for scheming on the Apollo's scheming benchmark.

Flow TV continuously streams user-generated AI video clips, and their associated prompts, organized into thematic channels.