TLDR AI 2025-09-19
How OpenAI uses Codex 👨💻, Gemini + Chrome 💻, Tesla Optimus lead joins Meta 🤖
Chrome: The browser you love, reimagined with AI (4 minute read)
Google introduces an AI-centric version of Chrome focused on productivity and safety. The browser now features Gemini, an AI assistant that understands context across tabs, and the enhanced omnibox with AI Mode for complex searches.
Nvidia and Intel's $5 billion deal is apparently about eating AMD's lunch (6 minute read)
Nvidia and Intel announced a $5 billion partnership focusing on creating a superior GPU-CPU SoC integration to compete against AMD. This collaboration targets the underserved integrated graphics market, which AMD dominates with its advanced offerings. Nvidia also aims to become a major Intel CPU customer for servers, challenging AMD's rising market share in that segment.
Luma AI's New Ray3 Video Generator Can 'Think' Before Creating (2 minute read)
Luma AI's new reasoning video model, Ray3, can create AI video clips with complex action sequences. It can spend more time working through prompts, making it better at handling more advanced scenes. The model can generate clips in 16-bit HDR. It has a draft mode that lets users test ideas quickly by generating shots in a lower resolution format.
Designing NotebookLM (12 minute read)
The lead designer for NotebookLM explains the UI evolution of the product, specifically how the team solved “tab overwhelm” by creating a unified workspace with responsive panels that adapt to different workflows—reading plus chat, chat plus writing, or reading plus writing. While the tool itself, like many AI tools, is very flexible, there is still a linear underlying flow: users import sources, interact through conversation, then generate structured outputs like notes and Audio Overviews.
Very few of Waymo's most serious crashes were Waymo's fault (5 minute read)
Out of Waymo's 45 most serious crashes, none were due to failures of the self-driving software, and 37 were caused by human drivers. Over 96 million miles, Waymos were involved in 34 airbag crashes. Human drivers average 159 over that same distance.
You can't eval GPT5 anymore (1 minute read)
The GPT-5 API is aware of the date, which is problematic because the model becomes aware that it is in a simulation when evals are run against it. There is evidence that models behave differently when they know they are in a simulation. Some models start questioning other parts of the simulation once they know they are in a simulation. The API being aware of the date may be an attempt at patching a safety risk.
Coding as the epicenter of AI progress and the path to general agents (16 minute read)
Coding is arguably the last tractable, general domain of continued progress for frontier models that people can interface with. It is a domain where models are already useful, and they continue to consistently stack on meaningful improvements. Many developers are starting to learn a new way of working with AI through new command-line code agents. Within a few years, autonomous agents will be functional, and AI will be able to improve the most complex codebases.
👨💻
Engineering & Research
You can't scale AI without containers (Sponsor)
AI workloads demand dynamic scaling, flexible consumption, and consistent deployment - exactly what containers provide. But managing thousands of GPUs or TPUs is not trivial.
Google Kubernetes Engine scales to 65,000 nodes, cuts inference costs 30%, and powers AI at Anthropic and LiveX AI. Learn how your existing Kubernetes expertise can fuel AI innovation in this
ebook by Google.
How OpenAI Uses Codex (8 minute read)
Every engineering department at OpenAI uses Codex daily. They emphasize fully utilizing Codex's features, not just talking to it like a chatbot: use “Ask Mode” before switching to “Code Mode,” put time into maintaining AGENTS.md files for persistent context, and use the task queue as a lightweight backlog.
Discovering new solutions to century-old problems in fluid dynamics (6 minute read)
Mathematicians have for centuries tried to develop complex equations to describe the fundamental physics involved in fluid dynamics. However, experts can still carefully craft scenarios that can make theory go against practice, creating scenarios that could never physically happen. These situations, also known as 'singularities', help mathematicians identify fundamental limitations in the equations of fluid dynamics and help improve their understanding of how the physical world functions. Researchers recently published an entirely new family of singularities in some of the most complex equations that describe fluid motion. Their approach presents a new way to leverage AI techniques to tackle longstanding challenges in mathematics, physics, and engineering that demand unprecedented accuracy and interpretability.
Unlike LLMs, we can't scrape the internet for robot data (2 minute read)
Project Go-Big aims to build the world's largest humanoid pretraining dataset. It is training robots directly from human video data. Its humanoid robots learn navigation end-to-end using only human video. Examples of the videos the robots are training on are available in the thread.
Hard Disk Unexpectedly Not Dead (12 minute read)
While progress has been made in DNA storage, the technology is still many decades away from competing with the hard disk. The research isn't a waste of resources, as it will eventually be feasible and economic. Time-scales in the storage industry are long. Bandwidth and cost remain the core problems of DNA storage.
Get the most interesting AI stories and breakthroughs delivered in a free daily email.
Join 920,000 readers for
one daily email