TLDR AI 2024-07-15

If you're looking to get in front of a highly engaged tech audience including software devs, AI/ML engineers and AI enthusiasts, consider advertising in TLDR. We have 10 newsletters to choose from and a subscriber base of 4+ million tech professionals to effectively reach your target audience. Learn more about running your first campaign with us.

OpenAI is developing advanced reasoning capabilities for its AI models under a secret project called "Strawberry," aiming to enable the AI to autonomously scour the internet and plan for complex tasks.

Chinese attempts to lure domestic developers away from OpenAI will now be a lot easier as OpenAI has blocked users in China from accessing its tools and services. The tool was already blocked through the government's firewall, but developers were previously still able to use virtual private networks to access OpenAI's tools. OpenAI has not elaborated on the reason for its sudden decision. The decision may result in an acceleration of development by Chinese AI companies.

Bret Taylor, the Chairman of the Board of OpenAI, has introduced a new Agent Engineer role at his company, Sierra. According to a new blog post from one of the first in the role, the Sierra team sees agent engineering as a new discipline within AI engineering.

AuraFlow is a new open model trained by Fal. The model was trained using muP and is a 5.8B parameter model.

A model that dramatically outperforms state of the art for detecting when a language model hallucinates in its generations.

MARS is a new framework for text-to-image (T2I) generation that features the Semantic Vision-Language Integration Expert (SemVIE).

Distributed Low-Communication (DiLoCo) is a system from DeepMind that has been replicated by Prime Intellect. It enables cross datacenter training while maintaining GPU usage.

Answer AI has released a new lightweight and portable library for low-level GPU computations built on WebGPU. It can be used for writing cross GPU kernels and offers portable instructions.

ViTime is a foundation model for time series forecasting (TSF) that leverages visual intelligence rather than traditional numerical data fitting.

OpenAI's revenue is estimated to be $3.4B, much of that coming from its ChatGPT services.

Meta's optimizations in tail utilization of its machine learning inference services reduced failure rates by two-thirds, improved compute efficiency by 35%, and halved p99 latency. These enhancements ensure Meta's ad delivery systems can handle increased workload without needing extra resources while maintaining service level agreements. Continuous improvement strategies include predictive scaling and utilizing IPnext, Meta's unified platform, to manage the machine learning model lifecycle.

Gradient-Boosting RL (GBRL) brings the advantages of Gradient Boosting Trees (GBT) to reinforcement learning.

Poe has launched a feature called Previews that lets people create interactive apps directly in chats with chatbots.

AI Web Designer allows users and private companies to redesign their websites in seconds using our AI model.

Meta Platforms will release its largest Llama 3 model, featuring 405 billion parameters and capable of understanding and generating both images and text, on July 23rd.