TLDR AI 2025-03-10

Former OpenAI researcher Miles Brundage criticized the company for downplaying its cautious approach with GPT-2, which he claims aligns with today's deployment strategy. OpenAI's recent document outlines a continuous approach to AGI development, but Brundage warns this could lead to dismissing valid safety concerns. Competitive pressures could tempt OpenAI to prioritize faster releases over safety, raising questions about long-term risks.

The US Army's TRADOC is using an AI tool, CamoGPT, to identify and remove DEIA references from training materials per an executive order by President Trump. CamoGPT, developed by the Army's AI Integration Center, scans documents for specific keywords and has about 4,000 users. The initiative is part of a wider government effort to eliminate DEIA content, leveraging AI for increased efficiency in aligning with national security objectives.

Stability AI has announced a strategic partnership and investment from WPP, aiming to integrate generative AI into advertising and media production.

Step Law is a unified optimal hyperparameter scaling law that generalizes across diverse model shapes, architectures, and data distributions. This means, using these results, one can predict how models will likely perform before training.

SeqFusion is a framework that sequentially selects and fuses pre-trained models for zero-shot forecasting. Unlike conventional approaches, it minimizes data use to enhance privacy while maintaining competitive accuracy on diverse temporal patterns.

Adam has been the dominant optimizer for years in deep learning. However, recently the community has found that Muon might be a viable alternative. It accomplishes many of the same things as muP without requiring modifications to the model. This post describes some of the theoretical motivations behind the optimizer.

Join Sourcegraph's monthly live stream to get a live demo of their brand new Code Review Agent (EAP) - designed to make reviewing code faster, more precise, and less tedious. The team will also demo agentic chat, auto-edit, and other tools that help enterprises navigate complex codebases with ease. Join on 3/13 (Thursday) to see it all in action

This project introduces a federated learning-based approach for neural feedforward control, allowing multi-agent systems to improve tracking performance while preserving data privacy.

A lightweight 30m parameter model that can be used to augment any language model to enable it to understand and generate speech in response to general purpose queries. Critically, it doesn't require tuning of the underlying model and therefore can transfer across models easily.

Segment Anything (SAM) is an industry-leading model for visual analysis and segmentation. However, when there are two similar-looking objects in a video it can get confused. This new work tackles these “distractors” with additional memory augmentation and training.

The Gemini team has trained and released an excellent embedding model for text. It tops benchmarks and is reasonably priced while also boasting excellent speed.

Most video understanding models operate one frame at a time, which makes temporal questions somewhat challenging. STORM, which uses Mamba adapters, adds temporal attention operations. This post compares it against Qwen models.

TLDR is hiring part time curators to launch our TLDR Data newsletter.

The ideal candidate would have deep experience working directly with data warehouses, data pipelines, data lakes, and other modern cloud infrastructure.

Time commitment is ~2-3 hours/week paid at a rate of $100/hr. To apply please send your LinkedIn or resume to jobs@tldr.tech along with a couple sentences on why you'd be a good fit!

The US military has awarded a significant contract to Scale AI and partners including Anduril and Microsoft to integrate AI agents into operations for decision-making in military workflows. The Thunderforge project aims to enhance strategic planning speed and accuracy while maintaining human oversight. The Pentagon plans to eventually deploy this AI system across all its combatant commands.

Learn how Ai2 dramatically increased compute capacity and performance on NVIDIA Hopper GPUs to support open source innovation using Cirrascale Cloud Services accelerated by NVIDIA. Read the case study

AMD has sent a number of large computers with MI300X GPUs in them for the tiny grad team to use to test its drivers.

Google has opened applications for its AI for Energy accelerator, a 10-week equity-free program supporting startups integrating AI into energy solutions.

A federal judge ruled that authors can move forward with their copyright lawsuit against Meta, which alleges unauthorized use of their books to train AI models.