TLDR AI 2025-01-02

Nvidia acquired AI orchestration firm Run:ai for an estimated $700 million and plans to open-source its software. This move aims to enhance GPU cloud orchestration, offering more flexibility and efficiency for AI infrastructures. Nvidia's acquisition aligns with its strategy to broaden its AI ecosystem and address antitrust concerns.

Deepseek, a Chinese AI startup backed by the hedge fund High-Flyer, has gained attention by outperforming OpenAI on reasoning benchmarks and initiating price wars with its efficient AI models. Led by CEO Liang Wenfeng, Deepseek prioritizes open-source foundational technology and leverages extensive computing resources without external funding. The startup focuses on AGI research, challenging prevailing innovation norms in China while attracting top domestic talent.

Sam Altman is working to shift control of OpenAI from its founding nonprofit to a for-profit structure to better compete with tech giants. Discussions involve setting fair compensation for the nonprofit and accommodating external stakeholders like Microsoft. OpenAI faces a deadline to restructure within two years or risk converting recent investments into debt.

PrefixQuant is a new method that improves LLM quantization by isolating outlier tokens offline, eliminating the need for costly per-token dynamic quantization.

GCA-HNG is a framework for generating more effective hard negatives by considering global sample correlations instead of just local ones.

A new study assesses the ethical decision-making of 51 LLMs in autonomous driving scenarios, analyzing alignment with human moral judgments across various models, including GPT, Claude, and Llama.

Xmodel-LM is a compact and efficient 1.1B language model pre-trained on around 2 trillion tokens. Trained on a self-built dataset (Xdata), which balances Chinese and English corpora based on downstream task optimization, Xmodel-LM exhibits remarkable performance despite its smaller size. It notably surpasses existing open-source language models of similar scale.

Automatically figuring out what Sparse Autoencoder features mean can be challenging. This tool aims to help with some clever clustering and probability management algorithms.

SAG is a self-supervised framework designed to improve the generalization of optical flow and stereo methods in real-world applications. Unlike traditional approaches, SAG uses advanced reconstruction techniques to generate datasets from RGB images. It quantifies the confidence level of these results to address imperfections.

YouTube and CAA are partnering on a program to help talent manage AI-generated fakes on the platform using early-stage likeness management technology. This tool will let actors and athletes identify unauthorized AI replicas and submit removal requests. The collaboration aims to protect IP rights while testing and refining AI detection systems before a broader rollout.

Sakana AI's "universal transformer memory" technique optimizes language models by using neural attention memory models to reduce unnecessary tokens, improving efficiency and performance. This approach cuts compute costs, facilitates faster processing, and proves beneficial across various tasks using open-source models. Researchers demonstrated significant memory savings and performance enhancements, especially in handling long sequences for text, code, and multi-modal tasks.

Olympus provides a comprehensive framework for evaluating AI creativity across multiple domains, offering insights into generative model capabilities and limitations.

TAMP is a multi-scale integrated Transformer model designed to enhance non-ideal CT (NICT) imaging.

Engineered Arts has restructured as a U.S. company and raised $10 million to expand its humanoid robot operations, bringing total funding to $16.2 million.

NVIDIA's Jetson Orin Nano Super Developer Kit boosts generative AI performance, with a new model priced at $249.