TLDR AI 2025-02-19

HP acquired most of Humane's assets for $116M, leading to the discontinuation of Humane's AI Pins. The AI Pins will lose functionality by February 28. Customers are advised to transfer data. Humane's team will form HP's new AI innovation lab, HP IQ.

The former CTO of OpenAI, along with many extremely talented scientists and engineers, have joined to make another AI company. The goals are somewhat vague, but it seems to be a product and foundation model company, with a focus on infrastructure.

Meta is launching LlamaCon, an open-source AI developer conference, on April 29. The event will showcase advancements in the Llama AI model ecosystem, followed by Meta Connect in September for XR and metaverse developments.

DeepSeek has entered the Attention Alternative arena with this novel algorithmic method to speed up quadratic Attention. They get as much as an 11x speed up with no loss in overall performance.

A great benchmark of over 1,400 freelance software engineering tasks from Upwork, valued at $1 million USD total in real-world payouts. SWE-Lancer encompasses both independent engineering tasks - ranging from $50 bug fixes to $32,000 feature implementations - and managerial tasks, where models choose between technical implementation proposals. Top models earned $400k.

Researchers present a quantitative analysis of how ReLU neural networks fold input space, revealing self-similarity patterns. They introduce a novel metric to study these transformations and provide empirical results on benchmarks like CantorNet and MNIST.

Over 5 million tech pros start their day scanning TLDR for the latest news—so when an ad appears, it gets noticed.

That's why companies run TLDR ads to reach founders, engineers, and decision-makers without the noise of traditional ad platforms. Plus TLDR provides free ad copywriting and performance insights. Learn more.

A new approach to multi-task learning mitigates negative transfer by dynamically grouping tasks and updating them sequentially during training. The method, based on proximal inter-task affinity, significantly improves performance over existing multi-task optimization techniques.

CAMEL improves reinforcement learning efficiency by integrating LLM-generated suboptimal policies with dynamic action masking.

Perplexity has post-trained R1 to remove Chinese censorship. They do so in a way that doesn't harm underlying reasoning. It is Perplexity's first open weights release.

This article provides a comprehensive overview of xAI's Grok 3.

Google is introducing new tools for flood experts in Flood Hub, including an inundation history map and basin view, while collaborating with aid groups like GiveDirectly and the IRC to assist flood-affected communities.

Great X article by PufferLib maintainer. It dives into the key differences between types of RL and gives a great guide for base hyperparameters.

OpenAI has updated the "chain of thought" for its o3-mini model, enhancing transparency in its reasoning steps for both free and paid ChatGPT users.

Researchers at Stanford and the University of Washington have trained an AI reasoning model for under $50 using distillation and modifications to an Alibaba AI model.

OpenAI is considering building data center campuses in 16 states as part of President Trump's Stargate initiative, in partnership with Oracle, SoftBank, Microsoft, Nvidia, and Arm, aiming to invest up to $500 billion over four years.