TLDR AI 2025-03-31
xAI acquires X 🤝, Perplexity addresses rumors 🗣️, ARC-AGI-2 & ARC Prize 2025 💰
Gemini 2.5: Our most intelligent AI model (1 minute read)
Gemini 2.5 Pro, an advanced AI model, is leading LMArena benchmarks by a significant margin. It enhances performance and accuracy through improved reasoning capabilities. The model "thinks" by analyzing information and making informed decisions, building on Gemini 2.0 Flash Thinking advancements.
Announcing ARC-AGI-2 and ARC Prize 2025 (12 minute read)
The ARC Prize has launched ARC-AGI-2, a challenging benchmark aimed at advancing general AI systems. Current AIs score significantly lower compared to humans. The accompanying ARC Prize 2025 competition, hosted on Kaggle with a $1 million prize pool, aims to drive open-source innovation by rewarding efficiency and capability in solving ARC-AGI-2 tasks.
xAI acquires X in $80B all-stock deal (1 minute read)
xAI has officially acquired X in an all-stock transaction that values the combined company at over $110 billion.
👨💻
Engineering & Research
Tools and Weapons: Microsoft's Story, Told by Its CEOs (Sponsor)
Hosted by Microsoft Vice Chair and President Brad Smith, the
Tools and Weapons podcast explores technology's global impact. In the latest episodes, Bill Gates, Steve Ballmer, and Satya Nadella reflect on Microsoft's 50-year journey—past, present, and what's next.
Listen to the podcast here.Mobile-VideoGPT (GitHub Repo)
A lightweight multimodal video model under 1B parameters that features dual visual encoders and token pruning for real-time inference on edge devices.
Multimodal Adaptation Methods (GitHub Repo)
A curated list of approaches for multimodal adaptation that covers traditional domain adaptation, test-time adaptation, and newer methods.
Reasoning augmented generation code (GitHub Repo)
Traditional Retrieval-Augmented Generation (RAG) systems rely on a two-step process: first, semantic search retrieves documents based on surface-level similarities. Then, a language model generates answers from those documents. While this method works, it often misses deeper contextual insights and can pull in irrelevant information. ReAG – Reasoning Augmented Generation – offers a robust alternative by feeding raw documents directly to the language model, allowing it to assess and integrate the full context. This unified approach leads to more accurate, nuanced, and context-aware responses.
Awesome Vision-to-Music Generation (GitHub Repo)
A curated and regularly updated list of methods, datasets, and demos for turning visual inputs into music. It covers both academic and industrial advances in V2M.
OpenAI reshuffles Sam Altman's job once again (2 minute read)
OpenAI has expanded Brad Lightcap's role to oversee operations and partnerships, allowing CEO Sam Altman to concentrate on research and product development. Mark Chen and Julia Villagra have been promoted within the company amid leadership restructuring following recent executive departures. OpenAI is also transitioning to a for-profit model, leading to a lawsuit from cofounder Elon Musk.
Tim Cook says China's DeepSeek AI is 'excellent' during visit (3 minute read)
Despite DeepSeek AI's security and privacy issues, Tim Cook praised it as "excellent" during his China visit. The AI, developed in China, rivals top global models at lower development costs but faces investigations in the US and Europe. Cook, who is attending the China Development Forum, often has to make diplomatic remarks about China due to Apple's business interests there.
Get the most interesting AI stories and breakthroughs delivered in a free daily email.
Join 920,000 readers for
one daily email