AI Updates on 2025-12-25

Alibaba releases Qwen Image Edit 2511 and Qwen Image Layered in ComfyUI, featuring enhanced editing with better consistency and the ability to decompose images into editable RGBA layers @Alibaba_Qwen
Liquid AI releases LFM2-2.6B-Exp, an experimental 3B parameter model built using pure reinforcement learning that achieves 42% on GPQA benchmark and outperforms DeepSeek R1-0528 (a model 263x larger) on IFBench, with consistent improvements in instruction following, knowledge, and math benchmarks @liquidai

NVIDIA acquires Groq for $20B through a non-exclusive licensing agreement, with founder Jonathan Ross and key team members joining NVIDIA to integrate Groq's inference technology while GroqCloud continues operating independently @JonathanRoss321
Big Tech companies are using licensing deals instead of traditional acquisitions to avoid antitrust scrutiny, with key staff joining the acquiring company while leaving a "zombie company" behind, following similar patterns seen with Google's Windsurf and Character acquisitions @GergelyOrosz
The US blocking Adobe's $20B acquisition of Figma has led large companies to avoid traditional acquisitions due to regulatory uncertainty, opting instead for licensing arrangements that don't trigger antitrust investigations @GergelyOrosz
NVIDIA strategically announced the Groq deal on Christmas Eve, timing the announcement during a period when there's minimal tech news coverage and most people are offline to minimize press attention @GergelyOrosz

François Chollet clarifies that the ARC-AGI series is not an AGI threshold but a compass pointing research toward the right questions, with ARC-AGI-1 testing minimal fluid intelligence, ARC-AGI-2 probing deeper reasoning complexity, and ARC-AGI-3 (launching March 2026) evaluating interactive reasoning and autonomous goal-setting @fchollet
Current image generation models still struggle with specific tasks including counting and precision (keys on a piano, ladder rungs), subtle movements (shifting furniture slightly), and rotations (rotating objects 90 degrees) @nlevin
Terence Tao suggests that while genuine artificial general intelligence may not be within reach of current AI tools, a weaker but valuable type of "artificial general cleverness" is becoming reality through pairing janky internal methods with strong verification filters that reject bad outputs at scale @rohanpaul_ai

GPT-image and Gemini demonstrate capability in incorporating measurements from websites and embedding furniture in place reasonably well for interior design tasks, though small tweaks after initial placement don't work well in either model @nlevin