AI Updates on 2025-10-07

Google releases Gemini 2.5 Computer Use model that can navigate browsers by clicking, scrolling and typing, setting new benchmarks with faster speed and safety features @GoogleDeepMind
OpenAI introduces gpt-image-1-mini, a new image generation model that is 80% less expensive than their large model @simonw
xAI launches Imagine v0.9 video generation model with massive upgrades in visual quality, motion, and native audio generation capabilities @xai
Alibaba's Qwen3-VL secures 2nd place in vision leaderboard and becomes first open-source model to rank first in both pure text and visual leaderboards @Alibaba_Qwen
LiquidAI releases LFM2-8B-A1B, an 8.3B parameter MoE model with only 1.5B active tokens designed to run on phones and laptops @maximelabonne

JPMorgan has reached AI equilibrium, spending $2 billion annually on AI development while saving the same amount, with plans to gain first-mover advantage through agentic AI at all levels @AndrewCurran_
Perplexity has overtaken Grok in web traffic with 168 million visits in the last 28 days, showing competitive dynamics in AI search @exec_sum
OpenAI reveals their top 30 customers who have used over 1 trillion tokens, demonstrating massive enterprise adoption @deedydas
Anthropic's next big bet is India, identified as one of their fastest-growing markets worldwide @TechCrunch
IBM incorporates Anthropic's Claude large language model family into their software development products @TechCrunch
Cohere launches Partner Program to accelerate global AI adoption and deliver measurable business outcomes through industry collaboration @cohere
HuggingFace community added 1 million new repositories in the past 90 days, with 40% being private repositories showing increased enterprise adoption @ClementDelangue

Motion Picture Association urges OpenAI to take immediate action to address copyright infringements by Sora 2, stating it's OpenAI's responsibility to prevent infringement @AndrewCurran_
Microsoft Research discusses red-teaming effort that exposed and secured a biosecurity vulnerability in AI-driven protein design, highlighting dual-use risks @MSFTResearch
Ethan Mollick notes that ChatGPT now refuses to do many things that Claude is happy to address, showing divergent safety approaches @emollick

Tesla releases FSD Supervised V14.1 with new arrival options allowing users to select parking locations and a new Driver Profile Sloth mode for more conservative driving @Tesla
Cursor introduces plan mode where AI can write detailed plans before starting complex tasks, allowing agents to run for significantly longer periods @cursor_ai
ChatGPT iOS app now supports video input including audio transcription through drag and drop functionality @AndrewCurran_
Google's Computer Use model becomes available in preview via API, enabling automated browser navigation @AndrewCurran_
Figma announces context integration coming to OpenAI's Codex, enhancing design-to-code workflows @figma
Copilot Vision helps users navigate software applications in real-time, demonstrated with video editing in Filmora @yusuf_i_mehdi

Google DeepMind introduces CodeMender, an AI agent that automatically fixes critical software vulnerabilities, potentially boosting developer productivity and security @demishassabis
Open-weights models like DeepSeek V3.2 Exp are reducing the gap to proprietary frontier models on agentic workflows, with DeepSeek surpassing Gemini 2.5 Pro on Terminal-Bench Hard evaluation @ArtificialAnlys
Research paper "Readability ≠ Learnability: Rethinking the Role of Simplicity in Training Small Language Models" challenges conventional wisdom about model training approaches @chrmanning
Stanford scholars are building a multimodal foundation model of cells to reveal protein-gene interactions and disease causes @StanfordHAI
PyTorch community explores combining quantization with 2:4 sparsity for greater LLM compression while maintaining accuracy on hardware-accelerated deployment @PyTorch