AI Updates on 2025-09-28

Qwen3-Max is now available and ready for users to build applications, with new capabilities including Code Interpreter and Web Search for data fetching and visualization @Alibaba_Qwen

BigTech companies will spend $345B on capex for AI buildouts this year, representing a 2.5x increase in just 2 years, with OpenAI's Stargate promising $500B by 2029 representing ~25% of projected $2T spend @deedydas
OpenAI is reportedly spending $150M+ per year on Datadog, more than 2x what Datadog itself spends, highlighting the massive infrastructure costs of AI companies during rapid growth phases @GergelyOrosz
Hollywood studios are quietly embracing AI technology under the radar, with multiple public announcements about high-profile AI projects expected at the beginning of the new year according to Luma AI's Dream Lab LA head @AndrewCurran_
NVIDIA CEO Jensen Huang claims the company checks in more open-source AI models and datasets than anyone except AI2, positioning NVIDIA as a major contributor to open AI development @natolambert
Every researcher on the Google Veo 3 paper, described as the world's best video generation model, is not from the USA, highlighting global talent distribution in AI research @deedydas

Ethan Mollick demonstrated using ChatGPT Codex to recreate a lost Maxis simulation game (SimRefinery) from just an article and screenshot, building a playable prototype without touching any code directly @emollick
Claude Code successfully debugged a complex macOS Finder issue that grew to 8GB in size through ~10 iterations over 30 minutes, demonstrating new debugging capabilities that didn't exist before AI agents @GergelyOrosz
Scott Aaronson published his first paper where a key technical step in the proof came from AI, specifically using GPT-5-Thinking, describing the AI's contribution as "clever" by academic standards @AndrewCurran_
AI models can now solve most common CAPTCHAs better than humans, with the main reason CAPTCHAs still work being that major LLMs often refuse to complete them rather than lacking capability @emollick

DeepMind's new paper "Video models are zero-shot learners and reasoners" demonstrates that generative video models are to vision problems what LLMs were to NLP problems - single models capable of solving a wide array of challenges @simonw
The progression from "agents are nowhere close to working" to "general purpose agents are actually useful for a range of tasks" has occurred in less than a year, with significant improvements in tool use, work steps, and error reduction @emollick
RL research is becoming like pretraining/modeling with a huge vibe shift, as most published RL research hasn't been using enough compute to make decisions matter as much, though this is slowly changing @natolambert
Anthropic researchers predict crossing parity with human experts within "probably only a few months," with the company having stated in 2023 that 2025/26 models could automate large portions of the economy @AndrewCurran_