AI Updates on 2025-09-01

Apple releases FastVLM and MobileCLIP2 models that are up to 85x faster and 3.4x smaller than previous work, enabling real-time vision language model applications including live video captioning locally in browsers @ClementDelangue
Microsoft releases upgraded VibeVoice Large ~10B Text to Speech model with MIT license, capable of generating multi-speaker podcasts in minutes @reach_vb
Tencent releases Hunyuan-MT-7B open translation model supporting 33 languages including 5 ethnic minority languages in China, with full pipeline from pretrain to ensemble refinement achieving SOTA performance @AdinaYakup

Research finds firms using AI are hiring fewer junior employees while not impacting senior roles, comparing companies across industries that hired for AI projects versus those that haven't @emollick
Evidence suggests junior hiring in AI-intensive fields has slowed down in the US, though establishing direct causation to AI remains challenging due to multiple macroeconomic factors @emollick
Users report canceling Anthropic subscriptions in favor of OpenAI's Codex, citing better limits and precision for coding tasks @steipete
Analysis suggests most of the ~150k Indian Masters students graduating in the US will not find jobs, with 70% studying CS/Engineering but insufficient tech jobs to meet demand, compounded by visa restrictions @deedydas
Runway is building a robotics-focused team and fine-tuning existing models for robotics and self-driving car customers @TechCrunch

Alimama Creative uses Qwen-Image and Qwen-VL to transform plain product shots into high-converting posters through fully automated creative pipeline, handling rewrites, prompts and visuals from SKU to ad in seconds @Alibaba_Qwen
User creates Gemini 2.5 Flash powered app that processes episode transcripts, show notes, and raw video to write step-by-step flows with perfectly timed screenshots, then posts via API to CMS @clairevo
Ethan Mollick demonstrates using nano banana to recreate the Bayeux Tapestry showing the Norman Conquest in war photography style, showing improved fidelity in capturing details compared to previous years @emollick
Lovable specializes in helping people build apps and websites through vibe-coding, particularly for users with no coding experience, letting them guide AI models as they produce code and websites @TechCrunch

GPT-5 Pro demonstrates impressive capabilities by critiquing a 2010 academic paper, suggesting methodological advances, identifying a previously unspotted error, and spontaneously running Monte Carlo simulations and sensitivity analyses @emollick
Both GPT-5 Pro and Gemini 2.5 Pro Deep Think are described as very impressive models for hard problems, though potentially undersold during launches as labs may not fully understand the market for slow, deep-thinking models @emollick
OpenAI's Codex merged 350K PRs in its first 34 days and has since merged over 1M PRs with explosive usage growth @AnjneyMidha
Growing movement to build LLMs in low-resource languages aims to expand AI access for underserved populations and address digital divide that prevents communities from accessing AI's economic benefits @StanfordHAI