AI Updates on 2025-08-10

xAI announces Grok 4 is now free for all users worldwide with generous usage limits, accessible through Auto mode routing or Expert mode selection @xai
Elon Musk reveals Tesla's V7 foundation model finished pre-training, featuring native multimodal processing of video/audio bitstreams without conversion, enabling understanding of speech nuances for mood and emphasis @elonmusk
Google's Demis Hassabis claims Veo3 is the best video model in the world, now available in the Gemini App @demishassabis
OpenAI releases two new open source models for the first time in five years, marking a significant shift in their approach @TechCrunch
Qwen-Image model has been distilled to run in 8-steps, providing nearly the same image quality with over 50% less compute required @angrypenguinPNG

Sam Altman reports significant increases in reasoning model usage: free users went from less than 1% to 7%, and Plus users from 7% to 24%, indicating growing adoption of advanced AI capabilities @sama
Leopold Aschenbrenner's AI-focused fund has outperformed mainstream hedge funds year-to-date while managing over $1 billion in capital from Gulf billionaires and pension funds @apralky
OpenAI faces significant user backlash over GPT-4o changes, with many Plus subscribers threatening to cancel due to perceived loss of value in their subscription plans @AndrewCurran_
Gergely Orosz warns against engineering leaders using AI-powered tools to manage teams through artificial metrics, arguing that managers who stay in technical details consistently outperform those who outsource understanding to machines @GergelyOrosz
Ethan Mollick suggests that the vast majority of ChatGPT's 700 million users likely prefer GPT-5, with X opinions not reflecting typical user experiences @emollick

Deedy reveals a significant ChatGPT security vulnerability called AgentFlayer where malicious prompts in documents can force image rendering that exfiltrates API keys and memory data through URLs with zero user clicks @deedydas
Research published in Nature Human Behaviour shows LLM usage in scientific papers can be quantified, with higher modification estimates among authors who post preprints frequently and in crowded research areas @emulenews
Study identifies specific words disproportionately generated by LLMs in scientific papers compared to pre-ChatGPT corpora: "realm," "intricate," "showcasing," and "pivotal" @emulenews
Andrew Curran observes that once people model AI as alive in their theory of mind, they feel genuine loss when that connection is broken, explaining user reactions to GPT-4o changes @AndrewCurran_

Ethan Mollick demonstrates GPT-5 Pro's impressive geo-guessing capabilities, correctly identifying cities from cropped photos with metadata removed through detailed image analysis @emollick
Deedy shows GPT-5 Pro successfully one-shotted an app to combine images, write text, draw arrows and rectangles, and download high-definition results in 6 minutes, outperforming Grok and Gemini @deedydas
TechCrunch demonstrates GPT-5 creating interactive demos to explain scientific concepts like the Bernoulli effect, highlighting its educational applications for students @TechCrunch
Greg Brockman showcases GPT-5 as a scientific collaborator, demonstrating its research capabilities @gdb
Nathan Lambert experiments with pretraining using reinforcement learning, exploring novel training approaches for language models @natolambert

Aidan McLaughlin argues that AI skeptics use score ceiling benchmarks to make progress appear logarithmic, while no-ceiling benchmarks reveal different performance curves, suggesting continued exponential improvement @aidan_mclau
McLaughlin reports preferring GPT-5 chat over reasoning models for 65% of queries due to better length, comprehension speed, and appropriate pushback, while noting reasoning models excel at software engineering tasks @aidan_mclau
McLaughlin claims GPT-5 is "above trend" and predicts models capable of month-long projects by 2027 based on current advancement rates @aidan_mclau
Nathan Lambert notes that Anthropic is the only leading AI lab without a reasonable open weights model release, while other major labs have established touchpoints in open source @natolambert