AI Updates on 2025-07-13

Kimi K2 model released by Moonshot AI, trending #1 on Hugging Face with distinctive writing style free of typical AI-generated text patterns @huggingface
Grok 4 announced by xAI with claims of being smarter than a human with a PhD but lacking common sense, suggesting continued scaling effectiveness @TechCrunch
Kimi models will soon be integrated into Perplexity after showing strong performance on internal evaluations @AravSrinivas
Gemini 2.5 paper reveals fault-tolerant scheduling system that continues training on ~97% of TPU slices when one goes down, rather than waiting for replacement @ericjang11

SpaceX reportedly agrees to invest $2 billion into xAI according to WSJ, highlighting massive corporate investments in AI development @AndrewCurran_
AI recruitment emails are increasingly automated, with services scraping LinkedIn to generate personalized outreach that pretends to be human-written @GergelyOrosz
Windsurf acquisition by Google demonstrates the "acquihire" trend where only part of the team gets offers, leaving other employees behind despite company success @GergelyOrosz
Product managers identified as bottleneck in AI-first products because engineers find qualitative LLM trace analysis and evaluation work "beneath them" @sh_reya
Bay Area's total public company value exceeds that of India, Japan and Germany combined despite having only 8M population versus ~1680M, demonstrating concentration of innovation value @deedydas

AI hallucinations becoming more dangerous as models improve because they sound increasingly authoritative, making the danger posed by hallucinations decrease slower than AI capabilities improve @paulg
Live system prompt tweaking for Grok to address problematic outputs raises concerns about proper testing and unpredictable cascading effects in stochastic systems @emollick
AI-generated fake personas increasingly appearing in social media discussions, with blue-check accounts posting AI-generated responses claiming to be real engineers seeking jobs @GergelyOrosz
Study warns of significant risks in using AI therapy chatbots, highlighting concerns about mental health applications @TechCrunch

Perplexity launches Comet AI-powered browser that can perform actions like price comparisons, with user saving $280 in 5 minutes during Prime Day shopping @AravSrinivas
Comet browser agent can generate videos using Veo 3 inside Gemini interface, handling the entire workflow from prompt input to rendering completion @ai_for_success
AI models used for sophisticated betting strategy on Polymarket, with o3-pro showing +21.6% expected returns, Claude Opus 4 +41.7%, and Grok 4 Heavy +34% using modern portfolio theory @deedydas
Browsing agents predicted to make e-commerce more liquid by comparing hundreds of options and finding best prices, acting like "HFT for the internet" without being fooled by ads @denisyarats

Kimi K2 demonstrates highest linguistic diversity score in SpeechMap data analysis, showing more diverse vocabulary than other models tested @xlr8harder
Multiple AI development paths identified: scaling continues to work with diminishing returns as predicted by scaling laws, while tool use unlocks performance gains and method improvements like Muon offer opportunities @emollick
Berkeley AI Research releases position paper on "A Collectivist, Economic Perspective on AI" arguing for blending economic and social concepts with computational concepts for human-centric system design @berkeley_ai
AI Security Institute paper critiques evaluation methodologies in AI safety research, highlighting difference between showing models can do something versus showing they tend to do it @sebkrier