AI Updates on 2025-08-16

OpenAI releases updated GPT-5 personality that is warmer and friendlier based on user feedback, with subtle changes like "Good question" or "Great start" without increased sycophancy @OpenAI
Google releases Gemma 3 270M, a hyper-efficient compact model designed for edge devices and task-specific fine-tuning @demishassabis
Anthropic announces new capabilities allowing their latest AI models to protect themselves by ending abusive conversations @TechCrunch

Paul Graham confirms that vibe coding (AI-assisted development) is here to stay, with infrastructure company founder reporting many vibe-coded apps are making money and the technology will only improve @paulg
Developer reports that some programmers are becoming significantly more prolific with AI coding tools, suggesting hiring decisions may increasingly favor AI-proficient developers @alexgraveley
Deedy explains how AI startups with $0 revenue can achieve $500M-$1B valuations through secondary share sales, creating a "get rich quick scheme" for founders and early employees @deedydas
Gergely Orosz observes that many services are struggling to effectively communicate the value of their AI features to customers, with unclear upselling attempts for "unlimited AI" @GergelyOrosz
OpenAI reportedly seeking $500 billion valuation, which would make it the world's most valuable startup, surpassing SpaceX @AndrewCurran_

Joanne Jang encourages AI professionals to define their personal ethical "line" - a boundary where they would leave their company if knowingly crossed and not walked back @joannejang
Simon Willison highlights 15 major prompt injection vulnerabilities discovered in AI products including ChatGPT, Cursor, GitHub Copilot, and others, demonstrating ongoing security risks @simonw
Ethan Mollick notes the AI research community's lack of dialogue with experts from economics, sociology, history, and psychology, missing opportunities to apply well-understood principles to AI development @emollick
Research shows doctors with AI outperform those without in diagnostics, but AI alone outperforms doctors, raising questions about optimal human-AI collaboration systems @emollick

Cursor CLI adds MCP (Model Context Protocol) support, Review Mode, file compression, and other UX improvements for AI-assisted development @cursor_ai
OpenAI enables Gmail and Google Calendar integration for ChatGPT Plus and Pro users globally, providing more contextual responses @OpenAI
Google Gemini App introduces chat history search functionality for both mobile and desktop users @GeminiApp
Qwen demonstrates advanced vision capabilities including object detection, weight estimation, and calorie calculation from meal photos with structured JSON output @Alibaba_Qwen
Jeremy Howard showcases SolveIt, a new development environment that fuses literate programming, live variables in AI prompts, and instant function-to-AI-tool conversion @HamelHusain

MIT CSAIL develops the first provably efficient method for machine learning with symmetry, potentially advancing drug and materials discovery by recognizing that symmetric transformations leave data fundamentally unchanged @MIT_CSAIL
Nathan Lambert ranks most memorable AI models: Claude 3.5 Sonnet for personality, o3 for search behavior, o1 pro for robustness, Gemini 2.5 pro for long context, and GPT 4.5 for personality @natolambert
Ethan Mollick observes that the new GPT-5 personality tends to give sandwich feedback (positive-criticism-positive) and is better at pushing back while being less sycophantic than GPT-4o @emollick
Google's Genie 3 can generate interactive worlds from text descriptions that users can explore in real-time, with potential applications in filmmaking, gaming, and agent training @a16z