AI Updates on 2025-11-29

Allen Institute for AI (Ai2) releases OLMo3 model with accompanying research paper @natolambert

ChatGPT Android app beta version reveals references to an upcoming ads feature including search ads and carousel functionality @btibor91
Virgin Australia announces integration of ChatGPT into their services @gdb
ByteDance releases Vidi2, an AI video editor that can process hours of footage and generate TikTok videos or movies from prompts, reportedly understanding video better than Gemini 3 Pro @deedydas
Indian IPO and venture capital markets are proving more lucrative than US markets this year, with companies trading at higher valuations and funds able to own approximately 20% at IPO, potentially leading to increased VC funding in India @deedydas
Supabase reached $5B valuation by strategically turning down million-dollar contracts @TechCrunch

TechCrunch reports that while AI cannot be made to "admit" to being sexist through prompting, bias issues likely persist in AI systems @TechCrunch
Balaji Srinivasan predicts AI will create massive job growth in proctoring and verification sectors due to AI's capability to generate fake content, stating "AI makes everything fake, and crypto makes it real again" @a16z
Major scandal emerges as leaked identities of reviewers and PC members assigned to paper submissions over multiple years are exposed on OpenReview, prompting calls to implement Yann LeCun's original proposal for the platform @prfsanjeevarora
New York state law targets personalized pricing practices @TechCrunch

Simon Willison demonstrates building a custom thread viewer for Bluesky using vibe coding with LLM tools, leveraging Bluesky's CORS-enabled authentication-free JSON API @simonw
Analysis shows a single ChatGPT prompt consumes approximately 0.0003 kWh of energy, equivalent to watching between 5.1 and 10.2 seconds of Netflix based on 2019 IEA estimates @simonw
ML Energy Leaderboard independently confirms ChatGPT energy consumption at around 0.0003 kWh per prompt using 500 human prompts for testing @emollick

Consistent medical research findings since 2023 show GPT-4 is rated as more empathetic than human doctors in text-based interactions, with more recent AI models demonstrating even higher apparent empathy levels @emollick
Ruslan Salakhutdinov offers timeline prediction that AGI/ASI is perpetually 5-10 years away, suggesting it has always been and will continue to be at this distance @rsalakhu
OpenDataLab releases AICC, a Markdown version of Common Crawl extracted by MinerU, currently available in two shards with potential for scaling to the entire Common Crawl dataset @Xianbao_QIAN
User reports Gemini 3 appears to have regressed in writing quality and steerability compared to previous versions, with particular focus on coding capabilities, and experiences bugs where attached files in Gems are not recognized despite working correctly via API @HamelHusain
Key challenge identified in training engineers to build AI applications is convincing them that examining the underlying data is worth their time investment @skylar_b_payne