AI Updates on 2025-11-28

AI Model Announcements

  • DeepSeek releases first open-source model capable of winning IMO Gold in mathematics, using a generator-verifier-meta-verifier loop in natural language rather than formal proof systems like Lean, with potential applications across science and code domains @deedydas
  • Google announces Gemini 3 with enhanced capabilities including interactive app creation, visual learning features, and improved shopping assistance for Black Friday deals @GeminiApp
  • Shane Legg demonstrates Gemini 3 Pro with Thinking mode can create interactive simulations including double pendulum, orbital mechanics, and black hole accretion disk visualizations through natural language prompts @ShaneLegg

AI Industry Analysis

  • Andrew Ng provides comprehensive analysis of AI investment landscape, arguing the AI application layer is underinvested while infrastructure for model training may be experiencing a bubble, with VC hesitation stemming from difficulty picking winners rather than lack of opportunity @AndrewYNg
  • Ng reports infrastructure providers are supply-constrained for inference capacity despite low AI penetration, with agentic coding tools like Claude Code, OpenAI Codex, and Google CLI driving increased demand for token generation as market adoption grows @AndrewYNg
  • Paul Graham reports a startup using AI extensively operates with 6 employees instead of 16, representing a 2.7x productivity increase from AI implementation @paulg
  • Sam Altman states OpenAI is making a "very aggressive infrastructure bet" with new partnerships across energy, chips, and distribution, predicting significant economic value if model capability projections prove correct @a16z
  • Ben Horowitz argues crypto is the missing network layer for AI, providing money, identity, and provenance against deepfakes while AI provides the computational machines @a16z
  • Gap launches AI agent at full scale across four brands (Gap, Banana Republic, Athleta, Old Navy) handling order tracking, returns, and gift cards across web, mobile, and voice channels @btaylor
  • Andrew Curran argues GPT-4 alone was sufficient for massive societal transformation, particularly in employment, with only application development and reduced hallucination/inference costs needed rather than AGI or ASI @AndrewCurran_

AI Research

  • Ilya Sutskever states that while scaling current approaches will continue improving without stalling, "something important will continue to be missing" from AI models, sparking discussion about experiential learning and unified factored representation @ilyasut
  • Leading AI researchers show surprising convergence on AGI/ASI timelines: Demis Hassabis predicts 5-10 years, Francois Chollet about 5 years, Sam Altman within "a few thousand days," Yann LeCun about 10 years, Ilya Sutskever 5-20 years, and Dario Amodei as early as 2 years, with consensus that current paradigm enables massive economic impact even without AGI @polynoamial
  • CMU researchers introduce framework using privileged guidance from existing solutions to enable on-policy RL learning on hard problems, prepending minimal solution prefixes to difficult prompts to generate reward signals that generalize back to unconditioned tasks @rsalakhu
  • DeepSeek's mathematical reasoning model uses pure natural language generator-verifier-meta-verifier loop with RL-trained components, avoiding formal proof systems and potentially extending to any verifiable domain where checking is easier than solving @deedydas
  • Alex Graveley emphasizes importance of quantifying model jaggedness (uneven capability distribution) as the main differentiator between useful models for accelerating progress @alexgraveley

AI Applications

  • Ethan Mollick demonstrates Gemini 3 Pro excels at generating fictional scenarios including device diagrams, satellite photos, operational reports, and narrative sequences with high coherence @emollick
  • Google's AI Mode with Gemini 3 Pro Thinking enables users to create interactive physics simulations including Doppler effect, orbital mechanics, black hole visualization, and fluid dynamics through natural language prompts @ShaneLegg
  • Gergelyorosz highlights new book "Frictionless" addressing the question "AI can generate code in minutes - so why does shipping software still take forever?" focusing on developer experience and organizational friction @GergelyOrosz

AI Ethics & Society

  • TechCrunch reports on emerging federal vs state showdown in AI regulation, highlighting tensions in the race to regulate artificial intelligence @TechCrunch
  • Gergelyorosz emphasizes that adding an LLM to backend systems introduces prompt injection vulnerabilities that software engineers must address as a code security concern @giudegio