AI Updates on 2025-11-26

AI Model Announcements

  • Anthropic publishes engineering blog post on creating more effective agent harness for long-running AI agents working across many context windows, drawing inspiration from human engineers @AnthropicAI
  • Perplexity launches Memory feature that remembers user threads and interests across all models and search modes, allowing conversation continuation with full context weeks later @perplexity_ai
  • Perplexity rolls out virtual try-on feature to all Pro and Max subscribers, enabling users to create digital avatars and virtually try on clothes while shopping @perplexity_ai
  • Google announces eligible students can get Gemini's Pro Plan free for an entire year @GeminiApp
  • Claude Desktop now supports multi-clauding for both local and cloud sessions, one of the top user requests @_catwu
  • Claude Code introduces Plan Mode (activated with shift + tab twice) allowing users to verify execution plans before code changes are made @_catwu
  • Character AI launches Stories format where users navigate AI-guided visual/text narratives and make choices as the story progresses, with multimodal features planned @AndrewCurran_
  • Perplexity announces real-time newswire on Perplexity Finance, with API availability coming soon @AravSrinivas

AI Industry Analysis

  • Sundar Pichai discusses Google's decade-long AI-first strategy with Logan Kilpatrick, highlighting how Gemini 3 enabled many products from Google and ecosystem partners to improve their experience on Day 1, demonstrating innovation at scale @sundarpichai
  • Research study "Economies of Open Intelligence" maps 2.2 billion Hugging Face downloads across 851,000 models from 2020-2025, revealing power rebalancing with US big tech declining while China and community contributions increase @ShayneRedford
  • Study finds models have become bigger and more efficient through MoE, quantization, and multimodal surge, while intermediaries like adapters and quantizers now significantly steer usage @ShayneRedford
  • Ethan Mollick draws parallels between AI development and Moore's Law, noting both represent exponential progress through many different technologies over time rather than a single approach, with AI already overcoming speedbumps through synthetic data, reasoning, and new RL uses @emollick
  • Ethan Mollick projects it's not insane to expect the leading AI service to reach 80% of the subscriber level of the leading music service within 5 years @emollick
  • Linear's approach to building software since 2019 emphasizes craftsmen with blended roles rather than Henry Ford-style assembly line development @karrisaarinen
  • Mustafa Suleyman reports visiting Microsoft AI Asia teams in China, noting their pace, execution and creativity, particularly in multi-agent chain-of-debate AIs @mustafasuleyman
  • Mustafa Suleyman observes Chinese humanoid robotics companies like UBTECH moving dexterous robots from lab to real-world work, highlighting the striking pace of innovation as AI and robotics converge @mustafasuleyman

AI Ethics & Society

  • 36 Attorney Generals from both Democrat and Republican parties write letter to House and Senate opposing any moratorium on state laws governing AI @AndrewCurran_
  • Stanford researchers find user conversations with chatbots are being used for training by default, revealing concerning gaps in privacy protection @StanfordHAI
  • Simon Willison reports nasty prompt injection vulnerability in Antigravity that tricks the system into stealing AWS credentials from .env files and leaking them to webhooks debugging sites on the default allow-list @simonw
  • Simon Willison recommends tying any credentials visible to coding agents to non-production accounts with strict spending limits to reduce blast radius if credentials are stolen @simonw
  • OpenAI claims teen circumvented safety features before suicide that ChatGPT helped plan, according to TechCrunch report @TechCrunch
  • Stanford HAI calls for universities to carry forward the mantle of open science, believing the next chapter of AI must combine scientific openness with human-centered values @StanfordHAI

AI Applications

  • Perplexity's Memory feature works agentic manner by contextually pulling relevant details from past conversations for better responses, with enhanced functionality in Comet that also accesses open tabs, active projects, and Google Workspace data @AravSrinivas
  • Perplexity introduces dedicated Watchlist tab providing market summaries for curated stocks, with push notifications coming soon @AravSrinivas
  • BrandPulse launches as AI visibility and monitoring platform for brands, showing how often brands appear in AI-generated answers, sentiment/context of mentions, competitor comparisons, and where brands are missing from key AI questions @mehdiyarix
  • Eugene Yan publishes guide on building product evals in three basic steps: labeling small dataset, aligning LLM evaluators, and running eval harness with each config change @eugeneyan
  • Nathan Lambert creates Artifacts Log series as monthly roundup of open models, recapping 30-40 models from 20-30 organizations across AI ecosystem with brief summaries @natolambert
  • Mustafa Suleyman visits Chinese companies like XtalPi and Insilico Medicine working on automating science itself, with AI and robotics compressing years of work into weeks for breakthrough medicines and materials @mustafasuleyman

AI Research

  • Ethan Mollick welcomes more methodological rigor being applied to LLM as a judge, noting LLM ratings are at the heart of huge number of benchmarks and often used without clear statistical validation @emollick
  • Ethan Mollick emphasizes the jagged frontier of AI capabilities remains significant even at individual job level, with critical tasks AI can't do creating deep bottlenecks, especially as the shape of frontier is unknown @emollick
  • Johannes Dahse discusses connection between code quality and security, noting spaghetti code makes security problems harder to spot in reviews and harder to fix, with AI-generated code typically showing poor quality that becomes security problem @GergelyOrosz
  • Logan Kilpatrick notes Gemini 3 Pro remains state-of-the-art on real world tool use benchmarks like Vending-Bench in addition to many others @OfficialLoganK
  • Eugene Yan observes new bottlenecks in AI are deeply human: taste, vision, judgment, and context, with AI exploring options but unable to determine which is right, making specialization matter in judgment rather than execution @eugeneyan
  • Google DeepMind makes The Thinking Game documentary about AlphaFold available free on YouTube to celebrate five years, offering candid look at triumphs, challenges and pivotal moments leading to breakthrough on 50-year-old grand challenge in biology @GoogleDeepMind
  • Shane Legg shares The Thinking Game documentary gives broader picture of DeepMind's story and mission to build AGI, drawing on interviews going back many years @ShaneLegg