AI Updates on 2026-01-08

AI Model Announcements

  • Alibaba releases Qwen3-VL-Embedding and Qwen3-VL-Reranker, achieving state-of-the-art performance on multimodal retrieval benchmarks with support for text, images, screenshots, videos, and 30+ languages @Alibaba_Qwen
  • OpenAI launches ChatGPT Health, a dedicated, private space for health conversations with enhanced encryption, per-user keys, data isolation, and exclusion from model training @nickaturley
  • Gmail enters the Gemini era with AI Inbox, AI Overviews for conversational questions, suggested replies, and proofread features powered by Gemini 3 @GoogleAI

AI Industry Analysis

  • Gemini surpasses 20% global AI website traffic share, reaching 21.5%, while ChatGPT drops below 65% to 64.5%, according to Similarweb's first 2026 tracker @demishassabis
  • a16z leads $28M seed round in Boltz PBC, whose open-source AI models for biomolecular research have been used by over 100,000 scientists, every top 20 pharma company, and thousands of biotechs @a16z
  • a16z announces $30M Series A investment in Protege, building real-world data infrastructure for AI development, serving majority of MAG7 companies and largest private AI players @a16z
  • Marc Andreessen describes AI as the biggest technological revolution of his life, clearly bigger than the internet, with comps to the microprocessor, steam engine, and electricity @a16z
  • Disney adds vertical video to Disney+ to accommodate Sora-generated shorts arriving later this year, with plans for user-generated content, leaderboards, and payouts @AndrewCurran_
  • Mistral awarded framework agreement by France's Ministère des Armées to use AI for strengthening defensive capabilities @AndrewCurran_
  • Snowflake announces intent to acquire observability platform Observe @TechCrunch
  • OpenAI acquires team behind executive coaching AI tool Convogo @TechCrunch
  • NVIDIA reportedly asking Chinese customers to pay upfront for H200 AI chips @TechCrunch
  • Perplexity launches Perplexity for Public Safety, offering law enforcement agencies Enterprise Pro free for 12 months for up to 200 seats @perplexity_ai

AI Ethics & Society

  • AI FOMO drives rushed deployments introducing security risks, worsened by safety revisionism where terms like red teaming are repurposed without adequate security rigor @AINowInstitute
  • Gergely Orosz warns that ChatGPT, Claude, and Perplexity were all wrong in their legal advice interpretation, emphasizing that AI cannot be relied upon for high-stakes decisions where accountability is needed @GergelyOrosz
  • Stanford research shows production LLMs can leak near-exact book text, with Claude 3.7 Sonnet reproducing 95.8% of Harry Potter and the Philosopher's Stone, demonstrating that safety filters can still miss memorized passages @percyliang
  • Ethan Mollick observes AI is causing homogenization of writing and loss of idiosyncratic academic writing styles, though overall clearer communication is generally positive @emollick
  • Research suggests online data quality, including MTurk, is dropping due to LLMs, creating an existential crisis for behavioral sciences @emollick

AI Applications

  • Wade Foster at Zapier uses Granola transcripts to reverse engineer company culture and build interview rubric agents that provide structured feedback on every candidate @clairevo
  • Brian Lovin uses Claude to create interactive explainer for how terminal UIs work, demonstrating AI as a learning tool for technical concepts @brian_lovin
  • Developers can now generate and animate 3D characters in under 5 minutes using Nano Banana Pro, Hunyuan3D 3.1, Mixamo, and Claude with three.js @deedydas
  • CrowdStrike collaborates with NVIDIA on specialized fine-tuning of Nemotron open models for security reasoning, outpacing generalized advanced models in accuracy @NVIDIAAI
  • NVIDIA releases Nemotron Speech ASR for low-latency voice agents, achieving 24ms transcription finalization and under 500ms total voice-to-voice inference time @NVIDIAAI
  • Google AI Studio team ships UI improvements including seamless file drag-and-drop, easier tool selection, better mobile support, and design consistency @OfficialLoganK

AI Research

  • Research shows RL (reinforcement learning) is naturally robust to catastrophic forgetting in continual learning, achieving 60% final average accuracy compared to 54% for sequential SFT, without using replay buffers @cwolferesearch
  • RL-based continual learning abilities do not come from KL divergence penalty, as both GRPO training with and without KL divergence achieve similar performance levels @cwolferesearch
  • Andrej Karpathy releases nanochat miniseries v1, demonstrating compute-optimal training following Chinchilla scaling laws with parameter-to-token ratio of 8, achieving GPT-2 comparable results for approximately $500 @karpathy
  • Francois Chollet announces Pallas integration in Keras, allowing developers to write high-performance hardware kernels in Python that lower to Mosaic for TPUs or Triton for GPUs @fchollet
  • NVIDIA Blackwell architecture delivers 2x+ token throughput on GB200 NVL72 with new TensorRT-LLM upgrades for MoE performance @NVIDIADC