AI Updates on 2025-10-18

Google introduces grounding with Google Maps in the Gemini API, bringing data about 250 million places together with Gemini to create new experiences @OfficialLoganK
Google releases upgraded Veo 3.1 model with enhanced realism and richer audio, now available in Flow by Google, Gemini app, Google Cloud Vertex AI and the Gemini API @sundarpichai
Google's nano image editing model is now available in Search with Lens & AI Mode, NotebookLM, and the Gemini App, with rollout to Google Workspace Slides and Google Photos coming soon @sundarpichai
Google AI Studio ships new feature allowing users to save and re-use system instructions, making it easier to test and reproduce outputs with Gemini @OfficialLoganK
Google releases C2S-Scale 27B foundation model built with Yale and Gemma for cancer research, along with DeepSomatic open-source AI model for genetic analysis @sundarpichai
Microsoft Research introduces SimPoly, a machine learning force field for polymer simulation that accurately computes polymer densities and glass transition temperatures @gncsimm
Keras now supports model quantization with just one line of code, supporting int4, int8, float8, and GPTQ modes for both custom and pre-trained models from KerasHub @_avichawla

Gergelyi Orosz observes that OpenAI internally still focuses on "getting to AGI" as a guiding principle, while Anthropic feels more grounded in improving step by step based on conversations with engineers at both companies @GergelyOrosz
WhatsApp bans general purpose chatbots from using its Business API, impacting AI assistant services like Perplexity's WhatsApp integration @TechCrunch
Perplexity recommends users switch from WhatsApp assistant to their Telegram assistant "askplexbot" following WhatsApp's policy changes @AravSrinivas
Deedydas notes the emergence of billion dollar seed rounds for AI companies including Lila Sciences, General Intuition, Periodic Labs, Thinking Machines, SSI, and Sierra @deedydas
Ethan Mollick reports that in companies he talks to, leaders are not following new AI developments or thinking about AGI, but instead focusing on steady accumulation of valuable use cases and process adjustments @emollick
Google AI Mode in Search is now fully rolled out to 200+ countries and territories in 43 languages, with users asking questions nearly 3x longer than traditional searches @sundarpichai

Amanda Askell notes that people often conflate AI erotica and AI romantic relationships, suggesting one is clearly more concerning than the other @AmandaAskell
Andrew Curran highlights a concerning example of AI-generated video showing Chuck Schumer saying a real quote, but the video itself was artificially created since it wasn't said on camera @AndrewCurran_
TechCrunch reports that the AI-generated video was posted on Senate Republicans' X account, potentially violating X's policies against "deceptively synthetic or manipulated media that are likely to cause harm" @TechCrunch
TechCrunch covers controversy around White House's David Sacks and OpenAI's Jason Kwon for their comments about groups promoting AI safety @TechCrunch
A viral "Definition of AGI" paper is revealed to contain fake citations that do not exist, with different articles present at the specified journal/volume/page numbers @m2saxon

Gergelyi Orosz shares his experience using Claude Code to build landing pages instead of using templates or Webflow, finding it more efficient for frontend work he doesn't specialize in @GergelyOrosz
Orosz demonstrates using Claude for configuration tasks like setting up static sites on Netlify, eliminating the need to look up and re-learn infrastructure setup procedures @GergelyOrosz
TechCrunch features a new iPhone app called Endless Summer that uses AI to create photorealistic vacation photos starring users without requiring actual travel @TechCrunch
Simon Willison creates a vibe-coded tool for displaying OpenAI's Responses JSON from deep research API calls in a more readable format, built using Claude Code @simonw
Scott Belsky predicts that "whatever technology sees the most will remember the most, and memory will reign above all else in the next era," positioning Google well but noting potential wild cards like local models and browser innovations @scottbelsky

Ethan Mollick emphasizes that early results like GDPval show today's AI models are good enough to create major transformations over 5-10 years as companies learn to deploy and integrate them into processes @emollick
Mollick backs up his belief that fine-tuning is mostly useful in narrow situations, remaining skeptical that it's the right solution for many problems where prompting alone might suffice @emollick
Andrej Karpathy provides detailed commentary on his recent podcast appearance, discussing AGI timelines, reinforcement learning limitations, and the "cognitive core" concept for improving LLM generalization @karpathy
Karpathy critiques current RL approaches, stating "you're sucking supervision through a straw" with poor signal/flop ratios, and advocates for alternative learning paradigms beyond traditional reinforcement learning @karpathy
Nathan Lambert notes that Karpathy's view that "Reinforcement learning is much worse than the average person thinks" is mostly correct, with too many people claiming RL will solve everything @natolambert
Simon Willison explores OpenAI's o4-mini-deep-research model via their Responses API, documenting his findings and building evaluation tools @simonw
Interconnects AI reports on the latest open models, noting Qwen's strong presence and discussing methods for accurately monitoring Hugging Face downloads and the continued degradation of open datasets @interconnectsai