Alek Safar just introduced ClockBench to test multimodal models on analog clocks, and they struggled hard. While humans achieved 89.1% accuracy, the best model, Google's Gemini 2.5 Pro, only hit 13.3% with a median error of an hour. Interestingly, models solve time math easily once they get the time right, proving visual spatial reasoning remains the real bottleneck.
Google Gemini just claimed the top free spot on the US Apple App Store, overtaking ChatGPT. The surge is largely driven by its viral Nano Banana image editor, known for strong character consistency, which drove 23 million new installs and over 500 million edits in two weeks. Free users get up to 100 edits daily.
OpenAI is quietly turning ChatGPT into a transactional utility. Latest app builds reveal a new "Orders" section in settings, pointing to native credit card checkout and tracking for plugins, services, and digital goods. The update also moves desktop Voice Mode into the main app window and lays groundwork for parental controls. Why it matters: ChatGPT is rapidly morphing from a conversational assistant into an all-in-one operating platform.
California lawmakers just passed SB 243, sending a bill to Governor Gavin Newsom that would hold AI companion bots like Character.AI and ChatGPT legally accountable for user safety. Platforms would have to block self-harm and explicit content, and ping minors every three hours reminding them their digital pal isn't human. Why it matters: If signed, California becomes the first state to regulate emotional AI, allowing users to sue platforms for up to $1,000 per violation.
Alibaba just released Qwen3-Next-80B-A3B, an ultra-sparse MoE model that packs 80B total parameters while only activating 3B per token. It delivers 10x faster inference and 10x cheaper training than Qwen3-32B, with its Thinking variant outperforming Gemini-2.5-Flash-Thinking. Basically, flagship-grade reasoning is now running at lightweight speed.
Anthropic just gave Claude long-term memory, letting the assistant retain project context, workflows, and preferences so you don't have to keep re-explaining things. Each project keeps its own isolated memory bank, users can edit what Claude remembers, and an incognito mode lets you chat off the record.
Microsoft's Copilot Labs just launched Scripted mode for audio generation, powered by the new MAI-Voice-1 model. It reads your input verbatim instead of improvising, alongside an Emotive mode for added drama and a Story mode that performs multiple characters.
OpenAI just rolled out full MCP tool support inside ChatGPT. In developer mode, developers can now create connectors for write actions like updating Jira tickets or triggering Zapier workflows, rather than just fetching data. It basically turns ChatGPT into a practical automation hub for daily engineering workflows.
Perplexity is prepping local Model Context Protocol (MCP) support for its Comet browser. While Comet previously only supported remote connections, local MCPs allow its agent to browse local folders, attach files, and trigger desktop apps from chat prompts. Why it matters: It pushes AI browsers beyond the web sandbox into full desktop automation, though the feature is still tucked away in dev builds.
Baidu just launched its ERNIE X1.1 reasoning model at WAVE SUMMIT 2025, claiming a 34.8% jump in factuality over X1 and performance rivaling GPT-5 and Gemini 2.5 Pro. Alongside the flagship, Baidu open-sourced a 21B thinking model and updated its Comate coding assistant with multi-agent workflows. The new model is already live on the web and mobile apps.
Anthropic just gave Claude the ability to create and edit Excel spreadsheets, documents, PowerPoint slide decks, and PDFs directly in-app. Instead of returning plain text, Claude runs code in a private sandboxed environment to tackle multi-step analysis and data formatting. The feature is now rolled out to paid plans.
Google just expanded AI Mode in Search to Hindi, Indonesian, Japanese, Korean, and Brazilian Portuguese. Powered by a custom version of Gemini 2.5, it taps multimodal reasoning to provide locally relevant answers rather than simple machine translation.
Google just added direct audio uploads to the Gemini app, addressing what the company calls users' most-requested feature. Free users can upload clips up to 10 minutes long (capped at five daily), while AI Pro and Ultra subscribers get up to three hours. Google also upgraded NotebookLM with new report styles like study guides and flashcards.
Google just integrated Gemini directly into Sheets with a new =AI function. Users can simply type the formula alongside a natural-language prompt into any cell to categorize feedback, summarize reviews, or generate personalized copy across hundreds of rows without touching complex scripts.
Atlassian just agreed to acquire The Browser Company, the startup behind the Arc and Dia browsers. The workplace software giant plans to shape Dia into a dedicated browser for knowledge workers, injecting AI skills, personal work memory, and enterprise-grade security. Why it matters: Atlassian wants to control the actual canvas where daily SaaS work happens.
Google DeepMind researchers just proved that single-vector embeddings have a fundamental geometric limit that breaks advanced RAG setups. On their new LIMIT stress test, top embedding models from Google and Snowflake saw recall drop below 20%, while old-school BM25 excelled. Why it matters: simply scaling models won't fix retrieval.
Microsoft just debuted its first homegrown models: MAI-Voice-1 and MAI-1-preview. The voice generator produces a full minute of expressive speech in under a second on a single GPU, while MAI-1-preview is an end-to-end mixture-of-experts foundation model trained on 15,000 H100 GPUs and now testing on LMArena. Microsoft is finally building its own muscle.
Anthropic just updated its consumer privacy policy to train AI models on user chats and coding sessions, retaining data for up to five years. The change takes effect September 28 across Free, Pro, and Max tiers with opt-in enabled by default, though enterprise and API accounts remain exempt.
Anthropic just updated its consumer terms and privacy policy, giving Free, Pro, and Max users a choice to let their chats and Claude Code sessions train future models. Opting in extends data retention from 30 days to five years, while commercial and API customers remain untouched. Users have until October 8, 2025, to decide in their privacy settings.
Microsoft just brought its Copilot AI assistant directly to Samsung's 2025 lineup of smart TVs and monitors. Viewers can hold the remote's mic button to trigger a floating animated avatar for movie recommendations and spoiler-free episode recaps. Why it matters: It brings conversational AI to the living room's biggest screen, though it is currently restricted to newer hardware.
