Gemini operates your Mac, while agents learn skills from experience
Today’s updates focus on executing tasks: operating applications in the background, finding tools when needed and learning personal skills. Research brings questions about testable hypotheses, AI use in court filings and the distant effects of data-center electricity demand.
01
Products and updates
5 stories
2026-11-19Google / Josh Woodward
Gemini for Mac adds computer use in background windows
Google announced computer use in the Gemini app for Mac, allowing users to delegate tasks across folders, applications and the web. Gemini clicks, scrolls and types in a live background window while users retain use of their own screen and cursor.
The official demonstration turns a sketch into a rigged 3D model in Blender, showing a workflow that moves from source material into a specialist application. It is a vendor demonstration; the announcement does not provide an overall task success rate or specify subscription and regional eligibility.
2026-11-19Google
Gemini API previews on-demand tool schema retrieval
Google introduced Tool Retrieval in preview through the Interactions API for Gemini 3.6 and later models. The model initially sees tool names and descriptions, loading complete parameter schemas only when needed. Search, schema loading and the resulting function call can happen within one API request.
Developers enable server-side retrieval with tool_search and defer_loading on selected tools; remote MCP tools are supported too. There is no separate retrieval fee, but normal model token rates apply. Search adds a small fixed input overhead, and loaded schemas continue to count as input during the conversation.
2026-11-19Google
Genie 4 reaches developers through the Gemini API
Following the consumer Project Genie update, Google made Genie 4 available through the Gemini Interactions API and AI Studio. Developers can start with text, a reference image or both to generate continuous 720p worlds at 24 frames per second, with synchronized native audio covering ambience, effects and dialogue.
Real-time keyboard input controls exploration, while live text prompts can change weather, lighting, objects or events. Google lists interactive prototypes, spatial design and agent training as potential uses. Outputs carry invisible SynthID watermarks, and AI Studio offers an interactive canvas and starter applications.
2026-11-19Perplexity / Aravind Srinivas
Perplexity Brain learns personal skills from past sessions
Perplexity announced that Brain in Computer now creates and updates personal skills automatically. It reviews past tool calls, reasoning and results to identify recurring procedures, proposing changes once it has enough evidence. Shared built-in skills also undergo paired replay on failure scenarios and engineer review before deployment.
In an eight-week randomized test on production traffic, the company reports that 72.9% of users received at least one new personal skill, versus 6.5% in the manual-addition group. Tool-call failures fell 11% and dissatisfied threads fell 7.6%, with no measurable cost or latency change. These are Perplexity’s own reported results.
2026-11-18OpenAI / ChatGPT
ChatGPT adds live transcription to dictation
ChatGPT announced Live Dictation, which turns speech into visible text while the user is still talking. The update reduces the wait for transcription after an utterance and lets users see their spoken message taking shape sooner.
OpenAI describes dictation as faster and more reliable, but the announcement does not quantify accuracy or specify languages and platform rollout dates. It concerns speech-to-text input; it does not announce a longer recording limit or a separate voice-notes application.
02
Engineering and research tools
2 stories
2026-11-19OpenAI
OpenAI connects performance alerts to fixes and verification with Codex
OpenAI described Performance Factory, an internal system that uses cloud-based Codex agents to investigate ChatGPT slowdowns, propose fixes for engineer review and verify recovery after deployment. Over slightly more than three months, its regression workflow tracked over 1,700 investigations and identified root causes in more than 600 cases.
Separate adversarial reviews challenge diagnoses and proposed fixes. In a human review of 100 root-cause suggestions, 70% were rated accurate; the remainder needed refinement or had not established causality. Persisted task state lets work resume after reviews and deployments.
2026-11-19Ai2
Ai2 launches Theorizer in Asta and releases its code
Ai2 launched Theorizer in Asta and open-sourced its code. The agent searches papers around a research question, designs an extraction template and compares findings to propose explanations with citations and scope conditions. Asta covers token costs up to a limit through its web interface; users running the open library locally pay their own model costs.
In a Peruvian dew-point study, its seasonal explanation was rejected by the data but prompted researchers to examine distance from weather stations, exposing problems in the source dataset’s spatial estimates. The output is a hypothesis to test. Theorizer can miss prior research or misjudge novelty, so a proposed explanation is not an established scientific result.
03
Research and societal impact
2 stories
2026-11-19Stanford HAI / RegLab
Stanford calibrates AI detection for court filings
Stanford HAI and RegLab introduced PhantomDocket, combining documented legal hallucination cases with control filings, and ScoreMix to calibrate detection estimates. For the studied 2026 complaints excluding debt collection, estimated AI-use prevalence was 38% in Santa Clara County, 36% in San Francisco, 17% in Tulsa County and 16% in Oklahoma County.
Changing detector settings caused the share flagged in the same Santa Clara sample to range from 10% to 98%. The researchers say calibrated detection is more suitable for estimating prevalence across a population than determining whether an individual filing used AI. Substantial false-positive and false-negative rates remain even for stronger detectors.
2026-11-19MIT
MIT models the distant air-quality impacts of data-center electricity
MIT researchers combined electricity dispatch, atmospheric transport and epidemiological evidence to estimate that added pollution from U.S. data-center electricity demand in 2023 corresponded to about 650 premature deaths: 513 associated with fine particles and 137 with ozone. That is about 5.3% of estimated nationwide electricity-related mortality. These are modeled attributable effects, not individually confirmed deaths.
The framework identifies which generators respond to extra demand each hour instead of relying only on average grid emissions. Electricity demand was concentrated in Virginia and Texas, while some higher per-capita impacts occurred in the Midwest. Grid connections, downwind pollution transport and atmospheric conditions spread effects beyond the facilities’ locations.
04
Industry developments
1 story
2026-11-19The Decoder
Reports describe canceled overseas compute deals and shipment pauses
The Decoder, citing The Information, reports that Moonshot AI canceled a $150 million contract with a cloud provider in Indonesia and paused other negotiations. The planned rental would have provided Nvidia Blackwell Ultra B300 capacity from next year for training Kimi K4. The report links the cancellation to tighter U.S. licensing requirements for overseas cloud compute.
The same article cites the Financial Times as reporting that TSMC and Samsung paused certain processor shipments to Chinese customers while applying for U.S. licenses, affecting AI and some advanced smartphone chips. These reports concern particular contracts and supply changes; they do not establish that all overseas cloud access or every chip shipment to Chinese companies has stopped.