Perplexity brings its local agent to Windows on RTX PCs
Portable Computer can work with files on compatible PCs without sending every task to the cloud. The hardware requirement is substantial: at least 24GB of graphics memory.
The briefing
Perplexity brings its local agent to Windows PCs with enough graphics memory. A former Spotify executive’s startup raises $5.5 million for hands-on music apps, while Google opens a new way to explore AI use across jobs and countries.
Perplexity brings its local agent to Windows on RTX PCs
Perplexity’s local agent is available in its Windows app for compatible NVIDIA GPUs. NVIDIA says locally completed work avoids Computer credits, while tasks needing cloud support require permission before information leaves the device.
The Windows release extends support beyond Linux RTX PCs and NVIDIA DGX Spark systems. It runs on GeForce RTX and RTX PRO GPUs with 24GB or more of VRAM, rather than on Windows machines generally.
In NVIDIA’s announcement, Gerardo Delgado describes an agent that plans multistep jobs, analyses data and combines information across files. NVIDIA says work completed locally keeps sensitive information on the device and does not consume Perplexity Computer credits.
Local work, with a cloud option
The app sets up a local model, such as Qwen 3.8 27B, adapted for Perplexity Computer and optimised for RTX hardware. That removes some of the model selection and software configuration normally involved in running a local agent.
Connectors include Outlook, OneDrive, Word, Google Drive, Gmail, Slack and GitHub. NVIDIA describes jobs such as sorting open pull requests, tracing fees through financial documents and analysing where new users abandon a signup process. These are company-described capabilities, not independent test results.
Local operation is not an absolute boundary: the agent can identify tasks needing cloud support and ask permission before sending information off-device. DGX Station support remains planned.
People with qualifying hardware can tackle document-heavy jobs without spending Computer credits on locally completed work. The permission step also gives them a decision point before a task sends information to cloud models.
Market signal
Music startup raises $5.5 million for remixing apps beyond prompts
A Vinyl Bar in Shibuya has raised $5.5 million in pre-seed funding. Its music apps emphasise hands-on remixing, with an iOS mixer already available and collaborative features still in development.
Investors in the round include Mantis VC, SV Angel and BoxGroup, alongside former Spotify executive Dawn Ostroff, who has joined as an adviser. Ivan Mehta reported the funding and product details for TechCrunch.
The startup’s main iOS app, bop, lets users arrange instruments and effects on a grid to remix a base song, change its tempo and add one-off effects. Finished mixes can be exported to TikTok or Instagram.
Remixing first, prompts alongside
Its smaller tools offer different ways to play with sound. Speed Surfer changes the speed and filtering of browser audio through a Chrome extension. Stacks lets users switch parts of a mix on and off in a three-by-three grid, while Sampler turns YouTube videos into beat-based music samplers.
This is not a blanket rejection of generative AI. The company has added a prompt-based sound-creation feature, and Zonoozy says he is open to using AI to build products and support their infrastructure. His stated approach is not to put it into every interaction.
A multiplayer version of bop is in development, rather than part of the described existing experience. The company also aims to release a broader music-making sandbox during 2026, including tools for creating instruments.
For people making short music clips, the existing apps offer direct control over instruments, effects and tempo rather than requiring a written prompt. Collaborative remixing is still a development project, not an existing feature to rely on.
What to watch
Google opens interactive ATLAS explorer for AI use across jobs
Google has launched an interactive view of its AI & Economy ATLAS data. Users can explore adoption patterns, while accompanying scientist survey findings distinguish reported time savings from actual gains in discoveries.
Readers can explore how occupations from electricians to purchasing managers use AI, examine household tasks and compare adoption across countries. Google’s Zanna Iscenko and Scott Strand announced the new interface for the long-term ATLAS research project.
Google’s examples show why the measure matters. Computer and mathematical occupations account for 30% of work-related AI usage in the US. That is a share of usage, not a finding that 30% of those workers use AI.
Time saved does not equal discoveries made
Alongside the explorer, Google, Google DeepMind and MIT FutureTech present research analysing 2,600 specialised models and surveying more than 600 scientists in the US and UK. Nearly half of those surveyed reported using some form of AI daily.
Scientists reported saving just under seven hours a week. The research also found time spent validating AI outputs, a backlog of hypotheses and bottlenecks in physical experiments and clinical validation. The savings are self-reported, not a measured increase in discoveries.
The researchers found different roles for different tools: language models were used across fields and tasks, while specialised models were relatively more common in health and life sciences and in prediction, generation and simulation.
The explorer gives readers a way to compare AI use by job and location rather than treating adoption as one global number. For research teams, the survey points to validation and experiment capacity as constraints worth examining.
What to watch next
- People with qualifying hardware can tackle document-heavy jobs without spending Computer credits on locally completed work. The permission step also gives them a decision point before a task sends information to cloud models.
- For people making short music clips, the existing apps offer direct control over instruments, effects and tempo rather than requiring a written prompt. Collaborative remixing is still a development project, not an existing feature to rely on.
- The explorer gives readers a way to compare AI use by job and location rather than treating adoption as one global number. For research teams, the survey points to validation and experiment capacity as constraints worth examining.
The takeaway
People with qualifying hardware can tackle document-heavy jobs without spending Computer credits on locally completed work. The permission step also gives them a decision point before a task sends information to cloud models.
The editor’s view
For people making short music clips, the existing apps offer direct control over instruments, effects and tempo rather than requiring a written prompt. Collaborative remixing is still a development project, not an existing feature to rely on.
Sources & further reading
- Gerardo Delgado, NVIDIA Blog: Perplexity Portable Computer Is Now Available on Windows, Powered by NVIDIA RTX ↗
- Ivan Mehta, TechCrunch: A Vinyl Bar in Shibuya is a startup from a former Spotify leader for making music apps ↗
- Zanna Iscenko and Scott Strand, Google: New insights from Google’s AI & Economy ATLAS ↗