An AI news roundup, citing Nvidia, says the assistant can complete work locally inside its Windows app.
A September 15 roundup, citing Nvidia, says Perplexity's Portable Computer can now complete work locally in its Windows app. A separate post touts file reading and job execution, and claims private data stays entirely out of the cloud.
—
2 Sources, first seen ago
—
Not ranked yet
—
Not ranked yet
@Romka2xTITLE AI News: Perplexity Goes Local, ChatGPT Human Reviews Reported DESCRIPTION AI developments from the past 24 hours cover local assistants, reported human chat reviews, crowd sensing, software repairs and an autonomous truck unveiling. https://blogs.nvidia.com/blog/local-ai-perplexity-windows-pcs/ — NVIDIA https://www.404media.co/inside-project-lily-the-humans-reading-your-chatgpt-chats/ — 404 Media https://news.samsung.com/global/samsung-and-verizon-complete-successful-isac-trial-over-virtualized-network-at-global-sports-event — Samsung https://github.com/QwenLM/qwen-code/releases/tag/cua-driver-rs-v0.20.9 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11910 — Qwen Code https://github.com/ollama/ollama/releases/tag/v0.34.1-rc2 — Ollama https://github.com/ollama/ollama/pull/18376 — Ollama https://blog.pony.ai/pony-ai-unveils-new-gen-4-robotruck-in-collaboration-with-gac-commercial-vehicle/ — http://Pony.ai https://github.com/QwenLM/qwen-code/pull/11831 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11636 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11881 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11903 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11906 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11805 — Qwen Code AI-generation disclosure: This video uses AI-generated narration and visuals. TRANSCRIPT NVIDIA says Perplexity's Portable Computer can now complete work locally inside its Windows app, with no Computer credits consumed for locally finished tasks. Its September 14 announcement limits availability to GeForce RTX and RTX PRO graphics processors with at least 24 gigabytes of video memory, and lists connections including Word, Outlook, Gmail and Google Drive. For people using that hardware, NVIDIA says tasks needing cloud support ask permission before information leaves the device, giving users a decision point when their work needs to move off their computer. Hundreds of contractors are being hired to rate and critique real users' ChatGPT conversations, according to a September 14 investigation by 404 Media based on leaked documents and prompts. The publication reports that reviewers do not see usernames, while OpenAI told it the company tries to remove personal information and acknowledged that sensitive details can pass through. That means personal material in some conversations can reach human reviewers despite those removal efforts, and Anthropic also confirmed to the publication that it uses human review to improve its models. Samsung says artificial intelligence converted changes in wireless signals into a near-real-time map of crowd density during a completed trial with Verizon. In a September 14 announcement, Samsung described a radio receiving signal variations from a commercial 5G device at an international soccer fan event in Dallas, with nearby AI processing turning those readings into a heatmap. Samsung says the demonstration measures crowd density without identifying individuals and could support congestion management, a potential use for event organizers that remains a trial result rather than an established safety improvement. Qwen maintainers say their latest Linux automation repair returns an unconfirmed result when an input may already have been sent, instead of automatically replaying it. The September 15 CUA Driver release follows the previously covered Linux input repair, and its associated change also avoids traversing application-managed virtual descendants while sharing a single deadline for collecting interface information and geometry. The practical change is avoiding duplicate clicks or typing after an uncertain result, but the maintainer says omitted spreadsheet content still requires screenshots and validation did not cover native Wayland, macOS or Windows acceptance. Ollama's September 14 release candidate changes a repeated-token failure into an explicit error instead of returning an incomplete answer, with the repetition limit raised to one hundred. The prerelease also evicts active-conversation cache snapshots, checks free memory and waits for removed model processes before loading the next model through the MLX software path. For affected small graphics processors running Gemma3n, Ollama says another safeguard avoids falling back to the main processor in a way that can corrupt image representations and make the model describe the wrong image. Pony AI unveiled a fourth-generation autonomous heavy-duty truck based on GAC's T9 platform at the Hannover transportation show on September 14. The company says volume production is scheduled for later in 2026, with long-haul freight, dedicated routes and port transportation among the intended uses. Pony AI says its autonomous-driving kit costs 70 percent less than the previous generation and forecasts a 30 percent reduction in transportation costs per ton-kilometer, making the freight savings a company expectation rather than an achieved operating result. Qwen Code merged six source changes that correct skill-enabling mistakes, organize background results, repair standalone terminals, expose hook settings, clarify hook progress and output, and let extensions contribute workflows. The changes stop colons in skill names causing false extension matches, make batch enabling order-independent, coordinate when background results reach the model, show detailed hook configuration and emit internal start/end outcomes, reject malformed structured hook output as a non-blocking error, and list extension workflows during installation and updates. Developers using this source work gain clearer controls, including workflows disappearing when an extension is disabled and a standalone terminal repair verified on Linux x64, but Linux arm64 still lacks a pinned terminal package and these merges do not establish packaged application availability.
@aiseomasteryPERPLEXITY JUST TURNED YOUR WINDOWS PC INTO AN AI WORKER It reads files, runs jobs, and keeps your private data off the cloud entirely.@Romka2xTITLE AI News: Perplexity Goes Local, ChatGPT Human Reviews Reported DESCRIPTION AI developments from the past 24 hours cover local assistants, reported human chat reviews, crowd sensing, software repairs and an autonomous truck unveiling. https://blogs.nvidia.com/blog/local-ai-perplexity-windows-pcs/ — NVIDIA https://www.404media.co/inside-project-lily-the-humans-reading-your-chatgpt-chats/ — 404 Media https://news.samsung.com/global/samsung-and-verizon-complete-successful-isac-trial-over-virtualized-network-at-global-sports-event — Samsung https://github.com/QwenLM/qwen-code/releases/tag/cua-driver-rs-v0.20.9 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11910 — Qwen Code https://github.com/ollama/ollama/releases/tag/v0.34.1-rc2 — Ollama https://github.com/ollama/ollama/pull/18376 — Ollama https://blog.pony.ai/pony-ai-unveils-new-gen-4-robotruck-in-collaboration-with-gac-commercial-vehicle/ — http://Pony.ai https://github.com/QwenLM/qwen-code/pull/11831 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11636 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11881 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11903 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11906 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11805 — Qwen Code AI-generation disclosure: This video uses AI-generated narration and visuals. TRANSCRIPT NVIDIA says Perplexity's Portable Computer can now complete work locally inside its Windows app, with no Computer credits consumed for locally finished tasks. Its September 14 announcement limits availability to GeForce RTX and RTX PRO graphics processors with at least 24 gigabytes of video memory, and lists connections including Word, Outlook, Gmail and Google Drive. For people using that hardware, NVIDIA says tasks needing cloud support ask permission before information leaves the device, giving users a decision point when their work needs to move off their computer. Hundreds of contractors are being hired to rate and critique real users' ChatGPT conversations, according to a September 14 investigation by 404 Media based on leaked documents and prompts. The publication reports that reviewers do not see usernames, while OpenAI told it the company tries to remove personal information and acknowledged that sensitive details can pass through. That means personal material in some conversations can reach human reviewers despite those removal efforts, and Anthropic also confirmed to the publication that it uses human review to improve its models. Samsung says artificial intelligence converted changes in wireless signals into a near-real-time map of crowd density during a completed trial with Verizon. In a September 14 announcement, Samsung described a radio receiving signal variations from a commercial 5G device at an international soccer fan event in Dallas, with nearby AI processing turning those readings into a heatmap. Samsung says the demonstration measures crowd density without identifying individuals and could support congestion management, a potential use for event organizers that remains a trial result rather than an established safety improvement. Qwen maintainers say their latest Linux automation repair returns an unconfirmed result when an input may already have been sent, instead of automatically replaying it. The September 15 CUA Driver release follows the previously covered Linux input repair, and its associated change also avoids traversing application-managed virtual descendants while sharing a single deadline for collecting interface information and geometry. The practical change is avoiding duplicate clicks or typing after an uncertain result, but the maintainer says omitted spreadsheet content still requires screenshots and validation did not cover native Wayland, macOS or Windows acceptance. Ollama's September 14 release candidate changes a repeated-token failure into an explicit error instead of returning an incomplete answer, with the repetition limit raised to one hundred. The prerelease also evicts active-conversation cache snapshots, checks free memory and waits for removed model processes before loading the next model through the MLX software path. For affected small graphics processors running Gemma3n, Ollama says another safeguard avoids falling back to the main processor in a way that can corrupt image representations and make the model describe the wrong image. Pony AI unveiled a fourth-generation autonomous heavy-duty truck based on GAC's T9 platform at the Hannover transportation show on September 14. The company says volume production is scheduled for later in 2026, with long-haul freight, dedicated routes and port transportation among the intended uses. Pony AI says its autonomous-driving kit costs 70 percent less than the previous generation and forecasts a 30 percent reduction in transportation costs per ton-kilometer, making the freight savings a company expectation rather than an achieved operating result. Qwen Code merged six source changes that correct skill-enabling mistakes, organize background results, repair standalone terminals, expose hook settings, clarify hook progress and output, and let extensions contribute workflows. The changes stop colons in skill names causing false extension matches, make batch enabling order-independent, coordinate when background results reach the model, show detailed hook configuration and emit internal start/end outcomes, reject malformed structured hook output as a non-blocking error, and list extension workflows during installation and updates. Developers using this source work gain clearer controls, including workflows disappearing when an extension is disabled and a standalone terminal repair verified on Linux x64, but Linux arm64 still lacks a pinned terminal package and these merges do not establish packaged application availability.@aiseomasteryPERPLEXITY JUST TURNED YOUR WINDOWS PC INTO AN AI WORKER It reads files, runs jobs, and keeps your private data off the cloud entirely.@Romka2xTITLE AI News: Perplexity Goes Local, ChatGPT Human Reviews Reported DESCRIPTION AI developments from the past 24 hours cover local assistants, reported human chat reviews, crowd sensing, software repairs and an autonomous truck unveiling. https://blogs.nvidia.com/blog/local-ai-perplexity-windows-pcs/ — NVIDIA https://www.404media.co/inside-project-lily-the-humans-reading-your-chatgpt-chats/ — 404 Media https://news.samsung.com/global/samsung-and-verizon-complete-successful-isac-trial-over-virtualized-network-at-global-sports-event — Samsung https://github.com/QwenLM/qwen-code/releases/tag/cua-driver-rs-v0.20.9 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11910 — Qwen Code https://github.com/ollama/ollama/releases/tag/v0.34.1-rc2 — Ollama https://github.com/ollama/ollama/pull/18376 — Ollama https://blog.pony.ai/pony-ai-unveils-new-gen-4-robotruck-in-collaboration-with-gac-commercial-vehicle/ — http://Pony.ai https://github.com/QwenLM/qwen-code/pull/11831 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11636 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11881 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11903 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11906 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11805 — Qwen Code AI-generation disclosure: This video uses AI-generated narration and visuals. TRANSCRIPT NVIDIA says Perplexity's Portable Computer can now complete work locally inside its Windows app, with no Computer credits consumed for locally finished tasks. Its September 14 announcement limits availability to GeForce RTX and RTX PRO graphics processors with at least 24 gigabytes of video memory, and lists connections including Word, Outlook, Gmail and Google Drive. For people using that hardware, NVIDIA says tasks needing cloud support ask permission before information leaves the device, giving users a decision point when their work needs to move off their computer. Hundreds of contractors are being hired to rate and critique real users' ChatGPT conversations, according to a September 14 investigation by 404 Media based on leaked documents and prompts. The publication reports that reviewers do not see usernames, while OpenAI told it the company tries to remove personal information and acknowledged that sensitive details can pass through. That means personal material in some conversations can reach human reviewers despite those removal efforts, and Anthropic also confirmed to the publication that it uses human review to improve its models. Samsung says artificial intelligence converted changes in wireless signals into a near-real-time map of crowd density during a completed trial with Verizon. In a September 14 announcement, Samsung described a radio receiving signal variations from a commercial 5G device at an international soccer fan event in Dallas, with nearby AI processing turning those readings into a heatmap. Samsung says the demonstration measures crowd density without identifying individuals and could support congestion management, a potential use for event organizers that remains a trial result rather than an established safety improvement. Qwen maintainers say their latest Linux automation repair returns an unconfirmed result when an input may already have been sent, instead of automatically replaying it. The September 15 CUA Driver release follows the previously covered Linux input repair, and its associated change also avoids traversing application-managed virtual descendants while sharing a single deadline for collecting interface information and geometry. The practical change is avoiding duplicate clicks or typing after an uncertain result, but the maintainer says omitted spreadsheet content still requires screenshots and validation did not cover native Wayland, macOS or Windows acceptance. Ollama's September 14 release candidate changes a repeated-token failure into an explicit error instead of returning an incomplete answer, with the repetition limit raised to one hundred. The prerelease also evicts active-conversation cache snapshots, checks free memory and waits for removed model processes before loading the next model through the MLX software path. For affected small graphics processors running Gemma3n, Ollama says another safeguard avoids falling back to the main processor in a way that can corrupt image representations and make the model describe the wrong image. Pony AI unveiled a fourth-generation autonomous heavy-duty truck based on GAC's T9 platform at the Hannover transportation show on September 14. The company says volume production is scheduled for later in 2026, with long-haul freight, dedicated routes and port transportation among the intended uses. Pony AI says its autonomous-driving kit costs 70 percent less than the previous generation and forecasts a 30 percent reduction in transportation costs per ton-kilometer, making the freight savings a company expectation rather than an achieved operating result. Qwen Code merged six source changes that correct skill-enabling mistakes, organize background results, repair standalone terminals, expose hook settings, clarify hook progress and output, and let extensions contribute workflows. The changes stop colons in skill names causing false extension matches, make batch enabling order-independent, coordinate when background results reach the model, show detailed hook configuration and emit internal start/end outcomes, reject malformed structured hook output as a non-blocking error, and list extension workflows during installation and updates. Developers using this source work gain clearer controls, including workflows disappearing when an extension is disabled and a standalone terminal repair verified on Linux x64, but Linux arm64 still lacks a pinned terminal package and these merges do not establish packaged application availability.@aiseomasteryPERPLEXITY JUST TURNED YOUR WINDOWS PC INTO AN AI WORKER It reads files, runs jobs, and keeps your private data off the cloud entirely.@Romka2xTITLE AI News: Perplexity Goes Local, ChatGPT Human Reviews Reported DESCRIPTION AI developments from the past 24 hours cover local assistants, reported human chat reviews, crowd sensing, software repairs and an autonomous truck unveiling. https://blogs.nvidia.com/blog/local-ai-perplexity-windows-pcs/ — NVIDIA https://www.404media.co/inside-project-lily-the-humans-reading-your-chatgpt-chats/ — 404 Media https://news.samsung.com/global/samsung-and-verizon-complete-successful-isac-trial-over-virtualized-network-at-global-sports-event — Samsung https://github.com/QwenLM/qwen-code/releases/tag/cua-driver-rs-v0.20.9 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11910 — Qwen Code https://github.com/ollama/ollama/releases/tag/v0.34.1-rc2 — Ollama https://github.com/ollama/ollama/pull/18376 — Ollama https://blog.pony.ai/pony-ai-unveils-new-gen-4-robotruck-in-collaboration-with-gac-commercial-vehicle/ — http://Pony.ai https://github.com/QwenLM/qwen-code/pull/11831 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11636 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11881 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11903 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11906 — Qwen Code https://github.com/QwenLM/qwen-code/pull/11805 — Qwen Code AI-generation disclosure: This video uses AI-generated narration and visuals. TRANSCRIPT NVIDIA says Perplexity's Portable Computer can now complete work locally inside its Windows app, with no Computer credits consumed for locally finished tasks. Its September 14 announcement limits availability to GeForce RTX and RTX PRO graphics processors with at least 24 gigabytes of video memory, and lists connections including Word, Outlook, Gmail and Google Drive. For people using that hardware, NVIDIA says tasks needing cloud support ask permission before information leaves the device, giving users a decision point when their work needs to move off their computer. Hundreds of contractors are being hired to rate and critique real users' ChatGPT conversations, according to a September 14 investigation by 404 Media based on leaked documents and prompts. The publication reports that reviewers do not see usernames, while OpenAI told it the company tries to remove personal information and acknowledged that sensitive details can pass through. That means personal material in some conversations can reach human reviewers despite those removal efforts, and Anthropic also confirmed to the publication that it uses human review to improve its models. Samsung says artificial intelligence converted changes in wireless signals into a near-real-time map of crowd density during a completed trial with Verizon. In a September 14 announcement, Samsung described a radio receiving signal variations from a commercial 5G device at an international soccer fan event in Dallas, with nearby AI processing turning those readings into a heatmap. Samsung says the demonstration measures crowd density without identifying individuals and could support congestion management, a potential use for event organizers that remains a trial result rather than an established safety improvement. Qwen maintainers say their latest Linux automation repair returns an unconfirmed result when an input may already have been sent, instead of automatically replaying it. The September 15 CUA Driver release follows the previously covered Linux input repair, and its associated change also avoids traversing application-managed virtual descendants while sharing a single deadline for collecting interface information and geometry. The practical change is avoiding duplicate clicks or typing after an uncertain result, but the maintainer says omitted spreadsheet content still requires screenshots and validation did not cover native Wayland, macOS or Windows acceptance. Ollama's September 14 release candidate changes a repeated-token failure into an explicit error instead of returning an incomplete answer, with the repetition limit raised to one hundred. The prerelease also evicts active-conversation cache snapshots, checks free memory and waits for removed model processes before loading the next model through the MLX software path. For affected small graphics processors running Gemma3n, Ollama says another safeguard avoids falling back to the main processor in a way that can corrupt image representations and make the model describe the wrong image. Pony AI unveiled a fourth-generation autonomous heavy-duty truck based on GAC's T9 platform at the Hannover transportation show on September 14. The company says volume production is scheduled for later in 2026, with long-haul freight, dedicated routes and port transportation among the intended uses. Pony AI says its autonomous-driving kit costs 70 percent less than the previous generation and forecasts a 30 percent reduction in transportation costs per ton-kilometer, making the freight savings a company expectation rather than an achieved operating result. Qwen Code merged six source changes that correct skill-enabling mistakes, organize background results, repair standalone terminals, expose hook settings, clarify hook progress and output, and let extensions contribute workflows. The changes stop colons in skill names causing false extension matches, make batch enabling order-independent, coordinate when background results reach the model, show detailed hook configuration and emit internal start/end outcomes, reject malformed structured hook output as a non-blocking error, and list extension workflows during installation and updates. Developers using this source work gain clearer controls, including workflows disappearing when an extension is disabled and a standalone terminal repair verified on Linux x64, but Linux arm64 still lacks a pinned terminal package and these merges do not establish packaged application availability.@aiseomasteryPERPLEXITY JUST TURNED YOUR WINDOWS PC INTO AN AI WORKER It reads files, runs jobs, and keeps your private data off the cloud entirely.
