Muse Saves Money, Jev Gets a Voice: 10 AI Releases Worth Testing

What Matters This Week

The common thread across these releases is not a smarter chat box. It is AI moving into the interface where work already happens. Muse is handling financial errands, ChatGPT Voice can use connected apps, Sider Omni sits beside a Mac window, Clueso turns a rough recording into a finished explanation, and Glance Speedlab tests live camera understanding locally. Jev shows how a fast decision model can sit underneath voice-driven computer control, while NewsJack packages a repeatable PR process as agent skills.

The episode also shows the limits of launch-week evidence. Reported Muse savings are individual stories, the Jev computer-control app is a demo rather than an open repository, and the Systo store is a prototype. The reproducible projects are narrower: NewsJack publishes its skills, Glance Speedlab publishes its measurement setup, and Clueso documents the editable workflow it sells.

Watch the Original Episode

Credit: this guide analyzes Meta's Muse makes you money. And more, published by The Next New Thing on 24 September 2026, with Andrew Warner and Corey Ganim. The supplied transcript shaped the summary; official product pages and project repositories were checked separately.

The Release Map

Muse Is Selling the Outcome, Not the Model

The Muse stories are compelling because the output is measured in dollars rather than tokens. The episode cites public examples of users reporting a cheaper insurance policy, an unused subscription, unclaimed money, and lower medication costs. Singularity Research collected roughly 100 examples, which helps identify repeated use cases, but it does not turn every reported saving into an independently audited result.

Meta's product position is broader: Muse is a personal agent that can research, browse, use connected services, and ask for approval before consequential actions. Voice reduces the effort required to start. That matters for people who will never learn prompting syntax, but it also makes the approval screen more important. A spoken request to "fix my insurance" can change coverage, price, cancellation state, and future liability.

Start with a reversible financial audit: identify duplicate subscriptions or compare a bill, then ask Muse to show the evidence and proposed next step. Keep purchases, policy changes, cancellations, form submissions, and messages behind explicit approval. Muse is currently free in supported access, but pricing and availability can change.

The Jev Demo Is a Decision Layer Wrapped in an App

In the demo, a Mac user speaks a command and the computer reacts. The important distinction is architectural. TypeSafe describes Jev as a System One decision model: software sends a structured question and receives typed choices with probabilities and confidence estimates. Jev does not supply the microphone, screen capture, mouse control, or action policy by itself.

The surrounding app turns speech and screen context into candidate actions, asks Jev to choose, and executes the selected operation. That can feel instantaneous because the decision is narrow. It is a promising pattern for menus, routing, filtering, game inputs, and repeated desktop actions. It is not evidence that a classifier can plan any arbitrary computer task.

Clueso and NewsJack Productize Two Messy Workflows

Clueso takes a screen recording or slide deck and produces a cleaner script, AI voiceover, zooms, captions, branded video, and step-by-step documentation. Every stage remains editable. This is a good example of constrained automation: the source recording defines the task, the transformation is inspectable, and the finished artifact can be reviewed before anyone publishes it.

NewsJack applies the same productization idea to public relations. Its open Markdown skills cover strategy, newsworthiness, angle and headline generation, fact checking, journalist fit, news search, triage, monitoring, and coverage tracking. The repository documents which skills work in chat and which need a local agent or external data source.

The episode shows Jev filtering a large news set before a more capable model or person writes the final pitch. That is the right division of labor: use a fast classifier to remove obvious mismatches, then preserve human review for factual claims, journalist relevance, tone, and sending. Automating research can increase quality; automating indiscriminate outreach only increases spam.

ChatGPT Voice Can Now Reach Connected Work

OpenAI's 23 September release note says Voice now supports plugins and connected apps on web, iOS, and Android. In Work, voice can create documents, presentations, and spreadsheets, use connected apps, or continue an unfinished task in text after the call ends.

This is more than dictation. A spoken request can draw from email, calendars, Slack, and other tools available to the account. Availability still depends on the plan, app version, workspace controls, connected-service authorization, and the permissions held by the user. Andrew initially could not access the feature until updating the app, which is a useful reminder that launch-day availability is not uniform.

Test with a read-only job first: ask for a briefing from the next meeting, relevant email, and approved Slack channel. Verify each source. Add write actions only after the team can see what Voice accessed and where an approval appears.

Games, Window Sidebars, and Live Camera Understanding

Top Tennis is a useful vibe-coding case because it reached a playable multiplayer experience with tournaments and several game modes. The lesson is not that every game now takes three days. It is that a focused creator can prototype mechanics, networking, and distribution quickly enough to test whether the game is fun before building a large team.

Sider Omni takes a different interface approach. It attaches an AI sidebar to a Mac window, can keep separate context per window, and lets the user select among models. Sider says conversations and artifacts remain in a local agent runtime, while screen access is task-triggered and controlled through separate macOS Accessibility and Screen Recording permissions. Those claims are worth checking against the installed permissions and privacy policy before exposing sensitive work.

Glance Speedlab is the most inspectable camera project in the roundup. It runs a local camera-inference test bench, supports simple questions and choices, and records timing metadata only when enabled. Frames remain in memory and are sent to the selected local backend. The repository also publishes an important negative result: experimental 2B and 48-token profiles did not pass every fixed quality guardrail.

That makes it useful for low-risk triggers such as posture reminders, presence detection, or a prototype reaction game. It should not quietly watch employees, infer sensitive traits, or make disciplinary, medical, or security decisions from a webcam.

Systo Commerce and Opus 5.5 Show the Creative Ceiling

Systo Commerce turns a storefront into a walkable 3D environment with products, avatars, and advertising space. It is an imaginative prototype for discovery-led commerce, but a conventional product grid may still win on search, accessibility, mobile performance, comparison, and checkout speed. Test the experience against completed purchases rather than visual novelty.

The episode closes with Claude Opus 5.5 because several of the creative demos were attributed to it. Corey describes the model as particularly strong for writing, visual design, and animation. That is a practitioner's assessment. Anthropic's official release says Opus 5.5 performs at Fable 5.1 level on most work, costs about 40% less than Opus 5 on typical workloads, and generates output more than 30% faster than Opus 5. It lists $4 per million input tokens and $20 per million output tokens.

Use Opus 5.5 where a more polished first result can reduce review cycles. For routine classification or high-volume filtering, a specialized model such as Jev may be a better economic fit. One model does not need to win every layer of the workflow.

What to Test First

1. Low risk: transform an artifact. Give Clueso one rough internal tutorial and compare editing time, accuracy, and publish readiness with the current process.
2. Inspectable: run an open project. Review NewsJack's skills or Glance Speedlab's local privacy boundary, then test one bounded task with recorded inputs and results.
3. Read only: connect context. Use ChatGPT Voice or Sider Omni to summarize information without sending, buying, deleting, or changing anything.
4. Decision layer: add confidence gates. Put Jev in front of a queue where a wrong choice can be reviewed and recovered.
5. Consequential agent: require approval. Let Muse research a saving, but personally verify the quote, terms, and final account change.

For every pilot, measure accepted outputs, review time, retries, total cost, permissions used, and the worst plausible failure. A fast demo becomes a useful system only when the team can explain what it saw, what it decided, what it changed, and how to undo it.

Video Chapters

Sources and Useful Links

• The Next New Thing: Meta's Muse makes you money. And more; episode resources
• Meta: Muse overview; Muse; Singularity Research's user-story collection
• TypeSafe AI: Jev; voice-controlled Mac demo
• Clueso; NewsJack repository; NewsJack launch demo
• OpenAI: ChatGPT Voice; Voice plugins announcement
• Top Tennis; creator demo
• Sider Omni
• Glance Speedlab repository; project demo
• Systo Commerce prototype; Systo AI build post
• Anthropic: Claude Opus 5.5; PDoomVideo; Corey Ganim's Opus 5.5 example

The video was published on 24 September 2026. Product pages, repositories, and official release notes were checked on 25 September 2026. Availability, pricing, permissions, and beta behavior can change. Verify the live product and account controls before connecting sensitive data or allowing actions.