AI Documentaries, HackerTube and Grok Bot: What to Try First

Andrew Warner and Corey Ganim's weekly tour puts very different products under one “AI releases” label. GoMotion produces videos; HackerTube changes what you discover; Omnisend captures willing email subscribers; Grok Bot acts for you; Beacon tries to preserve what a coding agent learned. The useful question is not which demo looks most futuristic. It is which repeated job you can test, with its limits understood.

Watch Andrew and Corey's Release Tour

Source note, 5 November 2026: The hosts' demonstrations and opinions come from the video and supplied transcript. I checked the linked product pages, first-party documentation, founder posts and repositories for scope and access; I did not run these products or reproduce their business results. The episode is presented by Zapier. Prices, releases and plan terms can change after this check.

The Releases, Organized by the Job They Solve

This is a map, not a ranking. A live product, a research prototype and a service offer need different kinds of proof.

Make Better Videos, or Find Better Ones

GoMotion: an output engine, not a YouTube income switch

The hosts show GoMotion assembling narration, real footage, captions, music and motion graphics from a script. Its developer page also documents an API and MCP connection on paid plans. A good first test is one short, original explanation of something you know well: compare factual errors, awkward cuts, clip rights and editing time against your current process. Do not turn an unreviewed script into a factual documentary.

Andrew discusses monetization as an opportunity, not a result he verified for a new channel. YouTube's channel policy allows AI-assisted creative work, but rejects mass-produced or minimally transformed material without original value. Its AI disclosure rules can also apply to realistic altered content. A finished MP4 is only the beginning of that review.

HackerTube and FeedOff: make discovery more deliberate

Andrej Karpathy's own account says he built HackerTube around 6,454 videos repeatedly recommended in Hacker News comments, organized into 150 topics. It still plays YouTube videos; the product is a different selection and browsing layer. Its Hacker News taste is useful for some readers, but it is not an objective quality score or a parental-control system. The exportable dataset is the more interesting builder lesson: a small, curated source can make a familiar interface feel new.

Open Karpathy's original post on X

FeedOff takes the opposite route: keep the social service, remove Reels, Shorts, suggested posts or other feeds that pull you away. Its site describes an iPhone app, Apple Screen Time shield and a Chrome extension, with different controls by platform. The hosts speculate about download counts from the founder's post and App Store ratings, but a rating count is not a reliable download counter. Judge it by whether it preserves the messages, subscriptions or search you need without the distracting feed.

Original announcement: FeedOff founder on X; App Store listing.

Quiet Input and an Audience You Can Reach

Subvox is an experiment; FluidVoice is a different tool

Joe Blau's Subvox repository is more modest than the on-screen excitement suggests. It uses AirPods head-motion data to test whether silently mouthed words can be distinguished from a small set of user-trained templates. The README says it is not a language model, notes a roughly 25 Hz motion stream, and warns that repositioning the headphones can change the signal. This is a research prototype, not a universal way to dictate private emails without speaking.

Open Joe's original demo on X

Andrew mentions FluidVoice as his local-first dictation option. Its README says audio and transcript stay on the device unless a cloud provider is chosen, while anonymous usage signals are still collected and detailed analytics can be disabled. It uses spoken audio rather than Subvox's head-motion experiment. For the meeting side, see our FluidVoice and FluidMeet interview.

Omnisend: a comment becomes an invitation, not automatic consent

Omnisend Social Capture responds to a chosen keyword in a comment or DM, asks for an email address and adds the person only after a valid reply. Its setup guide requires a paid plan and an Instagram business or creator account, and includes customizable consent text. The hosts compare it with ManyChat; the useful pilot is one promised resource, one keyword, one welcome email and a measured opt-in completion rate. Never treat a public comment alone as permission to market by email.

Agents Can Act, but Memory and Approval Matter

The hosts show Grok Bot's phone-call demo negotiating a service plan, and a separate shopping announcement. These are product demonstrations, not proof that every account, seller or region supports the same flow. The current Grok Bot FAQ says access depends on plan and consequential actions may require approval. Its security guide recommends explicit ask-first rules for purchases, messages and account changes. A sensible first request is to research options and draft a recommendation, then let a person approve the call or purchase.

Open the phone-call post on X · Open the shopping post

Beacon addresses a less visible cost: repeating the same debugging work. Its repository describes collecting agent-session traces across multiple coding environments, then retrieving or distilling useful patterns into skills. That is promising when an issue recurs, but the event schema shows why privacy review belongs before installation: prompts, commands, files and tool events may be recorded. Test it on a disposable repository, inspect what it keeps, and verify a generated skill before trusting it on production code.

Open the Beacon announcement on X · Product Hunt launch

The Business Angle: Improve the Last Mile

AI Designer offers design references, a canvas and an MCP route into coding agents. The hosts discuss its founder's roughly $100,000 annualized revenue run rate; Shipped & Paid's dated case lists $8,338 MRR on 2 November from a payment-provider-linked profile. That is revenue, not profit, and not a prediction for another design tool. The lesson is that better taste, integration and distribution can be sold even when basic UI generation is cheap.

Tune-Up is the service version: take an AI-built site that looks generic and make it more credible. The hosts describe a fixed-price, fast-turnaround offer, but we did not verify its current terms or any conversion outcome. A strong version would show a before/after, mobile and accessibility checks, a clear revision boundary and a human handoff. It should not promise that cosmetic polish alone fixes a weak offer or unsafe code.

Loggit, discussed by Andrew as an early shared-finance app, illustrates a different last mile: making expenses understandable to both partners. A preview is not enough to recommend putting bank or brokerage credentials into it. Check availability, data handling and export before using real records.

Which One Should You Try First?

Copy a Website Tune-Up Pilot Brief

This is our original business prompt, inspired by the episode's service idea. It asks for buyer evidence and a limited, authorized deliverable rather than promising a new agency overnight.

Video Chapters

Create and curate

Grow and design

Agents and opportunities

Sources and Related Viewing

Checked 5 November 2026. The episode and transcript establish what the hosts showed and thought. First-party pages and repositories below establish documented features and limits; founder posts are announcements, not independent performance tests.