Skip to content
MakerGain
← Back to project

Ten million TikTok views, roughly one signup per 10,000: two dads with physics backgrounds turn commute scrolling into five-minute courses

Edited by MakerGain

Two dads with physics backgrounds turned commute scrolling into NerdSip. The company says any topic can become a five-minute course in about a minute. “Any topic” is both a selling point and a generation-and-review burden. Ten million views do not make a growth legend: roughly ten thousand views yield one signup, while revenue figures from different dates and sources do not form a continuous curve. The open question is whether users will return a second and third time.

On their daily commutes, Philip Sergelius and his cofounder Max each spent about half an hour in their phone feeds. Philip says in the interview[1] that they came away with nothing. Both have academic physics backgrounds, both work full-time in semiconductors, and each has two children. Product-building time was limited to early mornings, evenings, and weekends. They turned that unrewarding scrolling time into an app: enter a topic and get a course back in about a minute.

The app is NerdSip. Its website[2] describes an AI microlearning app built by two physics PhDs in northern Germany, offering 5-minute courses on iOS and Android. A user enters a topic—“how do black holes form?” for example—and the engine retrieves information before organizing it into 3, 5, or 10 short lessons. The package includes illustrations, quizzes, key takeaways, and a summary infographic that can be saved.

Image: From the official product display[3], obtained on 2026-11-18. Website promotional interface, used only to explain the product; effectiveness was not tested.

“Any topic” is the strongest selling point and the heaviest burden

Philip says they could have chosen a fixed category, like language-learning or book-summary apps, but instead let users enter any topic and generate a complete course. The advantage is that users do not have to fit into a category first. The cost is having no prewritten fixed catalog: courses must be checked both during and after generation.

The features page[4] spells out the sequence. Courses are generated from live search results, not written from model memory. Courses in the library are scored automatically every night on topic quality, engagement, and completeness; those that fail do not enter the swipe-card pool. Users can report something wrong with one tap, sending it to a queue “a real person will handle.” That last layer identifies only a real person, not who. The site has no answer for what happens to that layer as report volume rises, who handles it, or how increased volume is managed.

A game layer sits beyond the courses: streaks, leaderboards, friends, and item drops. Philip says he spent two weeks on a 3D character-and-equipment system that “nobody uses, but it's cool,” and thinks every app needs an overengineered feature nobody uses. The unavoidable review burden and the fun feature he wanted to build are choices made by the same small team.

Courses can be read aloud in podcast mode. Holding the microphone within a course lets users ask questions directly; the tutor answers about the current card and can be interrupted like a person. As of August 2026, individual conversations were capped at 2 minutes, with about 10 minutes per day on paid plans.

His reason for choosing Google models was practical: one API key covers LLMs, images, video, speech, and embeddings, avoiding separate integrations with multiple providers. He did not disclose what each type of call costs.

Who pays, and how much?

The free tier allows 1 sample course per day without an account. The website lists Full Access at €7.99/month or €49.99/year, with a 7-day free trial, 75 generated courses per month, 25 infographics, and unlimited audio lessons (website[2]). The US App Store separately lists NerdSip Gold at $7.99/month or $49.99/year, including a 7-day free trial. No payment is required during the trial; afterward the Apple account is charged unless auto-renewal is turned off at least 24 hours before the current period ends (App Store[5]). These are different channels and currencies, and the available sources do not say they are the same tier. Reading them as direct currency conversions would distort the comparison.

Ten million views, and roughly one signup per ten thousand

Philip says blog SEO brings more than 40% of new users, AI search accounts for 28%, and the remaining sources are scattered. Videos across four TikTok accounts have more than 10 million total views, converting at roughly 1 new account per 10,000 views. Recently, TikTok delivered 60 new signups in a single day for the first time. Total ad spend was under $500, split across small Reddit, TikTok, and App Store tests, none of which outperformed organic traffic. They also spent a few hundred US dollars buying backlinks from premium directories.

Putting ten million views beside “roughly one signup per ten thousand” at least shows that views are not a measure of channel efficiency in this business. It is not enough to write off TikTok either: attribution details are missing, and the day with 60 new signups happened only recently.

Revenue has two reference points. In the interview[1] dated August 5, 2026, Philip reports US$3,000 in July 2026 revenue. The same passage gives 10,000 downloads, 150 reviews averaging 4.8 stars, and just over 100 paying subscribers. The two wrote their first line of code in December 2025 and launched on March 1, 2026. The third-party TrustMRR record[6] shows $3,844 in trailing-30-day revenue, $1,970 MRR, and 384 active subscriptions as of October 30, 2026. That page also says the RevenueCat Legacy connection is unavailable and an API-key migration is needed to sync current revenue. Different sources, different dates, and a broken connection between them: those numbers do not make a continuous growth curve.

Revenue is not profit either. Costs include hosting, AI inference, and marketing; backlinks are treated as curated-content spending. The interview gives only ad spending below $500 and a few hundred US dollars for directory backlinks, not specific hosting or AI-inference amounts. Complete costs remain unclear.

Who appears when someone looks for an alternative?

Wondering has published its own comparison page[7]. Both products use AI to generate courses on any topic, but in different forms. NerdSip offers fast courses of 3, 5, or 10 lessons that can be browsed in minutes. Wondering generates complete courses from topics, links, or notes, followed over days or weeks, with 3-minute lessons. NerdSip's website says it has an Android app; Wondering is on the web and iPhone. NerdSip Gold costs $7.99/month with up to 75 new courses per month; Wondering Pro's web price is $14.99/month. The price gap is there, but they sell different timescales: one can be finished before your coffee cools; the other unfolds week by week.

NerdSip compresses “learn something” into a swipe, then uses games and voice to stretch that swipe. What really determines the size of the business is whether users will open the app for that action a second and third time.

Continue reading

AI video
Pexo shelved a well-received beta and gave users a way to mark up the video

Pexo plans and produces videos from a sentence or a PDF for small businesses and creators making science explainers and other educational content. Once users see the first cut, they can mark a shot or element and ask for changes. A November 12, 2026 interview reported Pexo's ARR at more than \$1 million.

AI development
Customers took 30 to 60 days to pay. Humyn Labs used \$21 million in venture debt to pay workers before customers paid

Humyn Labs records and labels videos of people doing physical tasks for AI labs that train robots. Customers pay by the hour of footage delivered. AI labs usually take 30 to 60 days to pay, and Humyn pays contributors daily or weekly. Its co-founder reported \$10 million in revenue from April through October 2026.

AI search
How WordTrace used daily phone recordings to deliver for a carmaker

WordTrace monitors brand mentions in generative engines, reruns questions, and tracks citations. Changsha Evening News reported on November 10, 2026 that WordTrace had upwards of 70 paying customers after roughly a year. A major Chinese carmaker asked for daily screenshots, photos, and screen recordings to prove the results appeared on a phone. The report kept the carmaker unnamed.

AI games
Making a Game With AI Paid for the Tools. What Were the 1,000 Hours Worth?

Solo developer zweipilot released the retro arcade game Pixel Darts: From Pub to Glory on Steam. He reported selling 202 copies in nearly ten weeks, enough by his estimate to cover AI tools and software. That leaves about 1,000 hours of work unpriced.