1. Everyone Has an AI Tool Now

At this point I’m fairly sure that if you shake any group of people between the ages of 22 and 40, at least one AI startup will fall out.

I’m not exaggerating. In the last year I have been pitched an AI tool that writes wedding speeches, an AI tool that tells you which of your plants are dying and why, an AI tool for “emotional accountability”, and an AI tool whose entire function, as far as I could work out, was to be an AI tool. Somebody once described their product to me for four uninterrupted minutes and at the end I still had to ask, “Sorry, what does it actually do?” and he said, “It’s a layer.”

A layer. Of what. On what. Underneath what.

So I have developed a reflex. When somebody tells me they’ve built an AI product, my face does a thing. It’s not a nice thing. It’s the face you make when a cousin announces he’s getting back into crypto.

Which brings me to Karam.

Karamjeet Singh is a friend of mine. He’s the one who builds things instead of talking about building things, which in 2026 is practically a diagnosable condition. He doesn’t post. He doesn’t do the founder-on-LinkedIn thing where a small update takes 900 words and includes the phrase “grateful for the journey”. For a long stretch he mostly went quiet, and then one day he sent me a link to AICaptionCut, an AI caption generator and video editor that runs inside your browser, and I did the face.

He couldn’t see the face. He knows me well enough to know it happened.

Standard disclosure, so nobody has to go digging: Karam is my friend. I’m not pretending to be a neutral laboratory instrument here. But nobody paid me, nobody sponsored this, nobody sent a brief, and nobody got to read it before you did. I wanted to understand what he’d built, why he bothered, and whether there was a real idea underneath it or just another layer.

So he came over, I made bad coffee, and I opened with the only question that mattered.

“Bruv. The internet is already drowning in AI tools. Why the hell did you build another one?”

“Because I was annoyed.”

“That’s it?”

“That’s honestly it. I had a video. I wanted subtitles on it. Every way of doing that involved uploading the file to a website I’d never heard of, run by people I couldn’t name, in a country I couldn’t guess, and then waiting. For subtitles. On a video I already had, on a computer that was sitting right there doing nothing.”

“So you built a whole product because you were mildly irritated.”

“Everything gets built because somebody was mildly irritated, man. The ones that fail are the ones built because somebody read a market report.”

I want to flag something about that before we go any further.

He didn’t say “market opportunity”. We’ll come back to why that matters.

2. So What the Hell Is AICaptionCut?

My first assumption, delivered with total confidence, was: “So it’s captions. It puts captions on videos.”

“Sure,” he said. “And a car is a chair that moves.”

“Oh, we’re doing this.”

“You asked.”

Okay. Let me lay out what it actually is, because I got this wrong and you probably will too.

You open AICaptionCut in a browser tab. No download, no install the app from play store, no account, no email, no sign-up wall that appears after you’ve already done twenty minutes of work. You drag a video in. It transcribes the audio and writes captions with word-level timing, which is the thing that makes karaoke-style highlighting land on the right syllable instead of lagging half a beat behind like a badly dubbed film. If it gets a word wrong, you double-click the word and fix it.

Then you pick a style. There are 63 on the site right now, sorted into buckets like trending, bold, clean, elegant, fun and minimal, and three of them are built specifically for Hindi. On top of that there are animations: pop, bounce, typewriter, blur, wave, slide, a few others, all of which combine with the word-by-word highlighting.

That’s the captions part. That’s what I thought the whole product was.

Then he mentioned the timeline.

“Wait. There’s a timeline?”

“There’s a timeline.”

“Since when?”

“Since it needed one.”

You can trim, split, join multiple clips into a single sequence, crop to 9:16 or 1:1 or 16:9, change speed, drop text on top. There’s colour grading, including LUTs and an auto colour correct. There’s a long-video-to-shorts workflow, auto-reframe, and clip finding. You can export subtitle files as SRT or VTT, or burn the captions in and export an MP4 straight to your disk.

“Bruv, that’s an editor.”

“It’s an editor.”

“Then why is ‘Caption’ in the name?”

“Because that’s what people type into Google at two in the morning. Nobody searches for ‘browser-based local-first video editing suite’. They search for ‘add subtitles to video free’. If I name it after my ambition instead of their problem, nobody ever finds it and I get to be very principled and completely unknown.”

“So the name is SEO.”

“The name is honesty about how people find things. The captions are the front door. The house is bigger than the door. That’s fine. Most houses are.”

3. Why Captions? Why Video?

Here’s the thing about captions that I genuinely didn’t appreciate until I started paying attention.

Captions aren’t an accessibility afterthought any more. On a phone, on a train, with the sound off, captions are the video. If the first two seconds don’t have text on screen, a serious chunk of people scroll away before your first sentence finishes. The thing that used to be a chore at the end of the edit is now closer to the whole point.

So I asked Karam what he’d seen that made him pick this specific problem, when there are a thousand problems in video.

“Because it’s the one everybody does and nobody enjoys. That’s the sweet spot. You don’t want the problem people find interesting. You want the problem people find boring, because boring problems get solved badly, and badly means there’s room.”

“Everyone already had a way to do this though. CapCut does captions. Instagram does captions. YouTube does captions.”

“They do. And every single one of them makes you hand over the file first. That’s the part I think is broken. Not the accuracy, not the styling. The transaction. You’re giving a company your unreleased footage in exchange for text. That trade made sense in 2016 when your laptop couldn’t run the model. Your laptop can run the model now. Nobody updated the trade.”

“Bruv, that’s a small problem.”

“It’s a small problem that happens to every creator, every video, forever. I’ll take that over a big problem that happens to forty people once.”

He also brought up Hindi and Hinglish, which turned out to matter more than I’d assumed. The tool supports fifteen languages, and one of them is Hinglish, which is Hindi speech written in Roman letters, the way approximately every Indian creator I know actually captions a reel. Nobody writes “मैंने सोचा” (Translation: “I thought”) in a reel caption. They write “maine socha”. Getting a speech model to do that consistently is a deliberate choice, not a default.

“Why was that worth the effort?”

“Because I live here, bro. Every tool built in San Francisco treats Hinglish as a transcription error. It hears code-switching and panics. Half the country talks like that. It’s not an edge case, it’s the main case, it’s just not the main case where the funding is.”

“So you built the thing you needed and then noticed other people needed it.”

“That’s the only way anybody’s ever built anything useful. The alternative is guessing about strangers.”

4. The Weird Thing About Privacy

This is where the conversation stopped being a favour to a friend and started being interesting.

Your video never leaves your computer.

Not “we delete it after 24 hours”. Not “we don’t look at it”. Not “we’re compliant, here’s a badge”. The file is read off your disk into the browser tab, the speech-to-text runs there, the styling runs there, the final render runs there, and the MP4 lands back on your disk. Nothing is uploaded, because there’s nowhere to upload it to.

The site tells you how to verify this yourself, which I appreciated more than any privacy policy I’ve ever not read. Open dev tools, press F12, watch the Network tab while you caption something, and your file appears in no request. Or the blunter test: load the site once, switch on airplane mode, and caption a video anyway. It works. A tool that uploads your footage cannot do that.

I tried the airplane mode thing. It worked. I felt like a hacker for roughly eleven seconds.

“Okay, Bruv. Explain this properly. Why build it this way when literally everyone builds it the other way?”

“You want the noble answer or the real one?”

“Obviously the real one.”

“The real one is arithmetic. A cloud caption tool rents GPUs by the minute. Every minute of your footage costs them actual money, every single time you press go. That’s why they meter you. By length, by upload size, by credits, by ‘media minutes’. It’s not greed, it’s that their costs scale with your usage, so your bill has to scale with your usage or they die.”

“And you?”

“I render on your machine. A three-hour video costs me nothing. A thousand users exporting at once costs me nothing. There’s no meter because there’s nothing to meter. I didn’t pick this because I’m precious about privacy. I picked it because it’s the only way one person can run a video product without a GPU bill that eats him alive.”

“So the privacy thing isn’t you being noble. It’s you being cheap.”

“The privacy is a side effect of being cheap, yes. And honestly that’s why you should trust it more, not less.”

“Go on.”

“Because a promise costs nothing. Every cloud tool promises they delete your footage. Maybe they do. You can’t check. You’re trusting a stranger’s internal process. I’m not asking you to trust me, I’m asking you to press F12. I don’t have your video. I can’t leak what I never had. That’s a smaller thing to ask of you, and small asks are the only ones that hold up.”

Now, I’m not going to sit here and pretend the architecture is free of downsides, and to his credit he didn’t either. His own site has a section saying, in effect, this product is probably not for you if you need certain things. I have genuinely never seen a product page do that.

So I made him say the downsides out loud.

One: your hardware is now the bottleneck.

“If you’re on a six-year-old laptop with four gigs of RAM, you’re going to have a worse time than someone on a new machine. A cloud tool gives everybody the same GPU. I give everybody their own, and some of those are terrible.”

“How terrible?”

“Depends entirely on the machine and I’m not going to invent a benchmark for you. The honest version is: on a modern phone or laptop it’s fine and people are surprised it’s fine. On something old and stressed it’s slow, and slow on a render is worse than slow anywhere else because you’re watching a bar.”

Two: long video takes proportionally longer, because it’s running on your hardware. Feature-length files will sit there and chew. That’s on the site too.

Three: no team collaboration, no cloud projects, no shared asset library. Projects save in your own browser storage. Which is excellent for privacy and useless if you’re an agency with four editors who need the same project open.

“That one’s not a bug I’ll fix next month,” he said. “It’s the tradeoff. You can’t have ‘nothing leaves your device’ and ‘my colleague can open it’ at the same time. Pick one. I picked.”

Four, and this is the one I like him for admitting: if you export dozens of videos a day and always want them clean, a subscription tool will probably cost you less. His own pricing page says that. Out loud. On the page where he’d like you to give him money.

“Bruv, you’ve written ‘don’t buy this’ on your own website.”

“I’ve written ‘here’s who this is wrong for’. Which is the same sentence, yeah.”

“Why?”

“Because the alternative is somebody pays, discovers it’s wrong for them, and now I’ve got an angry person and a refund. And because when you tell people the truth about the limits, they believe you about everything else. That’s not a marketing trick, it’s just cheaper than lying.”

One more wrinkle worth mentioning, because it’s the only thing that does travel over the network: the speech model itself downloads once from a CDN the first time you generate captions, then your browser caches it. That’s traffic coming to you, not your video going out. But it does mean the first run needs a connection, and it does mean something is being downloaded onto your machine, which people are entitled to know.

5. Wait, How Are You Actually Making Money?

Right. The money.

This is the part I care about most, partly because this article lives in my Money section and partly because software pricing has quietly become one of the strangest financial stories of our time and almost nobody talks about it properly.

Here’s how AICaptionCut charges you, as of right now.

The editor is free. All of it. Every caption style, every animation, colour grading, auto-reframe, clip finding, trimming, the lot. Nothing held back for a paid tier. Exports are unlimited and free at 480p with a small watermark in the corner. Subtitle files, SRT and VTT, are free on every tier including the free one.

If you want the watermark gone, you pay once, for that video. The cheaper unlock exports at 720p, the fuller one at 1080p, 60fps, top bitrate. If you export a lot, there’s a 30-day Unlimited Pass, which is still a single payment rather than a subscription. Nothing renews. No card is stored. Payment runs through Razorpay at the moment you export, and you see the price in your own currency.

One note for anyone reading this in six months: the site no longer publishes fixed rupee or dollar figures, it shows you your own currency at checkout. So if you’ve seen specific numbers in some listicle, go and look at the actual product before quoting them at me.

“Bruv, why didn’t you just charge ten dollars a month like a normal person?”

“Because I’d be charging you for a cost I don’t have. A subscription is how you pay for a meter that’s running. My meter isn’t running. Your machine is doing the work. Billing you monthly for that would be, what’s the word, a scam is too strong.”

“It’s a little strong.”

“It’d be renting you something you already own. That’s the polite version.”

I pushed, because “no subscription” sounds lovely right up until you realise recurring revenue is the only reason most software companies survive long enough to fix their bugs.

“Are you sacrificing recurring revenue for customer trust? Because that’s a beautiful sentence and a terrifying spreadsheet.”

“Yeah, it is. I’m not going to pretend otherwise. One-time pricing means I wake up every month at zero. No base. No floor. Every rupee has to be earned again by somebody finishing a video and deciding the watermark bothers them. Subscription companies wake up with money already in the account and spend the month deciding what to build. I wake up and have to be useful again.”

“Does it work?”

“It works at the size I am. Whether it works at a hundred times this size is a real question and I don’t have the data to answer it honestly, so I’m not going to give you numbers to make the answer sound better than it is.”

“That’s the most founder thing you could possibly refuse to do.”

“I’d rather be boring than caught.”

Then the question I’ve wanted to ask a founder for years.

“Why is every software company on earth obsessed with subscriptions?”

“Because a customer who forgets to cancel is worth more than a customer who buys once, and everyone downstream of that fact has organised around it. Predictable revenue makes you easier to value, easier to raise against, easier to plan. And once one company in a category goes subscription, everyone has to, because now you’re competing with a business that can afford to spend more than you to acquire the same person. It’s not a conspiracy. It’s a gravity well.”

I’ll give you my read, since it’s my article.

A normal person making content in 2026 is paying rent to the internet. Count it up. An editing tool. A design or thumbnail tool. A scheduler. Cloud storage, because your laptop is full. Maybe stock music. Maybe an AI writing thing. None of them feel expensive alone. That’s the design. Ten here, twenty there, sixteen for the design one, and you’re a hundred dollars a month into software before you’ve earned a single rupee from a single video.

I said that to Karam and asked whether his model actually answers it or is just a smaller version of the same trap.

“It’s an answer for some people and it’s worse for others, and I want to be specific about which, because this is where founders get slippery. If you post twice a month, paying once per video is dramatically cheaper than any monthly plan and it’s not close. If you post every day and you always want it clean at 1080p, a subscription beats me. That’s not me being humble, that’s just where the lines cross. I put it on the pricing page because I’d rather you work it out before you pay than after.”

“So who’s actually cheaper?”

“Depends on volume. Always has. Anyone who answers that question without asking about your volume is selling you something.”

For rough context on the subscription side of the market: third-party trackers in 2026 have reported CapCut’s US web pricing at around $9.99 a month for Standard and around $19.99 a month for Pro, with an annual figure quoted near $179.99, while CapCut itself says the final price varies by region, device and promotion. Adobe’s single-app Premiere plan has been reported around $22.99 a month on an annual commitment, higher month to month. Descript’s paid tiers have been reported starting around $16 a month billed annually and climbing.

Do not treat any of those as gospel. Check them at your own checkout. That being the only reliable number is its own small indictment of how this industry prices things.

“Does being the cheap option worry you? People assume cheap means worse.”

“It worries me a lot, actually. Free is the hardest price to charge. People think free means toy. The 480p watermark is partly there to make the free tier finite, but if I’m honest it’s also there so the paid thing has a job. I’ve gone back and forth on it more than anything else in the product.”

“And what does it actually cost you to run this?”

“A domain, hosting for what is basically a static site, a payment gateway taking its cut, and the CDN for the model. That’s the whole list. Compare that to renting GPUs by the minute and you understand the entire business in one line. I’m not clever. I just have a much smaller electricity bill than my competitors.”

6. Bruv, You’re Competing With CapCut

I wanted to be a good friend, so I said it as gently as I could.

“CapCut exists.”

“It does.”

“Adobe exists.”

“Famously.”

“Canva exists, VEED exists, Descript exists, and Instagram will slap captions on your reel for free inside the app you’re already posting from. Why should any normal human being care about yours?”

“Because none of them will do it without taking your file first, and a specific kind of person cares about that a lot.”

“Which person? And don’t say ‘creators’. Everyone says creators.”

“Fine. The freelancer editing a client’s product launch that isn’t announced yet. The person on a bad connection for whom uploading two gigs is an event they schedule. The person who posts four times a year and does not want a recurring charge on a card for it. The person on a plane. The person who has simply run out of patience for making accounts. That’s not everybody. It’s not even most people. But it’s real, and most of the industry structurally can’t serve them, because their whole cost model requires the upload.”

“Okay. Here’s the one that matters. CapCut ships on-device captioning tomorrow. Are you finished?”

“If they do it properly and put it in front of a billion people, I’m in a lot of trouble and I’m not going to pretend I have some clever answer. What I’d actually have is a head start and a reason to think they won’t do it properly.”

“Which is?”

“It costs them something to. Their storage tiers, their upsell, their account system, their retention numbers, all of it depends on the file being on their computer. Shipping a genuinely local version means giving up the meter they built the company around. They can. It’s just expensive for them and it’s free for me, because I never had the meter.”

“That’s a head start, not a moat.”

“Yeah. I know. Anyone who tells you they have a moat has been talking to investors. I have a head start, a very small cost base, and the ability to ship a change on a Tuesday afternoon without four meetings. That’s what I’ve got.”

“Is price your advantage?”

“No, and I’d be dead if it were. Someone with more money can go cheaper than me until I starve. Price is a consequence of the architecture, not a strategy.”

“So what can a one-man operation do that a giant genuinely can’t?”

“Be unprofitable on purpose in a specific place. Three Hindi caption styles are not a business case for anybody with shareholders. Answering support email personally is not scalable and that’s exactly why it works. And I can make a product that deliberately doesn’t collect anything, which is structurally impossible for a company whose valuation depends on an account system. Not because they’re evil. Because the org chart won’t let them.”

7. What Can This Thing Actually Do?

At some point I stopped interviewing and started just using it while he watched, which I recommend to anyone who wants to make a founder visibly uncomfortable.

I dragged in a clip from one of my sets. Bad lighting, decent audio, me talking far too fast. The kind of thing that breaks transcription.

Captions came up fast. Word timing was tight enough that the highlighting landed where it should. It missed a couple of words. I double-clicked them and fixed them, which is the correct amount of friction for fixing a word.

Then I went through the styles and lost about twenty minutes of my life. There are 63. The site says every preview is drawn by the same renderer that produces your final export, so what you see is what exports, which matters more than it sounds, because half the tools I’ve used show you a gorgeous preview and then export something rendered by what appears to be a fax machine.

I tried Cinematic. I tried Slam. I tried Pop Art, which made my material look considerably more fun than it is.

“Bruv, some of these are horrendous.”

“Some of them are horrendous, yeah.”

“Then why are they in there?”

“Because ‘horrendous’ is a taste opinion and taste is not universal. The ones I think are ugly are the ones that get used most. I built three styles I’m actually proud of and nobody touches them. That was a humbling month.”

Then I found the long-video-to-shorts workflow and the clip finding, and this is the feature I’d genuinely use, because the most tedious job in my life is watching back forty minutes of a set to find the ninety seconds worth posting.

“How good is the clip finding? And I want the real answer, because everyone overclaims here and the overclaiming is why nobody trusts these features.”

“It finds you candidates. It does not find you the funny bit. It’s working off what was said and where the energy is, and it’s decent at spotting a self-contained chunk with a beginning and an end. It cannot tell that the laugh at 14:20 is the payoff for a setup at 3:50. You still have to be the one who knows that. Treat it as a shortlist, not a decision.”

“That’s the most honest description of an AI feature I’ve heard this year.”

“It’s also the one that makes people actually use it. If I tell you it finds your best moment, you’ll test it, it’ll fail once, and you’ll never open it again.”

Auto-reframe was the other one. Shoot in 16:9, it crops to 9:16 and keeps the subject in frame rather than lovingly framing your left shoulder for ninety seconds.

“Where does it fall over?”

“Two people talking. It has to pick, and sometimes it picks wrong, and sometimes it picks correctly and then changes its mind. Fast movement out of frame is the other one. Single speaker, roughly centred, it’s fine. That covers most reels, which is why it’s in there.”

Colour grading is in there as well, including LUTs and an auto colour correct, which I did not expect to find in a thing called a caption tool.

And the whole thing runs on a phone. It installs to the home screen like an app, with a separate touch layout where the timeline is laid out for a thumb instead of a mouse.

“Okay. I’ll say it. That’s more than I expected.”

“Say it louder.”

“No.”

8. Is AI Helping Creators, or Just Helping People Make Crap Faster?

Now the fun part.

I put it to him bluntly: every barrier that used to exist in video has been dismantled, one by one, and the result has not obviously been a golden age. It’s been more content. An ocean of it, most of it indistinguishable from the rest.

“Bruv, be honest. Is AI helping creators, or is it just letting mediocre people be mediocre ten times faster?”

“Both, and I’m part of the problem, and I’ve made peace with that.”

“That’s a hell of an opening.”

“I’d rather say it than have you say it. Yes, there’s more rubbish. There was always going to be more rubbish. But the barrier never filtered for quality, man. It filtered for money and time. It never stopped bad creators. It stopped poor ones. The guy with a MacBook and a subscription was never funnier than the guy with a phone and no money, he was just louder.”

“And the flood?”

“The flood is real and I think it’s survivable. Every time production gets cheaper, there’s a glut, everyone panics, and then taste reorganises around whatever the machines can’t do. Happened with photography. Happened with desktop publishing. It’ll happen here.”

“Are you going to put video editors out of work?”

“Some of them. I’m not going to do the thing where I say AI won’t replace anyone, it’ll just empower everyone. That’s a line people say when they don’t want to finish the thought.”

“Finish it.”

“If your job is executing somebody else’s instruction, cut the silences, sync the captions, crop it vertical, match the colour, that job is getting eaten and a lot of it already has been. That work has a correct answer, and computers are excellent at things with correct answers. If your job is deciding what the video should be, you’re fine, and you might be better off, because the boring half of your day is disappearing.”

I ran my own theory past him. AI is good at the parts of editing that are essentially transcription of intent. What it’s bad at is the part with no correct answer. How long to hold on a face after someone stops talking. Whether to cut on the laugh or a beat after it. Whether the joke lands better at 0.9 speed. Which of two nearly identical takes has the thing in it.

“That’s roughly where I am, yeah. There’s no waveform for ‘hold this one second longer because it’s uncomfortable and the discomfort is the point’. That’s taste, and taste is just accumulated guesses about what an audience will feel. You can’t extract it from audio. You can only get it from having been wrong in front of people a lot.”

“Which is my entire career.”

“Which is why you’re the one who should be cutting your own sets, and why I’d like to make the boring ninety percent faster so you spend your time on the ten percent that’s actually you.”

“Last one. Do audiences get tired of it? Is there a point where people start actively preferring the rough, obviously human thing?”

“I think there’s already a premium on evidence that a person was there. Not rough for the sake of rough, people aren’t nostalgic for bad audio. Evidence of a specific human making specific choices. As the polished stuff gets free, the signal moves. In five years I think the creator economy looks like a very small number of people with genuine taste, a very large amount of machine-assembled filler, and audiences who’ve got extremely good at telling them apart in about two seconds.”

9. AICaptionCut Is Not the Endgame

About two hours in he said something that reframed the whole conversation, and it’s the reason this article exists rather than sitting in my drafts.

He wants to change how video editing works. Not add a tool to it. Change it.

So naturally I made fun of him.

“Bruv, ‘I want to revolutionise the industry’ is the single most founder sentence in the English language. It means nothing. My uncle wants to revolutionise the industry and he sells generators.”

“Fine. Then let me say it without the word.”

“Please.”

“Right now, editing software is a set of tools you operate. Timeline, razor, keyframes, effects. You learn the tools, then you translate what’s in your head into the tools, and the translation is most of the work and all of the barrier. What I want is for the software to understand what’s in the footage, so the translation gets shorter. Not so it makes the video for you. So that when you know what you want, saying it takes ten seconds instead of two hours.”

“Give me an actual example or I’m binning this section.”

“You’ve got forty minutes of a set. You want the bit about your landlord, vertical, captioned, with the dead air gone. Today that’s you, scrubbing, for an hour, and then thirty minutes of fiddling. I want that to be one instruction and then twenty minutes of you doing the part that needs you, which is deciding whether it’s actually funny and where it should end.”

Here’s the thing I started to understand as he talked, and it’s why the name is slightly misleading.

Captions are the wedge, not the product.

Think about what captioning a video requires you to have already built. Audio transcribed. Word-level timing. Knowledge of when things are said. The ability to render text over video, frame by frame, on somebody else’s machine, fast enough that they don’t leave. Build that and you haven’t built a caption tool. You’ve built a machine that understands what’s in a video and can re-render it locally.

Once you have that, clip finding isn’t a new feature. It’s a consequence. Auto-reframe is a consequence. Cutting silences, finding the section where somebody says the thing, assembling a rough cut, it all sits on the same foundation.

I put that to him.

“That’s about right, but I want to be honest about the order. I didn’t draw that on a whiteboard and then build towards it. I built captions because captions annoyed me. Then I noticed that once the transcript and the timing exist, half the other features are nearly free, and the roadmap kind of revealed itself. It looks like strategy in hindsight. At the time it was mostly me noticing what was already sitting there.”

“So what comes next? And not the roadmap slide. What do you want this to become?”

“I’m not going to announce features I haven’t shipped, because that’s how you end up with a comments section asking where the thing is. What I’ll tell you is the direction: more of the edit, less of the upload. Everything that’s currently a manual chore on a timeline, done locally, on your machine, without an account. That’s the whole thesis and it doesn’t change.”

“If someone handed you an unreasonable amount of money tomorrow?”

“Better models running locally. That’s it, that’s the answer. The entire ceiling on this idea is what will run on a normal person’s laptop, and that ceiling has been moving fast in my favour for three years. I’d spend all of it on making the on-device side better, because if that keeps improving, everything else I want becomes possible without me doing anything clever.”

“And in ten years, if you get everything you want?”

“Somebody sits down with two hours of footage and leaves forty minutes later with six good videos, having made maybe fifty actual decisions, all of them creative. No upload. No subscription. No account. The machine did the typing and the person did the thinking.”

The bit I’ll flag for you, reader, is this. Whether or not you ever use AICaptionCut, the argument underneath it is worth understanding. The last decade moved everything to the cloud because that’s where the money was, and a chunk of that migration was never actually necessary for the user. Your laptop is astonishingly powerful. Your phone is astonishingly powerful. The reason your video travels to a server in Virginia to have text put on it is not physics. It’s a business model.

If that’s right, and if the models keep shrinking, then an entire category of monthly bills is sitting on a foundation that could give way.

That’s a much bigger claim than “I made a caption tool”.

10. Okay, But What If Google Just Does This?

Back to reality, because I enjoy reality.

“Google ships this inside YouTube Studio next quarter. Free. Right where people already are. What happens to you?”

“I lose most of the casual users almost immediately. That’s just true. A default beats a better product roughly always. What I’d keep is the people who specifically don’t want their footage anywhere near a platform, which is a smaller business but not a zero business.”

“Which competitor actually keeps you up?”

“Not the big ones. The one that keeps me up is some person exactly like me who has the same idea and is slightly better at getting found. Giants are slow and predictable. Another obsessive with a laptop is neither.”

I want to name the structural problem honestly, because I don’t think it’s small.

Distribution beats product about nine times out of ten. A worse tool inside an app with a billion users will out-compete a better tool that you have to go and find. That’s not a moral fact about the world, it’s just how defaults work. And a one-person operation with no team, no ad budget and no investors is up against companies that can lose money in this category indefinitely purely to keep you inside their ecosystem.

So the question isn’t whether the product is good. It’s whether good is enough, and historically it often isn’t.

“Fair,” he said. “My honest distribution plan is search and word of mouth, which is a polite way of saying I don’t have a distribution budget. People have a specific problem, they type it into Google at midnight, and I try to be the thing that comes up and then actually works. That’s slow and it’s unglamorous and it’s the only channel I can afford. If that doesn’t compound, I’ve got a nice tool and a very small audience, and I’d rather find that out than pretend.”

11. The Questions Founders Usually Dodge

This was my favourite section to ask and, I suspect, his least favourite to answer. I warned him in advance that if he gave me a PR answer I’d print the PR answer and label it as one.

“What does it currently suck at?”

“Long video. If you throw an hour-long file at it, you’re going to sit there. It’ll work, but you’ll feel every minute of your own hardware doing the job. And anything involving more than one person on the same project. There’s no version of that story that’s good right now.”

“Which feature are you least confident about?”

“Clip finding. It’s useful and I stand by it, but it’s the one where the gap between what people hope it does and what it does is widest, and that gap is where disappointment lives.”

“Dumbest business decision so far?”

“Spending weeks on caption styles I personally thought were tasteful instead of the ones people were obviously asking for. I was designing for a version of the user who was really just me with better opinions.”

“What did you underestimate?”

“How much of building a product is explaining it. I thought the hard part was the render pipeline. The hard part is a sentence on a homepage that makes a stranger understand in four seconds why they should care.”

“What did you overestimate?”

“How much people care about privacy in the abstract. They care enormously in the specific, the second it’s their client’s unreleased footage. In the abstract they mostly shrug. I built a lot of my messaging around a principle when I should have built it around a situation.”

“Biggest misconception about the product?”

“That ‘runs in your browser’ means toy. People assume browser means weak, limited, the demo version of a real thing. It’s the single most expensive assumption I have to fight.”

“Who should absolutely not use it?”

“Agencies. Teams. Anyone who needs to hand a project to a colleague. Anyone editing feature-length. And anyone exporting ten clean videos a day, because a subscription will cost you less and I’d rather you spend the money there than resent me.”

“What almost made you quit?”

“There was a stretch where I’d fixed the hard technical thing, the thing I’d been stuck on for weeks, and I shipped it, and nothing happened. No spike, no emails, nothing. Silence after a technical win is worse than any bug. A bug at least tells you someone’s there.”

“What scares you about AI? Not the sci-fi. The boring practical stuff.”

“That it makes the floor so cheap that nobody can build a business above it. If everything I do is free inside three other products in two years, the thing I’m good at stops being worth money. That’s not an existential fear, it’s a Tuesday fear, and it’s much more likely.”

“What scares you about running a software company?”

“That it’s just me. If I get ill, the product doesn’t get better. Nobody covers for me. That’s the actual cost of staying small and people romanticise it far too much.”

“One thing you wish you’d known before starting?”

“That the build is the easy part and I was using it to avoid the hard part, which is telling people.”

“What would you do differently if you started again tomorrow?”

“Ship something ugly three months earlier and let people tell me what it should be.”

12. So What Happens Next?

Near the end I asked the question I ask everyone who builds something, because the answer tells you whether they’re building a company or performing one.

“What does success look like? A number or a picture. Not a feeling.”

“Enough that it pays for itself and pays for me, and I get to keep working on it without having to raise money from someone who’d want it to become a subscription. That’s genuinely it. It’s not a glamorous target and it’s the only one I actually want.”

“What would make you shut it down?”

“If somebody shipped the same thing, properly local, properly free, inside a product people already use, and did it well. Then the job’s done and I’d rather go and be annoyed by something else than defend a worse version out of pride.”

“Do you want a team? Standups? A Slack?”

“A couple of people eventually, maybe. Not a company with a mission statement. I’ve seen what happens to products when the headcount needs feeding.”

“If someone reading this wants to be useful, what’s the actual useful thing? And don’t say share the article.”

“Use it on something real and email me when it gets something wrong. Not a nice message, a specific one. The language it mangled, the style that broke, the phone it was slow on. Support goes to me, not a ticket queue, and that’s the only reason the product has improved at all.”

13. Ending

We talked for a lot longer than either of us planned, which is how you know an interview went well and also how you know I have no scheduling discipline.

Somewhere near the end I realised my opening question had been slightly unfair. “Why the hell did you build another AI tool” assumes the category is the problem. It isn’t. Nobody complains there are too many restaurants. They complain that most of them are bad.

The AI flood isn’t really a flood of tools. It’s a flood of layers, of the thing my four-minute friend was pitching, of products that exist because a model exists rather than because a problem exists. The test was never whether something uses AI. It’s whether you can say what it does in one sentence, whether that sentence is about you instead of about the technology, and whether the price bears any relationship to what it costs to run.

By that test, AICaptionCut passes, for now. Whether it survives contact with companies four thousand times its size is a completely different question and neither of us knows the answer.

So I asked him one last time, at the door, with my shoes half on.

“Bruv. Final answer. Why’d you build another AI tool?”

“I told you. I was annoyed.”

“That’s still it?”

“That’s still it. Everything after that was just me refusing to stop being annoyed for long enough to get a real job.”

Then he asked if I was going to caption this interview when I posted it.

I said I’d think about it.

He said he had a tool for that.

I told him to get out of my house.

Discover more from niranjanpathak.info

Subscribe to get the latest posts sent to your email.