They did it...

About This Video

A 13-minute video from Matthew Berman arguing that consumers are about to build their own software. He starts with a personal story about disliking Adobe Acrobat, then lays out three conditions he believes have now been met, shows examples of cloned and custom software, and addresses the maintenance objection. Figures and product names below are as he says them aloud and the auto-generated transcript spells some names phonetically, so check them against primary sources before relying on them.

Summary

Berman's thesis is that the cost of building and maintaining a piece of software with AI has dropped below the cost of buying it. He says three things had to line up: models had to be good enough, cheap enough, and easy enough to use through autonomous agents. He believes all three came together over the last month. He supports this with his own habit of testing models by asking them to clone Microsoft Excel, which the latest models now do in about 90 minutes to a couple of hours, and with projects such as a free PDF reader and an open-source Adobe-style suite. He concludes that software will tend to become open source and that people will build exactly what they need.

Why Build Your Own Software

• The trigger (00:00). Setting up a new PC, Berman needed to read a PDF. Windows lacks a good built-in reader, so he installed Adobe Acrobat and disliked it: bloated, full of features he did not need, pushing other products, and asking him to log in.
• The decision (00:41). Instead of putting up with it, he made his own PDF reader, then started building other software he needed.
• Others doing the same (01:03). He says someone cloned the entire Adobe suite, including Premiere and After Effects, and he expects consumers to build their own software more and more.
• Three conditions (01:24). Models must be good enough, cheap enough (burning tens of millions of tokens to replace a $10 to $30 product makes no sense), and easy enough, which he says agent-wrapped models now are.
• The turning point (01:54). He calls Opus 5.5 the peak, the point at which he could tell a model to make or clone a piece of software and it would do it. Falling token prices and falling cost to complete a task mattered as much as quality.

The Build-Versus-Buy Calculus

• The comparison (02:14). Weigh the cost of building and maintaining custom software against the cost of buying it, often as an ongoing subscription.
• His claim (02:38). He believes the token cost to build and maintain software is now lower than simply buying it.
• Exact features (02:50). Building for yourself means you get precisely the features you want, even ones no one else would care about.

Testing Models by Cloning Software

• A ladder of tests (03:06). Berman has tested models for a couple of years. He began with Snake in Python, moved to a Rubik's cube simulation as models improved, and then switched to rebuilding software from scratch.
• Clone Excel (03:46). He prompts a model with a simple instruction to clone Excel. Older models, around GPT-5 and before the model he names Fable, got close but with rough edges and missing features.
• The latest generation (04:11). The newest models clone Excel end to end, with every feature, in about 90 minutes to a couple of hours rather than days. He made it browser-based and says he will link the latest version.
• The pricing contrast (04:38). He cites standalone Excel at $180 for a lifetime license, or a subscription of about $7 to $10 per month that includes Word, PowerPoint, and more. He says he spent maybe $10 in tokens to rebuild it with full feature parity and can open source it for free.
• Value heading to zero (05:17). He argues the value, and the cost, of software is approaching zero, because anyone with a $200 per month subscription to Claude or ChatGPT can build and release software, and others can fork it.
• Open source by default (05:49). He compares the pattern to open source and says he believes all future software will be de facto open source.

Sponsor Segment

From roughly 06:14 to 07:16 Berman describes the hardware he uses to build all of this, a Microsoft Surface Laptop Ultra powered by Nvidia RTX Spark. He says Nvidia now makes both the CPU and GPU (a Grace CPU and a Blackwell RTX GPU), with up to 128 GB of unified memory and native CUDA. He says he runs local models, plays games, and edits video at the same time, and that it stays fairly cool and quiet. He mentions a 15-inch touchscreen and a link in the description. Treat this as a promotional segment.

Decompiled Clones and the Earlier Thesis

• ArtCraft (07:21). The most well-known example he cites is ArtCraft, an open-source and free Adobe-style suite with Photocraft (Photoshop-like), VectorCraft, and FilmCraft (Premiere-like). He says he covered it in an earlier video.
• The method (07:45). The creator pointed AI at existing software and told it to iterate until it cloned the software exactly.
• SaaS Is Dead (08:07). A couple of years ago Berman made a video proposing that enterprise software would collapse into agents plus a traditional database layer holding ground truth. He says he never expected consumers to build their own software, but now the math favors building over buying.

Berman's Own Projects

• Folio (08:53). A PDF reader and editor that supports e-signing, drawing, notes, and AI questions about the document. He describes it as lightweight, fast, free, and open source, and says it can be forked.
• Outlier Tracker (09:29). He used a tool he calls One of Ten to see how other creators' videos and channels perform, but says it added features he did not need, required him to authenticate, and took up about 30% of the YouTube screen. He built Outlier Tracker in about 30 minutes with the video stats, tracking, and channel analysis he wanted.
• Excel clone (10:11). The browser-based version, which he says works perfectly and has the functionality a user would want.

Where It Falls Short

• The common objection (10:28). Building is not enough; you must maintain the software, keep it secure, and keep it up to date, especially given more hacks and the risk that vibe-coded software has holes.
• His response (11:01). He is not sure that is true. He thinks AI writes better software over time and that better models mean more hardened code. He describes a cost graph in which running an agent continuously to update the software and look for security issues is cheaper than buying from someone else.
• Open-source maintainers (11:25). Alternatively, you can download open-source software like ArtCraft, which other developers maintain by continuously spending tokens on it as token costs fall toward zero.

The Outlook

• A software renaissance (12:02). He expects every individual to build the software they need exactly, driven partly by love of the craft of building.
• A customizable operating system (12:13). He mentions a new operating system he names as Omarki (spelling as transcribed), created by the creator of Ruby on Rails, with AI built in so users can customize the OS continuously. He thinks there may be room for a new operating system.
• Call to action (12:42). Try open source, build a small app for yourself, and see how far you get. He also names DeepSeek v4.1 Flash (transcribed as "DeepSeq") as one of the best open-source models right now and points to a separate video.

Key Takeaways

• Berman's three conditions are model quality, low token cost, and agent autonomy.
• His central comparison is the ongoing cost of tokens to build and maintain against the price of a subscription or license.
• His evidence is anecdotal: his own clones and projects, plus ArtCraft. He does not present independent benchmarks.
• The maintenance and security objection is the main counterargument, and his answer is a prediction rather than a measured result.

Discussion Questions

• Which software you pay for or tolerate would you most want to replace with a custom build, and what features would it need?
• Berman says the cost of maintenance is lower than buying. What would you need to see to trust that for software that handles sensitive data?
• If software becomes de facto open source, who is responsible for security and long-term maintenance?
• How does a single person's tailored tool differ from software a whole organization has to support?