AI labs watch what you build on their models, then ship the product themselves. Here is how to avoid being the next casualty.
Posted
2 days ago
Duration
Format
Talking Head
sarcastic
Views
121.9K
1.6K likes
57 · 43
Big Idea
The argument in one line.
AI labs see what people build and ask on their platforms, so they ship those products themselves, and the safe moves are going deeper, going niche, guarding your SOPs and running open models for sensitive work.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You are building a SaaS or app on top of Claude or OpenAI models and worry the lab will ship your feature as a one-click plugin.
You run a company with proprietary SOPs, prompts and customer data and are deciding what is safe to paste into a cloud model.
You are an indie builder looking for AI product niches that big labs will not bother to enter in the next few years.
SKIP IF…
You want a technical walkthrough of installing and hosting open-weights models; this stays at strategy level.
You want evidence-backed analysis of lab data practices; the claims here are opinion and inference, not documented findings.
TL;DR
The full version, fast.
Anthropic and OpenAI keep launching vertical products (finance, legal, security, design, sales) that knock public software stocks down on release day. The argument is that labs learn what to build from how people use their models, even when training is opted out, and they already own distribution. The advice is to avoid thin wrapper apps, build in narrow, physical or regulated niches, keep valuable SOPs out of cloud prompts, and use a harness that lets you swap between open-weights models for sensitive work and Claude for everything else.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Opting out of training does not stop a lab from learning how you use its product, only from training its models on your chats.
Labs already have hundreds of millions of users, so a new product ships as a one-click plugin with no marketing cost.
Anthropic's terms forbid using Claude to build a competing product, while its IPO filing says Anthropic may compete with software companies.
If you feed an SOP into a cloud model once, it sits in a transcript the provider could use later to rebuild that process.
Physical-world work, regulated work and relationship-heavy businesses like vendors and sourcing are the hardest for a lab to disrupt.
AT&T reportedly sends about 25% of its AI usage to open models and wants 70% to 80%.
Running your own compute for open models lets you batch requests, which can cut the cost of intelligent models at scale.
A product that only visualizes slides, or anything similarly shallow, has a lifespan measured in months once a lab notices the category.
Opportunity windows from new model drops are more frequent and bigger than before, but they close much faster.
Keep open-weights models for sensitive work and hot-swap to Claude or ChatGPT only for less sensitive tasks.
Takeaway
Build where AI labs will not follow
WHAT TO LEARN
Labs learn from usage and own distribution, so your defense is depth, niche, guarded data, and the ability to swap models.
Labs launch vertical products right after seeing demand, so a thin wrapper that just reformats model output is likely to be replaced within months.
Opting out of training still leaves providers able to learn what people use their products for, so assume usage patterns are visible to them.
Prefer narrow niches, physical-world work, regulated work and relationship businesses, where accountability and sourcing make a one-click plugin impractical.
Treat your SOPs, prompts and customer data as your moat, and avoid pasting the most valuable ones into cloud chats where transcripts persist.
Use open-weights models on your own compute for sensitive work and switch to Claude or ChatGPT through a model-swapping harness for lower-stakes tasks.
Fast, short-lived opportunities tied to specific model or feature drops are the ones big labs ignore, and they favor builders who move quickly.
Glossary
Terms worth knowing.
Wrapper product
An app that mostly adds an interface or prompt on top of someone else's AI model, with little unique data or depth of its own.
Open-weights model
A model whose weights you can download and run on your own servers, so no prompts or data are sent to a third-party API.
Knowledge alpha
The speaker's term for a business's accumulated edge: SOPs, prompts, customer data and edge-case knowledge that make it unique.
SOP
Standard operating procedure: the documented steps a business follows to do a task.
Harness
A tool that wraps AI models in a working environment and lets you switch which model it uses, for example with a slash model command.
Economic Index
A periodic Anthropic report that analyzes over a million business API transcripts to see which tasks AI can automate or outsource.
The Script
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
17px
Claude is not only eating your math homework, folks, it is literally eating everything. This started in 2025, when they launched Claude for Financial Services. Then they launched Claude for Excel and Claude for Life Sciences.
In 2026, they launched Claude Cowork's Legal Plugin Suite. You know what happened? Right after, Thomson Reuters fell 16%.
Immediately after this, they launched Claude Code Security. A few days later, CrowdStrike, big company, down 11%. They've done the same thing with design.
I don't know if you guys remember Claude Design. Figma fell about 7 % the day that that was released. And the funniest thing about all this to me is that in the Anthropic filing for their IPO, they said, and I quote, Anthropic may from time to time...
compete with enterprise software and application companies. That sounds like code for we're coming for your ass. Claude for small business.
Claude science. They just did Claude for financial advisors. Claude sheets.
Claude whiteboards. Claude data. Claude meetings.
Claude for manufacturing. Where's Claude for doing my dishes? How about the Claude for cleaning my toilet?
You know, I say this sort of tongue in cheek, but ladies and gentlemen, rest assured, they're going to get there soon too. And it's obviously not just Anthropic. It's obviously not just Claude.
Large language model companies and labs have been doing this thing for the better part of the last couple of years. This month, OpenAI also launched ChatGPT for sales. If you look on their little plugins library, it sits right next to Apollo, Clay, Gong.
These are pre -existing SaaS companies with a fair amount of usership. You know, every day I'm scrolling through Twitter and I'm seeing somebody wake up and say, Claude just killed my startup. OpenAI just destroyed my product.
You know, I think a lot of the time these were probably just wrappers to begin with. But dude, if you look deep enough, isn't everything kind of a wrapper? If you've ever wondered what's really happening here, it's that, you know, these AI companies, they make models.
And other companies build products on top of those models. Because other companies have built products on top of the models, the AI companies get to learn what sorts of things their users are asking. And then they soak up all the IP.
And miraculously, the AI company typically starts to sell those products itself after a given amount of time. On most plans, free or paid, your conversations, like your private one -on -one chats with Claude, are actually training their models unless you turn that off. For example, their consumer terms.
Legitimately say, it may use your chats to develop other products and services, including training our models unless you opt out. And all model companies, including OpenAI, Google, XSI, and so on and so forth, all have the exact same settings. However, what people don't understand is that is just about training the models.
Because even if you opt out, the companies are still going to learn how you use their products. And ultimately, if you're working for another company that they may want to get into the industry of, you know, ultimately speaking, what your SOPs and best practices are. The only difference is they're not actually training the models.
Once every once in a while, Anthropic publishes this report called the Economic Index. Now, in this process, Claude will read over 1 million business API transcripts and then sort them by task and then do a bunch of analysis on them. The analysis is specifically to determine what tasks...
Claude and AI are actually capable of doing, basically outsourcing, offshoring, or automating entirely. So what does Claude and Anthropic more generally do with all of that analysis? Aside from, of course, publish it in their economic report.
I think I have an idea. And a really funny thing here is that Anthropic's own terms actually say you may not use Claude to build a competing product, but its filing says that Anthropic may compete with software companies. So the rule is sort of one -sided.
And it makes sense that they do this, to be clear. I mean, if I were a big daddy LLM company, I'd probably do the same thing. By reading all your transcripts, the AI companies can legitimately see demand, both for the number of times their connectors or plugins are used, but also more broadly, just what people are asking, what people want.
So that is the sure as shit clearest thing or signal that any company, let alone a large language model company, can capitalize on to make a lot of money. But at the same time, these guys also have distribution. They're used by hundreds of millions, if not billions of people now.
They already have all the users. They don't actually have to do any additional marketing. They also have the advantage of being able to use AI to create products really quickly to then serve to the users that are on their AI platform.
And then they get to pitch it not as, oh, here's a new startup we're starting. Literally, oh, here's just a one -click plugin that might be able to help you with that just a little. You know, I'm not trying to say this is good or bad, but I am definitely trying to make sure you guys know what's going on.
Virtually all of the sophisticated intellectual property that you or your company, dare I say, is currently feeding Claude and other models will soon be used against you. I mean, it is just sort of the economically optimal thing to do, right? If these large language model companies have all the data and the means to quickly replace you.
But even more broadly, if you're just using these products in your day -to -day work, eventually they're probably going to want to replace you, right? And this is why I think a lot of companies are now starting to run open models. In case you guys didn't know, an open await model are models that you can download and then run right on your own servers.
So you're hosting it entirely yourself, meaning you're not actually sending like any API requests or data over to cloud. And the reason why that is important to companies is because of knowledge alpha. Knowledge Alpha is basically like your SOPs, your prompts, your customer data, information about your edge cases, basically the stuff that your business has dutifully collected, presumably over the course of the last few years, that make it unique and powerful.
So AT &T, for instance, something like 25 % of all their AI usage currently right now goes to open models. Their stated goals are to get that to 70 % to 80%. I think they, as a big infrastructural company, one that probably deals with a lot of software, are particularly worried about enterprise data leak.
Because as we know, enterprise data is a gold mine. And I quote, and the tools are just a way to mine the gold. They've also done lots of surveys on this.
At least 40 % of companies surveyed said that they want open weights models in production. 60 % of them have already moved some work off of current systems onto open weights models specifically to avoid this. Now, you know, there are a lot of reasons why companies are doing this.
A big one is cost. In addition to paying like the reaper's toll of all of your freaking data being leaked. Obviously, it costs a lot of money to use really intelligent models at scale.
And when you buy your own compute and stuff like that and actually have it run internally inside of an enterprise, you get to take advantage of this batching effect and basically save a fair amount. But more broadly, yeah, this is this is going on right now. And there's a reason why so many people are reaching out to me and other people in the industry basically saying, yo, dude, do you know anything about setting up these?
custom open weights models, I don't think that we're comfortable sharing all this info with Anthropic, Big Daddy, OpenAI, etc. So what do you actually do with this information? Well, the first thing is just don't build a freaking wrapper product.
I mean, goddamn, are those wrapper products screwed? This comes for everybody. If you're going to build a product, it has to be a lot deeper.
You know, they just shipped personal finance, decks, you know, meeting notes, sales prep, presumably like another dozen other products that I'm not even keeping up to date with. But I mean, if you're making an app that just like visualizes slides, you're probably going to have to go a little bit deeper than that if you don't want to get replaced literally the next couple of months.
The second thing is to build products and services just where these companies don't really think it's financially profitable to go, at least not for the next few years. I mean, Anthropic says they're going to compete in select categories. What that means is they're going to compete in places where there is easy pickings and money to be made.
Maybe eventually, over the course of the next decade, they'll get everything. But at least in the short term, you know, pick a really narrow niche. I'd also recommend physical work.
Something that intersects with the physical world is ideal. How about regulated work, you know, work where somebody basically needs to be held accountable, businesses that deal with relationships, vendors, sourcing, these sorts of things aren't very disruptible. Another thing is to protect your SOPs.
So if you guys are going to use, you know, cloud models or open AI models in order to do work, it's interesting, I would probably not actually feed in my SOPs, despite how compelling and simple it may be you know it's sort of like the devil whispering on your shoulders just paste that thing into the cloud prompt everything will be okay but the reality is i mean even if you know you feed the sop once it'll be stored in a transcript that they could use at any point in time to rebuild that process so if your sop is like super crazy dynamite if that is really the moat of your business be careful with that and then finally use something that allows you to hot swap models like i would keep open weights models for all sensitive work Assuming you are in a big enterprise with those proprietary SOPs I talked about.
And then just hot swap it to Claude in and out when you're working on something that maybe is a little bit less sensitive. This is really annoying to do, to be clear. But there are a lot of harnesses out there that now offer this sort of thing as a default option.
And you can very quickly and easily just switch between what model you want to use with these open harnesses, which make it super easy to just be like, okay, slash model, let's change this. Okay, feed that in. When you're done, maybe you can open a new chat transfer.
or something like that, and then just type slash model and go back to like a cloud or chat GPT. The way that I see it, these AI market opportunities that, you know, are currently being filled by these LLM companies, you know, they're one of many. With AI being introduced to our economy.
the frequency of these opportunities is going way up. And so there are these vacuums that need to be filled that are constantly created by these, you know, 10 ,000 model and product drops a day. Also, the magnitude of like total value you can exploit from one of these opportunities is also through the roof.
But the total amount of time that those opportunities are available, like in the case of a wrapper product, has also shrunk a lot. And so, you know, AI companies are going to take the big obvious categories because it just makes sense. But those like niche fast opportunities, those ones that appear for like a month and then disappear, that are typically tied to specific features or products dropped, those ones are really ripe for the picking.
And I think anybody with like a growth hacker mindset can make, you know, tens of not hundreds of millions of dollars over the course of the next year or two, if they take advantage of that. So if you guys want help to find niches like that, just check out Maker School. That's more or less what we do.
Otherwise, please subscribe and I'll catch all y 'all in the next vid.
The Hook
The bait, then the rug-pull.
Anthropic launches a vertical product and a public software company drops double digits the same day. The speaker says that is not a coincidence, it is the business model.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
A behavioral-neuroscience grad turned AI-agency founder breaks down ten ways to use AI as a learning tool, and the one way people default to that quietly works against them.
A same-day walkthrough of Claude Fable 5.1's benchmark chart, why the real story is cost-per-task rather than raw score, and what the new safeguard numbers mean for how often the model refuses benign questions.
Nick Saraev built Clairvo — an AI power dialer at $1M ARR — using Claude Code, and this is the exact playbook: idea mining, the build loop, pricing strategy, and four moats that survive AI commoditization.
A ten-minute rapid-fire tour of thirteen real things people are doing with Claude, Grok bots and Google AI Mode right now, from haggling with Comcast to coaching your tennis swing.