Software development philosophies · An argument from my own pages · My opinion
What gets cheap, and what doesn’t
A post by Alex Lieberman shared Claude’s answer to which businesses get stronger as raw intelligence explodes. I point out where my own Lab pages disagree with it, and argue that when execution is cheap the margin moves to guaranteeing the work and to service contracts for continuous updates and security. This is my opinion.
The claim
My opinion
This page started from a post by Alex Lieberman and the answer he got from Claude, and I want to point out where my own pages disagree with it. The answer is about which businesses get stronger as raw intelligence explodes. My claim is narrower and about economics: when execution gets cheap, the margin moves to what stays expensive, and I think that is guaranteeing the work, and service contracts for continuous updates and security.
I’ve put this into practice. KFTCDA Elite AI Enablement now carries cover after the assistants go live. The response times and the improvement schedule are agreed in each engagement. Its page puts the first part this way:
“Cover for what I set up: if something I built stops working, putting it right is on me.”
This site is built with Claude’s help, so I’m disclosing that I’m quoting a model I also use. There are no figures on this page that put a number on any of this. I have none.
What the source is
A model’s answer, in a screenshot
The source is a post on X by Alex Lieberman on 2026-11-02 [Alex Lieberman (post)]. He wrote that he asked Claude what types of businesses outside the labs will get stronger as raw intelligence explodes. He called the response spot on, and the post carries a screenshot of the answer [Alex Lieberman (post)]. So the substance is Claude’s answer, which I read as a transcription of that screenshot.
That makes it a model’s opinion. It cites no sources. I can’t confirm that a screenshot is unedited, and I say nothing here about Lieberman or his businesses beyond what the post shows.
The answer lists nine kinds of business whose margin goes up. Owners of physical bottlenecks such as power, grid connections and chip supply. Liability-bearers and licensed entities, anyone who signs and can be sued. Owners of proprietary feedback loops, meaning outcome data that is not on the internet. Distribution and trust. Owners of the customer in fragmented physical services, where the software vendor selling to them does not keep the margin. Industries held back by experiments and not by ideas. Verification, security and provenance. Deployment and integration, with a caveat. Human-premium goods [Alex Lieberman (post)].
Where I agree
Two points in the answer match pages I already wrote. The first: when generation is free, “checking becomes the job” [Alex Lieberman (post)]. The second: someone who signs and can be sued keeps the value, because someone still has to be accountable for the answer.
“nothing ships through my loop without a receipt for this exact content”
“code for anything a script can decide, a person for anything that cannot be undone, and a model for neither”
“A person says yes, then it merges to production.”
I disagree: generation is not free
Where I disagree
The answer treats generation as free. On my agent-graph page it has a cost I can point to. It also has a budget, because a run that can’t stop is a run that can spend without limit.
“Tokens rise, because each worker loads its own context.”
“if the same gate has failed three times for the same reason, stop and report what was tried”
“They rarely cost the same, so the two cost sliders matter more than the threshold”
I disagree: more output does not by itself mean more demand for checking
Where I disagree
The answer says every tenfold rise in output creates demand for verification, and gives no source for it. My platform-team page collects evidence that more output doesn’t clearly mean better results. The first figure there is from a vendor that sells engineering analytics.
“AI does not clearly make developers faster.”
“Faros, which sells engineering analytics, found 98% more merged pull requests but 91% longer review times and 9% more bugs per developer.”
I read the Faros figures as a vendor’s own data, which is how that page labels them. I read the METR result on the same page as the weak evidence that page calls it. They show a cost of checking more output. They don’t show that anyone pays for it.
I disagree: software vendors do not all lose
Where I disagree
The answer says the operator who owns the customer keeps the margin, and the software vendor selling to them doesn’t. My Software Checkup rules say it differently. Some software stays worth paying for.
“Some software is worth paying a specialist for.”
“Anything on the won't-rebuild list, or a tool you use most of, or need for its certifications or ecosystem.”
“Build when your process has settled.”
I disagree: a model’s answer is not a finding
Where I disagree
The post calls the answer spot on. I wouldn’t take it as a finding. I don’t know that it has a source, and there’s the reason I give about models checking their own work:
“a model asked whether its own work is finished tends to say yes”
Where the margin goes
My opinion
Here is my own claim. If building and running the first version gets cheap, the cost doesn’t vanish. It moves to what stays expensive. I think that’s guaranteeing the work, with a contracted SLA, and service contracts for continuous updates and security, and that these become the profit margins.
The reason is that a guarantee is a promise someone has to keep when things change. A model can produce the answer, but someone still has to be accountable. That’s the answer’s own point about liability.
“A setup that was right last year can be wrong this year.”
“Patching, multi-factor sign-in and backups that have been tested remain the floor, and I would not skip any of them”
“I think there is no real way to be secure in the age of AI other than constant monitoring, threat analysis and agent swarms.”
“is a claim you can check: this named check passed, on exactly this content”
A guarantee rests on a record like that: a named check, passed on this content. Without the receipt the promise can’t be kept or tested. So I think receipts are what a guarantee is made of.
Where I could be wrong
The evidence against
No figure on this page prices a guarantee, a service contract or an update. The claim could be wrong in either direction. Cheap execution could bring the margin down everywhere, or the margin could move somewhere I haven’t listed.
“This is my opinion, and I have not run an experiment to prove it.”
Related on this site
- Designing an agent graphBudgets, gates and what a run costs.
- Traces, checkpoints and receiptsThe record a guarantee would rest on.
- Security in the age of AISecurity as ongoing work.
- The next platform teamThe evidence on AI and developer speed, labelled by who produced it.
- Who to hire nowThe engineer worth having on a team has changed because of LLMs, agent graphs and second brains: my points from other Lab pages, gathered and linked.
- Attention is the bottleneckThe scarce thing in a team that builds with agents is a person’s attention: a vendor’s design for an always-on agent, taken apart with my own Lab pages.
Sources
- Alex Lieberman (post) · Alex Lieberman (@businessbarista), X post of 2026-11-02 asking Claude which businesses outside the labs get stronger as raw intelligence explodes, with a screenshot of Claude’s answer; the answer is a model’s opinion with no sources, and the screenshot cannot be confirmed unedited (read 2026-11-02)
Sources as read on 2026-11-02.