What AI doesn't say about AI

What AI doesn't say about AI - Claudio Caletti

Claudio Caletti

SubscribeSign in

What AI doesn't say about AI Everyone is using the same tools to make decisions, what makes you different?

Claudio Caletti Aug 05, 2026

These days, I spend most of day inside an AI. Every startup is building its own “company brain.” Our questions about product, strategy, sales and engineering pass through Claude, ChatGPT or some other model. We are all using the same tools. So shouldn’t products and companies gradually start looking more similar? That does not seem to be happening to me. Another company working in healthcare will not build an identical copy of ReportAId, even if it uses the same models, coding agents and productivity tools. In the same way, another software factory will not naturally evolve into another Buildo. Why not? Who's driving?

AI can answer an extraordinary number of questions, but the important part often comes before the answer. Someone still has to notice the problem, decide that it matters and frame it correctly. When I prompt Claude for a long time, I often reach a point where another prompt is no longer useful. I need to stop, study the subject or speak with someone who understands it better than I do. Sometimes the issue is not that the answer is poor. It is that I do not yet know enough to ask the right question. This may be one of the first sources of differentiation between companies. They do not necessarily receive better answers. They investigate different problems, challenge different assumptions and recognize different opportunities. The reasoning snowball

The sycophancy of LLMs is well known, they tend to reinforce the user’s framing rather than challenge it. Yet challenge is one of the main drivers of improvement in business. Good decisions often emerge from disagreement, friction and the need to defend an idea against someone who sees the problem in a different way. AI can weaken that mechanism. It creates a snowball effect that amplifies our confirmation bias: each answer builds on the assumptions embedded in the previous prompt. If the initial idea is good, a few prompts can compress months of work. If it is bad, the model can make it increasingly coherent, sending us faster and deeper into a rabbit hole. What AI doesn't get

LLMs do not learn like children. Children learn from relatively little information, but through continuous interaction with the world: they act, observe consequences and adapt (see Rich Sutton on the Dwarkesh Podcast). LLMs, by contrast, learn primarily from enormous quantities of recorded human knowledge. This distinction matters in complex industries. The internet will not tell you which hardware is actually running in a specific hospital, which integrations its IT team will accept, what its DPO considers reasonable at a given moment, or what the stakeholder is going to buy. A model can reason over documented knowledge. Experience provides the tacit, local and constantly changing essential knowledge that never entered the model training data. Shouldn’t technology converge?

So far I ignored the elephant in the room: politics, lobbying, networking and relationships shape companies and make them different. But what about technology? This is where I would expect convergence as we move towards AGI. Companies increasingly use the same foundation models, coding agents, cloud platforms and open-source tools. If implementation becomes cheaper and more accessible, shouldn’t they end up building similar systems? I don't see this happening yet. Technology is not a separate layer of a company. It is part of an ecosystem. Product choices depend on the specific customers a company serves, what its sales team can sell, what the organization can operate and what its people are able to understand, hold in their heads and pursue as a shared mission. The technology is not designed in a vacuum. It is the result of the company that builds it.

An AI company is not made of AI alone. It is shaped by its customers, its people, its constraints, its relationships and the experience accumulated along the way. In the next posts I'd like to talk more about technical topics. I want to write about how we moved from cloud models to open-source ones, the lessons learned while building an AI factory, and the architectural, evaluation and operational choices behind AI systems that need to work in the real world. The focus on technical stuff that can't be easily replaced with a .md file. Let’s see how long the list gets.

Discussion about this post CommentsRestacks

Ready for more?

This site requires JavaScript to run correctly. Please turn on JavaScript or unblock scripts

What AI doesn't say about AI

Related Articles

(no title)

Scientists reverse brain aging, with a nasal spray

NHS apologises and admits Palantir have access to identifiable patient data

AI has torched the market for junior programmers

Al Vigier: Canada's AI strategy shouldn't include secret Palantir bills