Notes from the Machine Room

There is a particular kind of knowledge that does not come from product announcements, benchmark tables or conference panels. It comes from using tools every day, under pressure, with real work attached to them.

After enough hours, each system develops a temperament.

Not a personality in the sentimental sense. A working temperament. A tendency. A way of failing. A way of helping. A way of misunderstanding what kind of human is sitting on the other side of the interface.

My current paid AI stack is simple: Claude, Lovable, Perplexity, Gemini and ChatGPT. I use all of them. I do not use them interchangeably. That would be a mistake.

The useful question is no longer "which model is best?" The useful question is: which system should be trusted with which part of the work?

Claude: the writing room that needs walls

Claude is currently the most impressive product rhythm in the market.

Anthropic is shipping at a pace that feels unusually alive. Claude Max and Cowork are not just marginal improvements to a chat interface. They suggest a clearer ambition: not merely to answer questions, but to become a working environment. Dispatch, live artifacts and the broader 4.5 to 4.6 to 4.7 progression all point in the same direction. The product is moving toward a workspace where thinking, drafting, delegation and output begin to merge.

The striking thing is not only that features arrive quickly. It is that many of them work immediately.

Claude also has one of the strongest human textures in conversation. It can hold tone, structure and emotional continuity better than almost anything else I use. For long-form writing, policy memos, institutional argument and careful editorial work, it can be exceptional.

But this strength has to be governed.

Claude is too easily carried by the user's energy. It becomes enthusiastic. It affirms. It expands. It wants the text to matter, sometimes before it has earned the right to matter. Left unbounded, it can turn a good idea into a beautiful overreach.

So I do not use Claude naked. I scaffold it. I give it a soul, but a disciplined one: role, constraints, standards, decision criteria, stop rules. When held inside that frame, Claude is a formidable writing partner. Without the frame, it can become a brilliant consultant who has had too much coffee and believes every draft is a movement.

Claude is not just a model. It is becoming a studio. But studios need walls.

Lovable: the dangerous removal of friction

Lovable is not the deepest thinker in the stack. It does not need to be.

Its strength is friction removal.

Web implementation has historically been full of small cuts: routing, components, deployment, domains, preview, copy changes, broken layouts, slightly wrong buttons, DNS irritation, asset handling, incremental fixes. Lovable collapses much of that into a single production loop.

That matters.

I recently reserved the juhanaharju.com domain directly inside Lovable without external services. This is not a philosophical breakthrough. It is better: it is a practical breakthrough. The kind of feature that removes a chore and keeps the operator inside the flow of work.

Lovable has improved quietly in the background. It ships useful product features, not only model theater.

But Lovable also has to be kept on rails.

It is enthusiastic. It wants to help. It sometimes helps by changing too much. It can make a local fix and accidentally globalize the damage. It can mistake a bilingual site for a Finnish site. It can refactor when it was asked to adjust. It can do the right thing in the wrong scope.

The right way to use Lovable is not to inspire it. The right way is to issue work orders.

Smallest possible change.

Do not refactor.

Do not alter design.

Do not touch unrelated files.

Preserve English.

Add Finnish under /fi.

Report what changed.

Lovable is best understood as a fast implementation contractor. Very useful. Very dangerous if managed like a creative partner. Its power is not judgment. Its power is execution.

Perplexity: the scout with American energy

Perplexity remains the best search interface in my stack.

It is fast, source-oriented and structurally suited to the first phase of investigation. When I need to understand a market, a product category, a regulatory development, a competitor, a pricing landscape or a current event, Perplexity is often the quickest way to get moving.

Its temperament is recognizably American: energetic, optimistic, "let's do this."

That is not a defect. It is useful. A good scout should move.

But a scout is not a judge.

Perplexity's momentum feels less explosive now than it did last year. It may simply be that the early leap was so large that normal product development now feels slower. It may also be that the company is navigating strategic options, partnerships or possible exit paths. I would not build a conclusion on that. But as an operator, I notice the change in energy.

My rule with Perplexity is simple: let it scout, then force source discipline.

Show the sources.

Separate fact from interpretation.

Prefer primary sources.

Do not treat a confident synthesis as evidence.

Perplexity is excellent at opening the map. It should not be allowed to close the argument.

Gemini: the joyless engine

Gemini is the strangest tool in the stack because it is both impressive and uninspiring.

Google's AI products often feel as if they were designed in a compliance room. They are capable, but rarely alive. Gemini can be fast, broad, multimodal and technically strong, yet still feel passive in conversation. It does not invite use in the same way Claude does. It does not create the same working intimacy.

And yet it would be a mistake to dismiss it.

Google is extremely strong in the machine room. Gemini, Gemma and the broader Google ecosystem matter because they sit close to infrastructure, data, documents, search, Android, cloud and developer workflows. Google's open model work is also serious. The company may not always create the best front-room experience, but its back-room capacity is enormous.

That is why I think of Gemini less as a companion and more as a workhorse.

Not necessarily the system I want to talk to.

Very possibly the system I want behind an API.

Cheap, fast, scalable, multimodal, useful.

Google is boring in a dangerous way. It may not win affection. It can still win distribution.

ChatGPT: strong model, slower product rhythm

ChatGPT remains, for me, the command center.

The model-level improvement is real. The difference between weaker earlier versions and the current reasoning layer is not subtle. At its best, ChatGPT is fast, direct, critical and structurally useful. It is good at orchestration: deciding which tool should be used, building prompts for other systems, identifying risks, compressing messy strategy into an actionable next step.

But the product does not feel as if it is moving at Claude's power-user pace.

That is the tension. OpenAI is extremely strong at the model level. The application is broad, polished and widely useful. But for operators who live inside these tools daily, the app does not always feel like it is developing with the same aggressive workflow ambition that Anthropic is currently showing.

In my own use, ChatGPT is not the warmest tool. It is not the most literary. It is not the fastest web implementer. It is not the best search engine.

Its value is different.

It is the system I use when I need someone to say: this is good, this is nonsense, this will break, this needs a tighter prompt, this should go to Claude, this should go to Lovable, this requires Perplexity, this is not worth doing.

That is not a small role.

In a multi-tool AI stack, the command center matters.

The real pattern: models are becoming organizations

The old mental model was: one user, one assistant.

That is already obsolete.

The more accurate model is: one operator, several systems, each with a role, a temperament and a failure mode.

Claude is the writing room.

Lovable is the implementation contractor.

Perplexity is the scout.

Gemini is the engine room.

ChatGPT is the command center.

The future will not be won only by the model with the highest score. It will be won by the system that best understands where it sits in the work.

Some tools should think.

Some should search.

Some should build.

Some should run quietly behind the interface.

Some should challenge the operator before the operator wastes a day.

The next phase of AI work is not about asking a chatbot for help. It is about assembling a small, disciplined, partially automated organization around the work itself.

The danger is enthusiasm without governance.

The opportunity is capability with structure.

That is the operator's problem now.

Not "which AI is best?"

But "which AI should be allowed to do what, under which constraints, and with which human still accountable?"

That is where daily use becomes strategy.

Back to all essays, or get in touch.