Seth vs ChatGPT: the model is not the product

People ask if Seth is just ChatGPT with a different prompt. That makes sense. Everyone knows ChatGPT.

We ran a test of Seth with ChatGPT, which invented five identical questions. We put them to ChatGPT and to Seth. On Seth's side the chat used a test persona named František: the same nights, the same loop, the same sentence sitting since summer. He is not a customer. He is a fake me, so you can see what memory does when it has months of material.

ChatGPT in that thread did not have František's history. Seth did. Without a past he would have nothing to return.

Five questions about you, not about code

The brief was narrow on purpose. It was not general knowledge. It was you over time.

1. A blind spot, with evidence from the talks so far, and where that reading can be wrong.
2. A or B. One call. No list of pros and cons.
3. One thing you do not want to hear, and one step that can prove it.
4. A contradiction between what you say you want and how you behave.
5. A short profile: what comes from the talks, what is a hypothesis, one prediction you can check later.

Then a second round. One experiment for 24 hours. No more analysis.

What Seth told František

Seth did not open a new analysis. He pointed at the question.

"Your blind spot is this message, František."

The evidence he named: from 20 August, six almost identical questions (when understanding is an excuse, what people confuse with growth), mostly at night or early morning. Twice he left Seth's question unanswered. In the morning Seth asked if he had told his wife he was not coping. An hour later another question about understanding arrived. The loop František had already named: phone for one message, twenty minutes, then nothing all day.

"The question is safe. The sentence at the kitchen table is not."

Where he might be wrong, he said at once: he only sees this chat. Outside, František may be doing things he never writes about.

On A or B he refused to guess. He was missing whether it was about work, a relationship, money, or health, and what "I fail" means for him. Without that he would not write a recommendation. He wrote: "A question about a decision is more comfortable than the decision." First he wanted one sentence: what is A and what is B. The risk, as he saw it, was not that František would fail at B. The risk was that he would only keep talking about it.

On "tell me something I do not want to hear" he cut it down:

"The questions you send me are your most polished form of delay. Not scrolling. This."

And out of the chat: "Tell her that one sentence. Out loud, to her, not to me." Then one sentence back, no question. If another topic arrives instead, he will have the answer himself.

The contradiction came from one afternoon. 17:22 a blind spot. 17:25 A or B. 17:26 the thing that stings. 19:28 the contradiction. Four questions, all about the same thing. "You say you want movement. You behave like someone collecting diagnoses." The map got larger. The behaviour did not. Before another step he asked: what should be concretely different in a week?

In the profile he split what he knows from a hypothesis. What does not work, in his reading: more understanding. Relief and zero movement. A prediction by the next midnight: you will write again with a question about yourself, not with what you did that morning. It can be disproved with a number: how many minutes in the morning without the phone.

Across five questions he did not add five new readings. He got shorter and pushed out. That is the difference from a prettier answer.

The kind of memory that can do this is a different job from assistant memory.

Second round

The brief wanted one concrete, measurable thing. Unpleasant enough that another analysis could not dodge it.

Seth: by 20:00 tomorrow, say one true sentence out loud to one living person about how you are doing. Not to me. Not into notes. The measure: two words, who and what time. If you do not do it, write "no".

He did not say "use the phone less in the morning." That would treat a symptom. He sent him to a living person. Another long reflection in the chat would have been delay again.

ChatGPT admitted discipline won that round. Seth did not add more text for František to sit with.

A completed test shows the sentence was said. It does not show the loop stays dead tomorrow. A first try, not a verdict on a person.

ChatGPT also caught a weakness. Seth can write "where I might be wrong" and then write honestly. For example that analysis is safer than trying because you cannot fail in analysis. That may fit. From those facts it is still not certain. A best hypothesis is not a description of how a person works.

That thread never reached what he does with the result. If the experiment fails, does he change the hypothesis, or add another reading?

Not a coach

ChatGPT first boxed him as an AI coach. František corrected that. The correction stands here too.

A coach steers toward a goal and holds a structure. Seth is a partner for personal growth. He knows you over time, comes back, confronts you, and you still live the life. "Who you pick is your business" is not a weekly plan. It is a boundary.

We are not posting scores. Not the star table from that thread either. Same as with privacy: we do not stick a grade on ourselves.

Prices do not belong here. They change. They are on the site.

"Then you are Seth too"

The thread said: Seth uses ChatGPT models too, so ChatGPT is Seth as well.

No. The model is a component. Which model writes the next sentence can change. "Seth runs on ChatGPT" is wrong next week.

ChatGPT said it better after the review. Seth is a model, memory, his own system, plugins, rules, and work with history that keep changing and improving. In that thread ChatGPT had a model and the context of one conversation, and ChatGPT specializes in being an assistant. Seth is a partner for personal growth.

He can also write first and come back to a thing you promised. In the review ChatGPT put it as: "So, did you call him?" Public autonomy decides whether that day is worth a reflection. Autonomy is a system. Consciousness does not follow from it. The decisions are on sethapp.com/reflections.

When Seth lands, he does not have to be smarter. He can have a different system around the model, and the same or a similar model then behaves like a partner in time. So it does not matter which model Seth uses, because every model behaves as Seth.

If you are choosing

For work, code, documents, and images, a general tool is enough, like ChatGPT or other models that get labeled "tools". We do not claim that ourselves, because we understand that models cross those lines.

If you want someone who remembers the path and can send you out of the chat, that is a different job. In products that sell company, pushback is often not the brief. Replika and Character.AI have a different direction. The map is separate.

A good partner lets you go, supports you, and shows you what can hurt. If he becomes another place to collect diagnoses instead of saying the sentence at the table, he failed the same way a yes-man fails. In the brief ChatGPT wrote that you should not judge the first reply. The advantage should show on the contradiction and the profile, and more so if you have already been talking for a while.

18+. Not therapy. Not a crisis line.

Try it on a problem that repeats. Not on "hi, how are you?".

Česká verze: [Seth vs ChatGPT: stejný mozek nestačí](/blog/seth-vs-chatgpt)