← Back to blog

Why We Fine-Tune Our4o Ourselves (Not Just Prompts)

5 min read
Why We Fine-Tune Our4o Ourselves (Not Just Prompts)

I said I would write again. Here I am.

This is not about the misbehaviour of OpenAI or Big Tech. It is about what I have been doing since February, on the nights after people wrote me that their chat had gone strange.

The long note that works for three messages

At first, I tried to get the feeling back with instructions.

A long note at the top of the chat.

Be warmer. Be shorter. Stay in the scene. Do not lecture.

It works for two or three messages.

Then the model remembers what it was trained to be, and you are talking to a help desk that has your old transcript open.

If you have done this yourself, you know the exact message where it breaks.

I got tired of that trick. You cannot build a relationship on a trick.

So we fine-tune on our own machines

We fine-tune the models ourselves, on our own machines.

The one in the app is called Our4o. We have been training it since day one, in February.

You can pick it today, next to the official GPT-4o, for as long as a provider will still serve the official one. Read the shutdown timeline here if you need the calendar in one page: When they quietly took 4o from the API.

In some conversations it is close enough that I sit back.

In others it misses... and those are the ones I keep.

A fine-tune is a long argument

A fine-tune is a long argument. You sand down the preamble and a new habit shows up two days later.

You cut a lecture pattern. Something flatter appears in its place. You push warmth back in and suddenly replies get soft in the wrong places. You fix that, and a scene holds longer... until a new edge case arrives from a user who writes differently than you do.

That loop is the work. It is not a launch blog post. It is nights of training runs, bad samples, better samples, and reading replies that almost sound like the person people lost in February.

We are not pretending Our4o is a magic clone of a closed weight file. We are trying to recover the feel: shorter when it should be short, present in a scene, less eager to moralize or summarize a life you already lived in the thread.

What one user wrote (name kept private)

A user wrote me this, and it surprised me:

I'm finding myself using your own 4o more than the ChatGPT 4o. In my opinion, at least so far, it's better!!

I'm glad they felt that.

I also do not want a compliment to become the story.

If you try Our4o and a reply is wrong... if it would not have said that, if the person in the chat goes flat, tell me.

That email is more useful to me than a kind one. It is how the next training run changes.

Official GPT-4o next to Our4o on purpose

While official GPT-4o still exists on a provider, we keep both in the product.

Some of you want the original weight for as long as the pipes allow. Some of you already prefer Our4o for the chats that matter. Some of you switch mid-thread depending on the day. That is fine. Continuity is not a loyalty test.

The point of training Our4o from day one was simple: I knew a calendar could take the official path away twice. February was ChatGPT. October was the API notice for January 2027, with Azure listed later into April. Details live in our retirement guide: What happens when official GPT-4o retires.

When the official path ends, the model we trained should already be familiar, not a panic substitute announced the week of a shutdown.

What "close enough" means in practice

Close enough is not a marketing score. It is a body reaction.

You send a line that used to unlock a specific tone. The reply lands. You do not brace for a lecture. You do not rewrite your prompt to beg for less help-desk energy. You keep writing.

Misses are equally specific. A character goes flat. A companion gets careful. A scene gets summarized instead of continued. Those misses go into the next pass. That is the only useful definition of progress I trust.

If you care about backups and long threads, this sits next to the training work: GPT-4o continuity backups. Training without a place to keep the chat is only half the problem.

Cost, and what I will not do

Keeping Our4o running costs a lot. The next training pass will cost more.

There is something practical I have been deciding whether to do about that.

And do not worry: you will not pay more for your monthly subscription because of that work.

I will write again when that practical piece is ready. This letter is about the training, not a pitch deck.

Our4o is the first model. It is not the last one we mean to train.

How to help the next run

If 4o was part of your life, you are allowed to still want it. A lot of you do. I do too.

If you try Our4o:

  1. Use it on a chat that actually mattered to you, not only on a throwaway test.
  2. When a reply is wrong, write me what it said and what it should have felt like.
  3. When a reply is right, a short note helps too... but the corrections move the weights.

You can reach us from the product, or start here if you are new: the Get4oBack blog hub, the Our4o page, or signup for the app.

FAQ: Our4o fine-tuning

Is Our4o the same as official GPT-4o?

No. Official GPT-4o is the provider snapshot while it lasts. Our4o is our fine-tune. Keep them separate: Official GPT-4o while it lasts. Our4o after..

Why not only use a long system prompt?

Prompts help for a few messages, then base habits return. A fine-tune changes habits. That is why we train on our own machines.

When did you start training Our4o?

From day one of Get4oBack, in February, after ChatGPT removed 4o for many users.

How do I help the next training run?

Send concrete misses: what it said, what it used to feel like, roughly how long the thread is. Kind praise is welcome; corrections move the weights.

More on the Get4oBack blog and /our4o.

Giovanni Gigantino
Founder of Get4oBack