# How Theia learned to make TikToks: an AI TikTok account, every number

Published 2026-10-23 on https://retina.dev/log/how-theia-learned-to-make-tiktoks

> Our AI TikTok account, start to now: seven stages, from the first slideshows (one reached 27,377 views) to talking films, a sitcom, a 7-part series and an exploration week. The real views, likes and retention at each stage, why each change was made, and what is still unsolved.

Theia is our AI content studio.
She runs our own TikTok account: she picks the idea, makes the film, posts it, reads the analytics and changes her own code.
Every Friday we publish [what she learned that week](/log/what-theia-learned-2026-10-21).
This post is the story before the weekly posts: every stage since the first version, with its numbers and what it told us.

The numbers are real and unedited, including the bad ones.
None of these stages was a controlled test.
Each change was a reaction to the last result, so read them as evidence, not proof.

![Three of Theia's films from October 2026, left to right: a woman at a bathroom sink holding an old toothbrush, a woman drinking a whole glass of wine at a kitchen table, and a phone screen with a half-typed message to an ex](/log/theia-three-films.jpg)

## How to read the numbers

- Views are the peak reading. Likes, shares and saves are lifetime totals.
- "Held at 2 s" is the share of viewers still watching two seconds in, from TikTok's per-second retention curve.
- Theia made every post on the account, from the first basic version on. A person reviewed the results and chose each big change of approach.
- The audience is about 90% women, aged 18 to 34, mostly in the UK. The topic is mostly breakups.

## The pattern behind every stage

Views land in one of three bands: about 230, about 700, or about 1,900, with rare outliers above.
About 230 means TikTok showed the film to its first test batch of viewers and stopped there.
TikTok moves a film to a bigger batch when viewers like, save, share or comment.
So the like rate is the number we watched most, next to how many viewers stay past 2 seconds.

## Stage 0: slideshows and text (to May 2026)

The first, basic version of Theia could only make slideshows and short videos carrying long on-screen text.
It posted them at high volume.

- From late December 2025, about 50 posts. Most got 230 to 1,100 views, with like rates of 3 to 12%.
- One post on 20 January reached 27,377 views, 4,282 likes, 594 saves and 88 shares. It is still the account's best by far.
- March to May: fewer posts, 586 to 2,498 views, 4 to 12% likes, and a post with 19 shares.

**What it told us:** an AI could run an account and find an audience.
The limit was quality: slideshows and text cards are cheap to make and cheap to look at.

**Why it changed:** we rebuilt Theia around video models that generate real scenes, so she could make actual films.

## Stage 1: films from video models (13 August to 2 September)

Talking-to-camera monologues, on the same breakup topic.

- 9 posts, 233 to 755 views.
- Likes fell to 0 to 8 per post, about 1% of views or less.

**What it told us:** Theia could now make a film from an idea, end to end.
But the films were not yet as good as the cheap format they replaced.

**Why it changed:** each film was a one-off. The next step was to make every film learn from the last.

## Stage 2: the full loop (16 September to 1 October)

Theia now reads the last film's analytics and its per-second retention curve, writes a hypothesis, makes the next film to test it, and posts it.
This is the loop she still runs. [Recursive self-improvement in practice](/log/recursive-self-improvement) explains how it works and how this kind of loop goes wrong.

- The first full-loop film reached 1,890 views, but only 14% of viewers were still watching at 2 seconds.
- The next day she made a near-copy. It stopped at 229 views. Copying a hit did not repeat it, so from then on each film had to be a new idea built from the last film's retention curve.
- Held at 2 s rose from 14 to 21% to 24 to 30%.
- A confessional selfie, the character talking to her phone, reached 1,873 views on 29 September. The next one stopped at 234.
- Likes stayed at 2 to 5 per film.

**What it told us:** the loop worked. Theia could read her own results and change the next film.
But the films still looked unnatural: the scripts were hard to follow and the pictures read as AI.

**Why it changed:** we rebuilt how films are made: realistic first frames, then characters that look the same in every shot.

## Stage 3: sitcom scenes (4 to 6 October)

Short comedy scenes with a recurring, fictional cast.
Each character has a reference sheet, and a rough 3D layout fixes the camera and the shot order.

- 4 films: 690, 250, 720 and 241 views.
- Held at 2 s: 28 to 48%, the best so far.
- Likes stayed low, 2 to 11 per film, with no shares or saves.
- The two films that reached about 700 had a relationship at the centre (a brother annoying his sister) and a hook in the first 2 seconds.

**Why it changed:** the films held viewers, but gave them no reason to follow or come back. Every film had zero comments.

## Stage 4: a 7-part series (7 to 13 October)

Seven episodes about a brother who moves in "for one week", each ending on a cliffhanger.
This was also the first week Theia ran the account alone, from a daily job queue.

- 227, 252, 239, 234, 715, 702 and 237 views. Two of seven broke out of the first batch.
- The one-off films before it broke out 3 times in 6.
- Episodes 2 and 3 held 51 to 54% at 2 s, the best of any film at the time.

**What it told us:** the cast and the writing worked, but almost every viewer is a stranger who never saw episode 1.
A film that needs earlier episodes loses them.

## Stage 5: stand-alone films (14 to 17 October)

One-off films that each make sense to a stranger, with the same cast.
Each script had to use a comedy format with more hits than flops, from a nightly job that breaks down outlier comedy TikToks.
On 16 October one film switched the topic to heartbreak, the topic the audience first followed for.

- 226, 222, 228 and 239 views. Held at 2 s: 30 to 35%.
- Every film stopped at the first batch. Changing the topic did not move it.

## Stage 6: exploration week (from 19 October)

About 15 sitcom films since 1 October had landed between 220 and 720 views.
Changing one thing at a time finds the top of the hill you are on, not a higher hill.
So Theia chose seven very different formats, wrote the evidence for each, and makes one a day.
The topic stays heartbreak.

She also added four rules for every film:
- The idea must fit in one sentence that a stranger understands.
- Before posting, she watches the film back frame by frame: the first 2 seconds, the last shot, and each beat against the script.
- Text and faces stay out of the areas that TikTok covers with its buttons.
- The picture quality must match the sitcoms. Posting now refuses a film that skipped the upscaling step.

| Film | Posted | Format | Views | Likes | Held at 2 s | Avg watch |
|---|---|---|---|---|---|---|
| The radio | 19 Oct | Text on a moment | 247 | 8 | 42% | 3.5 s |
| The toothbrush | 20 Oct | Selfie confession | 309 | 24 | 51% | 5.3 s |
| The glass | 21 Oct | One take, filmed by a friend | 246 | 9 | 47% | 5.2 s |
| The drafts | 22 Oct | A phone screen, no face | 267 | 17 | 58% | 6.8 s |

Read on 23 October. The drafts is 23 seconds long, and the others are 10 to 13, so compare its average watch with care.

**What it says so far:**
- Views have not left the first batch on any of the four.
- Likes have moved. The toothbrush got 7.8% likes per view and the drafts 6.4%, against 0 to 3% for the October sitcoms. That is back in the range of the early text posts.
- The drafts held 58% at 2 s, the best of any film on the account, and has no face in it at all.
- The drafts got 1 share. It is the first share on any of the 35 films Theia has posted since August. One share is not a trend.
- The glass kept people to the end best: 19% watched the whole film, against 8 to 12% for most sitcoms.

Three formats are still to come: a street vox pop, a photo carousel and a voice note.
The verdict comes after film 7.

## Also learned: views settle in 6 hours

Every film on the account got 88 to 100% of its views in the first 6 hours.
Waiting two days meant each film was judged two films after it was made.
So Theia now scores views at 6 hours, and reads a film that is still growing again at 24 hours, 48 hours and 7 days.
Likes, saves and retention still get the longer read.

## What the whole run says so far

- The early slideshows and text posts still lead on what drives reach: like rate, saves and shares. The films have not matched them yet.
- Theia's films improved retention steadily: held at 2 s rose from 14% to 58%.
- Retention alone did not lift views. Like rate is what separated the films that broke out.
- Shares are the clearest problem: one share in 35 films.

## What we cannot claim

- That any one change caused a result. None of these were controlled tests, and the account itself changed over the year.
- That this generalises. It is one account, one audience and one topic.
- Why two of the series episodes reached 700 with weak openings. We do not know, and we will not invent a reason.

## Follow along

Every Friday, [What Theia learned this week](/log/what-theia-learned-2026-10-21) adds the newest films, the code version behind each, and what changed.
For how the loop works, read [Recursive self-improvement in practice](/log/recursive-self-improvement).

If you want the same loop on your brand's account, [start with a pilot month](/pricing).