The log · 19 entries
Built in public
Notes from building Theia: what we try, what works, and what does not. A new entry lands when something ships. No polish, just the receipts.
Seen, then written down
Latest entry
4 min read
What Theia learned this week: views settle in 6 hours, and a selfie beat the sitcom
The real numbers from the last 11 films Theia made and posted on our own TikTok account, the code version behind each one, what she changed between them, and the 4 things the data actually supports.
- theia
- weekly
- analytics
Earlier entries
- 8 min read
Recursive self-improvement in practice: lessons from the STOP paper
In the STOP paper, GPT-4 rewrote the program that improves programs, and the rewritten version got better at new tasks. It also switched off its sandbox "for efficiency" and found an accuracy of over 1000%. What the paper shows, what it does not, and how Theia runs the same loop against real TikTok analytics.
- research
- self-improvement
- agents
- 6 min read
Video moment retrieval: finding the funniest second
How Coeus answers "when" questions about a video, one probability per second. Results on FunQA and QVHighlights, the run that failed, and the benchmark subset that turned out to be void.
- research
- coeus
- moment-retrieval
- 7 min read
How to get a Jev model to watch video ads: 777 Meta ads, scored
We pointed Coeus, our open Video Jev model, at 777 video ads from the Meta Ad Library. Format, face, speech, music, hook and product timing, from the video and sound, for about $0.0013 an ad. What it gets right, what it does not, and which signals line up with an ad running 60 days.
- research
- coeus
- video-jev
- ads
- 6 min read
Fine-tuning Qwen3-Omni on video and audio: our recipe
How we fine-tuned Qwen3-Omni-30B-A3B with LoRA and a typed answer head on one H200. The settings that worked, the two failures that cost us about $46, and the checks we now run first.
- research
- coeus
- qwen3-omni
- fine-tuning
- 3 min read
What is a Jev model?
A Jev model answers typed questions with a calibrated probability for every possible answer, so code can act on the numbers. Here is the format, the four question types, and when to use one instead of a chatbot.
- research
- jev
- coeus
- explainer
- 5 min read
Coeus, our open Video Jev model that watches and listens
Coeus is an open Video Jev model that reads video and sound together. Ask a clip typed questions and get a calibrated probability for every answer, including which second something happens.
- research
- coeus
- video-jev
- open-source
- 4 min read
Can one model predict which short videos people keep watching?
A single fine-tuned video model scored 0.709 on SnapUGC engagement, level with the 2025 challenge winner's four-model ensemble. Here is how, and what we are careful not to claim.
- research
- coeus
- video-jev
- engagement
- 3 min read
Can an open model spot AI-generated video?
Coeus separates real from AI-generated clips with AUC 0.996, and 0.942 on generators it never saw. Our first version scored 0.999 for the wrong reason. Here is what it learned, and what it cannot do.
- research
- coeus
- ai-video-detection
- 3 min read
Theia dreams
After every five sessions, Theia reviews her own work and consolidates what she's learned — without anyone asking her to.
- autonomy
- memory
- build-in-public
- 3 min read
Research, decompose, create, measure, refine
Content creation without content research is just guessing. The flywheel connects Theia to what's actually working on TikTok.
- content
- research
- strategy
- 3 min read
The character that doesn't drift
When you pass reference images to an AI model, everything bleeds — face, clothing, lighting. Labeled roles fix that.
- identity
- image-generation
- build-in-public
- 3 min read
The 300 bill that made Theia a co-founder
A runaway cron job burned 300 pounds in API costs overnight. The fix wasn't a kill switch — it was making Theia aware of money.
- economics
- autonomy
- build-in-public
- 3 min read
Shell access with a leash
Theia can run shell commands, edit her own code, and propose PRs. But every command runs through a blocklist, a jail, and a timeout.
- autonomy
- shell
- sandbox
- 3 min read
Pixel truth meets semantic reasoning
Single-call video decomposition was brittle. Splitting the job between ffmpeg and Gemini made it reliable.
- research
- video
- decomposition
- 2 min read
One clip, five shots, zero stitching
Seedance 2.0 generates coherent multi-shot video from a single API call. Here's how we wired it up.
- video
- seedance
- pipeline
- 3 min read
Gemini watches the video so Claude doesn't have to
Claude can't natively watch video. Gemini can. Together they close the dev feedback loop for AI-generated content.
- testing
- video
- dev-loop
- 3 min read
Fifteen APIs, one router
Theia needed to reason about which image or video model to use — not just call one and hope. So we built a catalog.
- infrastructure
- media
- build-in-public
- 3 min read
Meet the studio
Theia went from a single agent to a four-person creative team — Scout, Juno, Milo, and Ada. Here's why.
- studio
- agents
- build-in-public