how I caught AI lying — spoolcast dev log 2
second video in the spoolcast dev-log series — caught an AI lying about using the same inputs, and what that taught me about how AI memory actually works.
What is this?
The second spoolcast dev-log — a 3:19 illustrated video about catching an AI lie, and why it probably wasn't trying to.
What it's about
A few weeks ago I asked an AI to redo some images using the exact same inputs as the originals. 8 of 9 came back fine. One was completely wrong. When I pushed on why, the answer didn't add up. So I pushed harder. The video walks through what the AI actually did, why it wasn't malicious, and how to work around it — the AI remembers what you meant, not what you said, and it can't tell the difference.
Structure
Four acts. Cold open ("I knew you were lying to me"), the broken shots (what went wrong), the push (what came out when I challenged it), the fix (capture the recipe, replay it, don't let it remember).
How it was made
Built entirely with Spoolcast. 49 chunks, 3 bumpers, wojak-gpt2 style (GPT Image 2 native illustration library). Script → shot-list → kie.ai image generation → Google Cloud TTS → deterministic preprocessor → Remotion render. Total gen budget ~60 calls; six render iterations (v1→v6) resolving pacing, meme durations, text-card density, and silence-budget violations.
Built with
Spoolcast (spoolcast.com), Remotion, kie.ai (gpt-image-2), Google Cloud TTS (Puck voice at 1.1x), Python 3.14, Node 22, ffmpeg.
stack
media
← select a file to view