01
New entries as new models ship.
When a new model or major update arrives, the catalog is fed what turns up: new mistakes, changed ones, and old ones that seem to have gone quiet.
The AI mistakes catalog · in development
A catalog of how AI goes wrong, built to keep growing.
FailureExtractor is a living reference to AI mistakes. It is not a one-time list. It is designed to grow as new models and updates ship, so it changes alongside the AI it describes.
- Built to grow, not a snapshot
- In development
- Price not set
Why a living catalog
AI changes. A list that doesn’t change goes out of date.
New models and updates keep arriving, and the mistakes people run into change with them. A catalog written once describes the models of the day it was written. A reference you can lean on has to keep moving.
That is the whole idea behind FailureExtractor. New mistakes are looked for when new models ship, and they are added to the catalog as they are observed. It is meant to stay useful by continuing to grow, not by pretending to be finished.
This describes how the catalog is designed to work. It makes no promise about how often it is updated or how complete it is.
How it is designed to grow
Three habits built into the catalog.
02
Built from observation.
Entries come from mistakes actually observed in real AI work, not invented examples. New ones are added by me, one at a time, as they appear.
03
Older entries stay.
The design keeps earlier entries and records where each was seen, so you can follow how mistakes shift from one generation of models to the next.
What an entry is designed to hold
Patterns, not anecdotes.
Related mistakes are grouped into families and variants. Each entry is designed to say what happened in plain words, how to recognise it, and the conditions that tend to bring it on.
- Family
- Variant
- What happened
- How to recognise it
- What tends to trigger it
- Where it was observed
- Countermeasures to try
Entries describe patterns of AI behaviour. They are not meant to contain anyone’s private conversations or personal details.
Entry format · illustrativeEXAMPLE
- Family
- False completion
- Variant
- Reports a task as finished without showing the check
- Recognise it
- “Done” appears with no file, test result or live check attached
- Tends to appear
- Sessions with several open items at once
- Observed on
- [model and version, date]
- To try
- Ask for the evidence before accepting “done”
Countermeasures
Two kinds of response to try.
Alongside each mistake, the catalog is designed to list things worth trying. They fall into two groups.
What you give the AI
Instructions, memory and skills: wording and standing guidance you add to your own AI setup so it handles a known pattern differently.
How you work
Workflow changes: a check you add, a step you reorder, a habit you change in how you run AI work.
Countermeasures are things to try, not promises. Models change, and nothing here guarantees a result.
Where it stands today
In development.
This page describes what the catalog is for and how it is designed. The first public entries have not been released yet.
- Stage
- In development
- Public entries
- Not released yet
- What you will receive
- Not decided yet
- Price
- Not set yet
- Launch date
- Not set
Joining the waitlist is free and doesn’t commit you to buying anything.
FAQ
Questions, answered plainly
Is this just a list of AI mistakes?
No. A list is a snapshot. FailureExtractor is designed as a living catalog that keeps growing as new models and updates ship.
How often is it updated?
No schedule is promised. New entries are added as new mistakes are observed.
Does it cover every AI model and every mistake?
No. A catalog like this can only record what has been observed. It makes no claim to be complete.
Will it make my AI stop making mistakes?
No. It describes mistakes so you can recognise them, and lists countermeasures to try. It doesn’t change how any model behaves, and nothing here guarantees an outcome.
What will I actually receive, and what will it cost?
Not decided yet. This page will say when it is.
How does it relate to Mothex?
They are separate products from the same publisher. Mothex is a method for running multi-session AI projects with written state. FailureExtractor is a reference for the mistakes to watch for. You can use either one without the other.
Who makes it?
Mateus Ventura, a working name, publishing as a one-person business. The catalog is meant to be judged on what it contains, not on who is behind it.
Is it connected to an AI company?
No. Lizvo is independent and isn’t affiliated with or endorsed by any AI vendor.
Follow it as it grows.
Join the waitlist and hear when the first entries are public. No commitment, and no price has been set.