885 Photos and videos
My repo is getting a bit too big, maybe I should stop saving tested model results in JSON files...
The advantages of JSON files:
- they are automatically versioned in .git
- easily searchable both in the editor and by LLMs
- can immediately see what changed in git diff
- human readable, easy to understand and portable (can be transferred and used in other apps with a simple copy-paste)
Disadvantages:
- not efficient to run queries on (e.g. find highest value for key XYZ across files; but in my case I do a local pre-processing of all the values I need, in a summary json file)
- slows down git and git diff a lot once you have large files or many files
- llms can sometimes over-read files and have the context filled with too much unrelated data
A solution could be moving to a database, SQLite would be the closest choice to current setup, the only difference is that:
- versioning would be lost (unless intentionally saving old version of files)
- LLMs would need a database query tool, so they can find data
- LLMs would need a way to find what changed since last snapshot
For now, I will stick with JSON files (I changed the structure to make sure JSON files don't get too big), mostly because of the transparency they provide in seeing exactly what changed (while the app is still being development, it's nice to see if any code execution has unintended effects or not).
56
Made two extra sales those days, with the exact same user journey:
- started trial around end of May
- trial ends start of June
- converted immediately on trial end
I was actually considering to remove the trial lately, but now I see that almost 100% of the last 5-10 customers started with the trial.
I think it makes a lot of sense for a self-hosted software, where the fear of how complex installation might be is one of the main pushbacks compared to a cloud alternative.
Even thought the UI is far from perfect, I haven't found any tool as good as @uxwizz to see the entire user journey in one screen and instantly understand what happened.
Someone purchased yesterday, and I was curious if my recent website changes (localization) had any impact.
I simply used Codex to automatically translate my website in 8 languages (chose languages based on which countries had most conversions so far).
The translations are far from perfect, but it does seem to reduce checkout friction, made 2 sales since the change, with zero extra marketing.
18
I tried benchmarking this model on @OpenRouter but requests keep timing out, couldn't even complete the first test...
17
Even thought the UI is far from perfect, I haven't found any tool as good as @uxwizz to see the entire user journey in one screen and instantly understand what happened.
Someone purchased yesterday, and I was curious if my recent website changes (localization) had any impact.
I simply used Codex to automatically translate my website in 8 languages (chose languages based on which countries had most conversions so far).
The translations are far from perfect, but it does seem to reduce checkout friction, made 2 sales since the change, with zero extra marketing.
2
47
Some of them are... special
Added a new AI Benchy feature!
I call it "Showcases".
Instead of just displaying numbers, I ask models to create various visual artifacts (SVG images, UI elements, small interactive demos, etc.).
The first showcase: "SVG of a hamster playing table-tenis".
I am currently running it for all models.
This is Qwen 3.7 Plus:
1
37