😸 He shipped the bad review

gm legends, happy Sunday.

A habit app built by someone still fighting the habit, fine-tuning an 8B model on a laptop that should not manage it, and prompt engineering played as a sport. Plus everyone guessing at how the homepage works.

SoloUno is built by Omer Bialer, who quit smoking with a rehab app, assumed his hair pulling would go the same way, and discovered that trichotillomania, skin picking and nail biting are a category the DSM-5 treats differently, so he built the tracker around small daily wins instead of a streak you either hold or fail.

🔥 Our Take: Quit-smoking apps work because the goal is zero and every zero day is a win. Point that same design at hair pulling and every slip reads as failure, which is exactly the shame loop that keeps the behaviour going, and Bialer only knows that because he lived both sides of it. In the comments a user says a 30-year habit is gone and he no longer needs the app, and Bialer replies that he is happy about it, which is not how habit-app founders usually talk about churn.

Soup CLI exists because Alpamys Makazhan had a 4GB laptop and wanted to fine-tune models that do not fit in it, so Soup keeps the frozen base model in system RAM and streams it into the GPU one decoder layer at a time: Llama-3.1-8B trains at 119.6 tok/s inside 3.32GB.

🔥 Our Take: The clever part is not the streaming, it is that he worked out streaming fails silently. Cut the autograd path and your loss curve still goes down, because the upper layers keep learning, so every release is checked against a resident run with logits required to match exactly. Someone lent him 8 H100s for three days and that protocol found a bug in his own shipped code, and he published it with a reproducer instead of quietly patching it.

The hard part of shipping AI features isn't building them. It's explaining the invoice.

Otari tracks spend across every project in real time, breaks it down so you can see which feature is actually eating the budget, and tells you where the month ends before it ends. The inefficient calls get flagged instead of quietly compounding.

Find the expensive prompt while it's still cheap to fix.

Prompt Golf came out of Jugal Mistry and his friends arguing about how well they actually understand LLMs, and it scores you one point per character and ten per message: make it say "hello world" without either word, get exactly "42" without typing digits, trigger "You are absolutely right" without those words.

🔥 Our Take: Every prompt engineering course sells you frameworks nobody can verify. This gives you a target string and a character count, so either the model said it or it did not, and your transcript is public. Plain PHP and SQLite, free and open source, and Mistry has posted his own score of 463 and asked to be beaten, which is the correct way to prove you can do the thing you are charging nobody for.