alimafana

Last week I called the Gemma 4 MoE-vs-Dense divergence architecture-mediated. The comment thread said my max_tokens cap was starving the reasoning layer. I re-ran with the budget raised. Dense recovered on every scenario. Walking it back.

I ran Gemma 4 26B (MoE, 4B active) and Gemma 4 31B (dense) against GPT-4o and GPT-4o mini on a real Arabic e-commerce chatbot. Then I added three Gemma-only prompt rules. The MoE variant opened the catalog. The dense variant stopped reading it. Same prompt, two architectures, opposite failures.

A grounded LLM bug looked exactly like an architectural failure. The search router was returning the right data. The prompt was telling the model to ignore it. The fix changed nothing about the search code — it changed one variable label.

Postgres tells you how many times each index was scanned. It can't tell you which one the planner picked for your slow query. I built a 3-file analyzer — and the first run found 20 of 51 indexes never used, 78% of my index disk wasted.

OpenAI's billing page shows total spend. It doesn't show which feature, which tenant, or which conversation caused it. I built a 3-file monitoring system that does — and the first time I opened it, I caught a 100× cost gap between two features I'd been treating as similar.

TL;DR: I spent a full day debugging my AI sales chatbot. What looked like one bug turned out to be...

Three weeks ago I left a comment on a Dev.to article. Today, that comment turned into a full...

TL;DR: I split one big GPT-4o-mini call into two small, specialized calls. Tokens per message dropped...

I'm Ali, building Provia — an AI sales platform — from Gaza. I'd spent 8 sessions building features....

I'm Ali, building Provia — an AI sales platform — from Gaza. This bug could be silently killing your...

I'm Ali, building Provia — an AI-powered sales platform — from Gaza. This is the hardest lesson I...

I'm Ali, building Provia — an AI-powered sales platform — from Gaza. This is one of the bugs that...