الذكاء الاصطناعي ساعدني باشياء كثير و واحد منها اني طلبته يفحص جهاز بالكامل اذا فيه اي شي خطير، المهم خلص الفحص و طلع لي برنامج تعدين بجهازي من 2023 ماكنت داري عنه يعدن بالخلفيه، قلت له سوي له هندسة عكسية و فعلا طلع لي المحفظة الي تستخدم جهازي للتعدين
Recently I've been developing a neural network training method that's faster than the traditional optimizer (AdamW) with the same or better accuracy. Me and opus 5.5 trained it, tested it on datasets it had never seen, and here's @claudeai's report of the process.
- On text, AdamW with weight averaging reaches the target a little sooner.
- Training the helper took about 10 GPU hours, once. Every training run after that saves the 13 tuning runs.
- In normal use the helper's image test loss is worse than AdamW's (it is overconfident), although its accuracy is higher. At the limit its loss is better too.