Ahmed

Software Engineer and free software advocate, my views are my own. English & بالعربي .

[Ahmed @ahmedsbytes]

@mtantawy الاسبوع الجاي في ناس من big movers هاتطلع ارباحها
بنسبه كبيره هايبقي ارقام كبيره و spaceX داخل
تخميني ان دي ناس بتاخد ارباح الصعود الكبير اللي فات وهاتشتري تاني يعتبر بخصومات قبل ما ارباح الاسبوع اللي جاي تنزل

[Ahmed @ahmedsbytes]

@ahmd3ssam big != good :(

[Ahmed @ahmedsbytes]

Most of the big models have this annoying thing, 1m context models , usually deviate from instructions
The phenomenon has two names:
"Lost in the Middle" (Stanford, 2023), models show a U-shaped performance curve. They're good at the beginning and end of context, terrible in the middle. At just 4,000 tokens, accuracy drops from 75% → 55% based purely on where the information sits.
"Context Rot" (coined mid-2025 on HN), models genuinely get dumber as context grows, even when the task stays the same.

[Ahmed @ahmedsbytes]

@ahmd3ssam tbh , you are over thinking it , they are just trying to go for small investors, since big investors will mostly avoid tesla much , and smaller ones will want to run to buy tesla.

[Emad Elsaid @emad__elsaid]

10 days of daily streaming on Twitch, and I haven't quit yet 😅
Honestly, when I started on 10 days ago I wasn't sure I'd stick with it. Turns out streaming my programming sessions and gaming is awesome. We already have regulars I know by name.
https://t.co/4isTRDi95h

[Ahmed @ahmedsbytes]

@Bashmohandes لسه حاجز focus time blocks علشان نفس السبب
وما أخذ بالقوه لايسترد بغير القوه

[Thariq @trq212]

https://t.co/R6exTuF7P8

[Ahmed @ahmedsbytes]

@mohewedy هاجربه الاسبوع دا , تسلم جدا جدا❤️

[Ahmed @ahmedsbytes]

استخدام التوكيز بتاعي الشهر اللي فات
~4.88B
دا علي deepseek pro v4 بتاعهم , كان تمنه 120 $
دا كان هايبقي علي OpenAI , اقل تقدير ~$4,900
وعلي كلود بيك , بارخص موديل عندهم ~$4,880
بيتعبني شوية
وشغلته علي claude cli
ومجبر اني اعمل context engineering ,
بس احسن ROI بالنسبه ليا

[Ahmed @ahmedsbytes]

@attiamo1 انا تحت امر معاليك , انت تأمر يا كبير❤️

[Ahmed @ahmedsbytes]

مش المفروض تشتري عربيه فيراري علشان تشتري الخضار بيها
مش المفروض تستخدم Famous Frontier AI Models لكل حاجه
https://t.co/5bcADrppTe

[Ahmed @ahmedsbytes]

@a6ni @CommandCodeAI تسلم يا غالي ❤️

[Ahmed @ahmedsbytes]

@rebelfigure86 Nope ,
1. U configure claude cli to use ds https://t.co/tF1ViIMLs7
2. the other way around. You configure hermes to use Claude, it has plugins to do this

[Ahmed @ahmedsbytes]

@rebelfigure86 True.
I used Hermes and Openclaw with DSv4, and the lesd annoying and more productive is : hermes + deekseek v4 pro + claude cli (you can make it work with deepseek v4 pro api enf points)

[Ahmed @ahmedsbytes]

You are being charged for "thinking" tokens as well on Opus, also opus is actually greedy token consumer.
Assume , we even have error margin of 20% in the calculations in favor of opus , still opus would be serverl times expensive.
What you can consider , is using opus for the planning step and glm or deepseek or other cheaper model for execution
https://t.co/iimngy180A

[Ahmed @ahmedsbytes]

اغلب الاستخدامات ، تكون 75% input / 25% output
عامة بدون تدخل ، اكيد في مهام بيكون opus بارع فيها
لكن بالنسبة لل complex reasoning ، حتي opus أحيانا كتير يحتاج للتدخل
في شركات بتحاول دلوقتي انها تخلط بين AI Agents
تستخدم opus 4.8 to make the heavy planning step
And Deepseek or glm or kimi for execution
https://t.co/iimngy180A

[Ahmed @ahmedsbytes]

@yghonem14 بالظبط

[Ahmed @ahmedsbytes]

@mkelzeer You can use it from other providers , hosted in the US or so ,
(or even any other cheap open weight model)
https://t.co/YvmjiKKXjr

Also openrouter has this option to , to select model providers based on privacy preferences but not their quantization unfortunately .

[Ahmed @ahmedsbytes]

@ahmd3ssam claude cli , using https://t.co/tF1ViIMLs7 (bunch of claude skills on top of it ofc)
+ hermes agent