Mohammad Alaggan, Ph.D.

Sr. Software Development Engineer at @AWSCloud (CloudFront). Opinions are my own.

[Mohammad Alaggan, Ph.D. @malaggan]

My handle on bsky is malaggan

[Mohammad Alaggan, Ph.D. @malaggan]

@AhmedGhazey @awscloud بعت لحضرتك DM

[Filip Jerzy Pizło @filpizlo]

I wrote a document describing how Fil-C's `longjmp`, `setjmp`, `setcontext`, `getcontext`, `makecontext`, and `swapcontext` implementations work while maintaining all of the memory safety guarantees

https://t.co/2iWGfKvAC8

[Jaana Dogan ヤナ ドガン @rakyll]

I think the hardest part of working for gigantic projects or companies, it's hard to take a break. When you stop making decisions, someone else is making decisions for you.

[k h ô i @khoiiiind]

In which Tom Leinster advocates for replacing ZFC with a topos-flavored version that feels much more natural to the working mathematician.

[Mohammad Alaggan, Ph.D. @malaggan]

الكود بيتعمل في يوم بالذكاء الاصطناعي واقعد اسبوع انضف فيه عشان يبقى جاهز للنشر

[Wyatt Benno @wyatt_benno]

They say "formal verification".
We say "What kind?" 👀

[Mohammad Alaggan, Ph.D. @malaggan]

@Islamx ايه ده هو حضرتك في فانكوفر؟

[OS Dev @OSdev_]

The UNIX Time-Sharing System (1974) - The Classic

Dennis Ritchie and Ken Thompson's most famous paper.

Introduces the UNIX philosophy
Hierarchical filesystem (/usr/bin/...)
Processes
Fork/exec model
Shell
Pipes
Multi-user design

This is arguably one of the most influential papers in software history

https://t.co/r01bihUdSg

[nathan.rs @nathanrs]

I found out the other day that any compression tool can be contorted to do language modeling. Turns out gzip can generate text that somewhat *resembles* Shakespeare. Short write up linked below

[Mohammad Alaggan, Ph.D. @malaggan]

حقيقة الأمر أن التعليم ونظمه انعكاس للمجتمع والنظام الاقتصادي ككل: كونه قائم على وظائف معظمها غير إبداعية ولا تتطلب سوى القدر اليسير أو أقل من التفكير النقدي.

[Mohammad Alaggan, Ph.D. @malaggan]

لما انتشرت الآلات الحاسبة، لم تعد مهارة الحساب اليدوي مفيدة. ولكن هل هدف التعليم هو غرز مهارات؟ أم هدفه غرز ملكة التفكير النقدي وتوسيع المدارك وسعة الإطلاع؟ لو هدفه هو الأخير فانتشار الذكاء الاصطناعي فرصة ذهبية وليست عائقا، بل وهو المطلوب لنجاح الطلبة في سوق العمل.

[سماهر الضامن @samaherdhamen]

أنا أعاني حقا منذ تعرفت الطالبات على تطبيقات الذكاء الاصطناعي. ما عدت أرى أخطاء في الكتابة، ولا محاولات تفكير، ولا مغامرات أو قراءة مغايرة للمألوف، ولا حتى فضولا في القاعات.. اختقت لحظات الدهشة، وأصبح كل شيء مسطحا ومألوفا و مهذبا.. المشكلة الحقيقية تظهر في أوراق الاختبارات. 90% من الطالبات، وأنا متفائلة، لا يستطعن كتابة سطر واحد سليم. طيب، لنفترض أن العملية التعليمية ماشية على أكمل وجه، ما الفكرة من أن أقيِّم بحثا مولدا بالذكاء الاصطناعي؟ ملف معايير التقييم يفترض أن أقيّم كتابة الطالبات ومهارات أخرى قابلة للقياس، لا مهارات استخدام أدوات الاصطناعي.. يبدو لي أن المؤسسات التعليمية أخذت على حين غرة، فإذا كانت الجامعات تسمح، أو تشجع ضمنيا، على استخدام أدوات الذكاء الاصطناعي، والطلبة يستخدمونه بالفعل، فماذا نقيّم نحن بالضبط؟
ليس لدي موقف متشدد من استخدام تطبيقات الـ AI، أنا أتساءل فقط.. شكرا لانتباهكم :)

[Mohammad Alaggan, Ph.D. @malaggan]

كارثة الذكاء الاصطناعي في التعليم كشفت مدى ضعف النظم التعليمية الحالية وفشلها في تقديم تعليم حقيقي. اصل المشكلة أن السواد الأعظم من المدرسين ووزراء التعليم لا يدركون اصلا ما ينبغي للتعليم أن يكون، وأنه ليس مجرد مفرخة لتقييم النشأ على مسطرة واحدة.

[Erik Meijer @headinthebox]

Probably this book should be forbidden https://t.co/dFZiSrCVHh. Read it before it disappears just like your favorite LLM ;-)

(via https://t.co/tva8fgPLRk, not on X)

[Mohammad Alaggan, Ph.D. @malaggan]

Very interesting anecdote

[Taelin @VictorTaelin]

before, let me be precise about the state of Bend2:

initially, Bend2 was written by me, with GPT 5.5, based on a new GPU compilation architecture that I expected would be 10x faster than Bend1. I used 5.5, progress moved FAST, and I quickly validated the architecture. everything worked.

then, I got stuck. I realized the code was in a poor state. not because GPT 5.5 failed my prompts. but because it always built the most *narrow* way to solve anything. for example, compiled closures would work for U32, but wouldn't mix with mutable arrays. likewise, mutable arrays would work on GPU shaders, but would fail on CPU, because, for whatever reason, both modes ended up being compiled in entirely different way. and so on

so, things worked, but code was poor. I was not happy with any of that, and that is VERY far from my quality standards. so, I considered my usage of GPT 5.5 a failure, decided to delay the launch, and started re-doing it MANUALLY.

then, opus 4.8 launched with 3x cheaper fast mode. that was VERY helpful with my "manual" refactor. I started using opus 4.8 to help me with it, but I didn't leave it working alone anymore. instead, I just told it to do something specific (ex: add a location field to LTerm variants, and include located errors on the checker), let it implement quicky (it is FAST) and READ the whole output.

whenever it failed - and, yes, it failed a lot, and did dumb shit a lot - I could fix it quickly. this worked well because I had a super fast coder, but I was still 100% in charge, and aware about everything that was going on. I quickly managed to rewrite the whole lang into a much cleaner codebase. I'd say 20% was completed in 2 weeks or so (?) with Opus

then, enter Fable. it was like Opus fast mode, except it actually made no mistakes. I was still reading its outputs, but it actually just... made no mistakes. I mean, there was a thing or another, but, mostly, it just did what I asked in the most reliable way possible. so, in 3 days I'd say I went from 20% to 60% or so. and I started feeling like I would complete it all this week, possibly as soon as tomorrow!

then, the ban came. I was obviously very frustrated, as you can all see. over the weekend, I tried resuming the job with Opus 4.8 and GPT 5.5. problem is, the codebase is now large enough that they start making TOO MANY mistakes. the rate of mistakes per call is too large, and, since I'm now reading every output, I could see most of my prompts were DEGRADING the codebase.

example: I ask Opus/GPT to finish the U64 implementation. it does. but at the same time, it breaks the unboxed term layout. because it either didn't read the whole codebase to fully understand it, or because it did, but then got overwhelmed by the context explosion.

so, basically, both models are not able to meaningfully work on Bend's codebase without introducing damage and undesirable side effects. it is just wasting my time. I ask them to do one thing, they break something else. I ask them to fix the thing they broke, they undo the thing they did initially. and again, this is 100% a context size issue. the whole project is just too large for them.

don't add the full context → they make bad assumptions
add the full context → they get overwhelmed / degrade

so, yes, 5.5 and 4.8 are NOT going to help me finish Bend2. so, the option I have left is to do it manually. but honestly it is too much code, my tiny hands are tired and, at least for now, I'd rather have a break. if fable is back, I'll use it to quickly finish the job. if it never comes back, I'll take a deep breath and finish it later this month.

ultimately the repo is in a VERY good state right now and I want to keep things as they are, rather than risking letting the AI introduce damage and fuck up again

[Mohammad Alaggan, Ph.D. @malaggan]

@Al_Qasem هو ممكن يكون الهدف الحقيقي اشتراط ربط الهوية بحساب مواقع التواصل الاجتماعي، عشان يعرفوا هوية اي حد اونلاين.

[Luiz Pessoa @PessoaBrain]

𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝗱𝘆𝗻𝗮𝗺𝗶𝗰𝗮𝗹 𝘀𝘆𝘀𝘁𝗲𝗺𝘀 𝗶𝘀 𝗵𝗮𝗿𝗱
But lots of progress being made; this one looks very interesting.
https://t.co/OkBDB8qaKS

[Yuri Sagalov @yuris]

Really important advice for aspiring founders to internalize:

“The way to get the very best startup ideas is not to look for startup ideas. If you're consciously looking for startup ideas, it will make you too conservative. You'll lop off the outliers. Because the very best startup ideas tend to sound so lame, at first, that you'd reject them if you were consciously looking for startup ideas…”

“…So how do you find startup ideas without looking for them? By working on projects with your friends. That's where the very best startups come from. Initially they're not even meant to be companies. They're just something people built because they thought it would be cool”

[Mohammad Alaggan, Ph.D. @malaggan]

Economic inequality is a problem only if the lowest level is not living a decent life with sufficient safety, both physical and psychological. Lifting the bottom half up should be the goal, instead of bringing the top few down - they are the ones creating most jobs and adding most value to the economy. The corrupt should be persecuted, regardless of their position in the economic hierarchy. Politicians vilifying the rich is merely a distraction from the real reason of inequality: their failure as politicians in creating a fair and balanced system. Instead of admitting it, they deflect the rage of the masses by redirecting it at an enemy other than themselves.

[Mohammad Alaggan, Ph.D. @malaggan]

@sharno3 اللي شايفه في امازون في ال Individual Contributors اللي principal وأعلى انهم متعمقين تقنيا اكتر من ال senior وما دونه