Forwarded from AARMN The Limitless
YouTube
How Not to Waste VRAM: LLM Quantization (Persian/Farsi)
This is the "LLM and Quantization" Lecture, Conducted by Guilan University on Research Week.
We go from What, to Why, to How. I break down why your GPU is screaming when you run local LLMs, and how quantization is the cheat code that makes AI accessible…
We go from What, to Why, to How. I break down why your GPU is screaming when you run local LLMs, and how quantization is the cheat code that makes AI accessible…
I could finally upload my talk on LLM & Tokenization in youtube! It's persian, but pinky promise, the next video will be English, probably on either of the following topics
1. LoRA, VeRA and DoRA
2. Muon optimizer
3. Some deepseek related stuff
4. An English version to my Persian conferences with some extra material added
1. LoRA, VeRA and DoRA
2. Muon optimizer
3. Some deepseek related stuff
4. An English version to my Persian conferences with some extra material added
In my Operating Systems course (as TA), I assigned a semester project: implement a modified CFS scheduler variant which I called ALFS (Anushiravan-level Fair Scheduler), this varient uses a min-heap instead of a red-black tree. The change makes it harder to copy while adding some fun.
One of my good friends, and in this sem, students, Sepehr, took it far beyond expectations, specially considering current semester incidents. He implemented ALFS in Zig and built an impressive repository that brings the idea to life.
I'm sharing this with pride—seeing a project that started as a lighthearted, somewhat meme-like concept evolve into the increasingly realistic ALFS feels incredibly rewarding. There is surely lots of room for improvement and a few bad choices of DS, but overall, seeing projects like this keep academic work interesting.
https://github.com/Alireza-Sobhdoost/ALFS-ZIG
One of my good friends, and in this sem, students, Sepehr, took it far beyond expectations, specially considering current semester incidents. He implemented ALFS in Zig and built an impressive repository that brings the idea to life.
I'm sharing this with pride—seeing a project that started as a lighthearted, somewhat meme-like concept evolve into the increasingly realistic ALFS feels incredibly rewarding. There is surely lots of room for improvement and a few bad choices of DS, but overall, seeing projects like this keep academic work interesting.
https://github.com/Alireza-Sobhdoost/ALFS-ZIG
GitHub
GitHub - Alireza-Sobhdoost/ALFS-ZIG: A high-performance, event-driven CPU scheduler written in Zig. Designed to replace the standard…
A high-performance, event-driven CPU scheduler written in Zig. Designed to replace the standard Red-Black Tree approach with a Min-Heap for better O(1) access and load balancing on modern Big.LITTL...
❤1
I also like to have a chat with any of my friends or fellow rangers who happen to know a ton about ViTs and Few Shot Classification using Latent generation, or know sb which worked/knows these typa things
Holy mother of databases
https://youtu.be/C7gJ_UxVnSk
https://youtu.be/C7gJ_UxVnSk
YouTube
1000x faster than your database - SpacetimeDB 2.0
Star the repo on GitHub!
https://github.com/clockworklabs/SpacetimeDB
My Twitter: https://x.com/TylerFCloutier
Our Twitter: https://x.com/spacetime_db
Our Discord: https://discord.gg/spacetimedb
Website: https://spacetimedb.com
Referral Program: https:…
https://github.com/clockworklabs/SpacetimeDB
My Twitter: https://x.com/TylerFCloutier
Our Twitter: https://x.com/spacetime_db
Our Discord: https://discord.gg/spacetimedb
Website: https://spacetimedb.com
Referral Program: https:…
🔥1
Mercury 2, A diffusion based LLM
https://youtu.be/Bqdf6Um_8OE
https://youtu.be/Bqdf6Um_8OE
YouTube
Mercury 2: The First Diffusion Model That 'Thinks'"
In this video, I test Inception's new Mercury 2, a diffusion-based large language model that introduces reasoning capabilities and generates text at 1,000 tokens per second. I demonstrate its speed and instruction-following through coding tests, and then…
❤1👍1
Forwarded from Sadra Codes
میبینم این روزها خیلی از بچهها دارن از یک دورهی سخت عبور میکنن. گفتم شاید بد نباشه کمی دربارهش حرف بزنیم.
برنامهنویس بودن ــ و کلاً کار کردن در حوزهی تکنولوژی ــ در کشوری مثل ایران، واقعاً از سختترین انتخابهاست. سالها با این جملهی کلیشهای که «آدم توی محدودیتها ستاره میشه» خودمون رو قانع کردیم و چشمهامون رو روی خیلی از واقعیتها بستیم.
اما واقعاً منظور از ستاره شدن چی بود؟
اینکه برای دیدن ۵ دقیقه ویدئو توی یوتیوب، یک ساعت درگیر پروکسی، VPN، تانل و هزار و یک داستان و کوفت و زهرمار باشی؟
اینکه به اینترنت آزاد دسترسی نداشته باشی؟
اینکه دستت از شبکههای پرداخت بینالمللی کوتاه باشه؟
ولی واقعاً لازم بود مادربزرگ من تو این سن دنبال سرور خارجی و کانفیگ باشه؟ 🙃
گاهی که به گذشته فکر میکنم، میبینم برای رد شدن از بعضی بنبستها و محدودیتها، چه ترفندهایی که نزدم. تازه الان میفهمم اون موقعها چقدر پشتکار و صبوری خرج کردم. چقدر علاقه داشتم که حاضر بودم برای چند دقیقه یاد گرفتن، ساعتها با فیلترینگ و محدودیت بجنگم.
مطمئنم خیلی از شما هم تجربههای مشابهی داشتید؛
چه در نرمافزار، چه در صنفها و شغلهای دیگه.
و در آخر، فقط اینو میخوام یادآوری کنم:
به کم قانع نباشید.
به حقتون قانع باشید.
قبل از ورود به هر کاری، حق و حقوقمون رو بدونیم.
نور بر تاریکی پیروز است.
برنامهنویس بودن ــ و کلاً کار کردن در حوزهی تکنولوژی ــ در کشوری مثل ایران، واقعاً از سختترین انتخابهاست. سالها با این جملهی کلیشهای که «آدم توی محدودیتها ستاره میشه» خودمون رو قانع کردیم و چشمهامون رو روی خیلی از واقعیتها بستیم.
اما واقعاً منظور از ستاره شدن چی بود؟
اینکه برای دیدن ۵ دقیقه ویدئو توی یوتیوب، یک ساعت درگیر پروکسی، VPN، تانل و هزار و یک داستان و کوفت و زهرمار باشی؟
اینکه به اینترنت آزاد دسترسی نداشته باشی؟
اینکه دستت از شبکههای پرداخت بینالمللی کوتاه باشه؟
ولی واقعاً لازم بود مادربزرگ من تو این سن دنبال سرور خارجی و کانفیگ باشه؟ 🙃
گاهی که به گذشته فکر میکنم، میبینم برای رد شدن از بعضی بنبستها و محدودیتها، چه ترفندهایی که نزدم. تازه الان میفهمم اون موقعها چقدر پشتکار و صبوری خرج کردم. چقدر علاقه داشتم که حاضر بودم برای چند دقیقه یاد گرفتن، ساعتها با فیلترینگ و محدودیت بجنگم.
مطمئنم خیلی از شما هم تجربههای مشابهی داشتید؛
چه در نرمافزار، چه در صنفها و شغلهای دیگه.
و در آخر، فقط اینو میخوام یادآوری کنم:
به کم قانع نباشید.
به حقتون قانع باشید.
قبل از ورود به هر کاری، حق و حقوقمون رو بدونیم.
👍1
Rant and NSFW language ahead:
The number of opportunities I lost and I am actively losing because of being born on a specific geopolitical landscape is seriously perplexing to me
No good free trial, no real job opportunity, no path to exit unless you have money, which you can't get cause you didn't exit to have a real job, so your best chances are often wasting away for less than 1000 dollar a month (and that is, super generous ones), while being an skilled worker, and if you are, off of a top echelon, and lucky enough to get inter-fucking-net, and not intra-fucking-net
Till another rant session, or maybe an episode on Iran experience for a developer/researcher in war, bye bye!
The number of opportunities I lost and I am actively losing because of being born on a specific geopolitical landscape is seriously perplexing to me
No good free trial, no real job opportunity, no path to exit unless you have money, which you can't get cause you didn't exit to have a real job, so your best chances are often wasting away for less than 1000 dollar a month (and that is, super generous ones), while being an skilled worker, and if you are, off of a top echelon, and lucky enough to get inter-fucking-net, and not intra-fucking-net
Till another rant session, or maybe an episode on Iran experience for a developer/researcher in war, bye bye!
👍1
Some updates,
I read VeRA paper, I am halfway in DoRA paper, I also started HQQ paper
We gonna have some good Quantization and PEFT related episodes coming-up
But when, is a question only my internet and editor-aryan's pickiness and luck will answer
Btw, this truce has nothing to do with the Internet so I can't care less about it, I'm just glad I have electricity and that's right about it.
Nothing more to say, much love and take care y'all
I read VeRA paper, I am halfway in DoRA paper, I also started HQQ paper
We gonna have some good Quantization and PEFT related episodes coming-up
But when, is a question only my internet and editor-aryan's pickiness and luck will answer
Btw, this truce has nothing to do with the Internet so I can't care less about it, I'm just glad I have electricity and that's right about it.
Nothing more to say, much love and take care y'all
👍1
Randrange Podcast
Holy mother of databases https://youtu.be/C7gJ_UxVnSk
I heard some news that this benchmark is faked, I cannot verify it as the best my internet can do at the moment is to open texts
But, I would put the link for you to verify and come to your own conclusions:
https://youtu.be/dCpEdmueKv0?si=gHm6kM9sFHGxMRwX
But, I would put the link for you to verify and come to your own conclusions:
https://youtu.be/dCpEdmueKv0?si=gHm6kM9sFHGxMRwX
YouTube
The database that's 3000x faster: PostgreSQL 18 (quick check on the SpacetimeDB 2.0 benchmark)
I double-checked the numbers in SpacetimeDB's 2.0 announcement video.
Pull request: https://github.com/clockworklabs/SpacetimeDB/pull/4457 (now closed)
I plan to move my efforts to a dedicated free repository: https://github.com/hogejo/spacetimebenchmark…
Pull request: https://github.com/clockworklabs/SpacetimeDB/pull/4457 (now closed)
I plan to move my efforts to a dedicated free repository: https://github.com/hogejo/spacetimebenchmark…
Hello guys, we will have a surprise episode, as soon as I can edit it. (Un)fortunately, I am currenly un-unemployed and therefore, I cannot be that active here (which I never was to begin with anyways!), but its a low edit episode, I just need to ask for some snippets of code from sb ;)
The lowest hanging fruit is picked
https://www.youtube.com/watch?v=DlCoCfTDzMI
https://www.youtube.com/watch?v=DlCoCfTDzMI
YouTube
Thinking Machines Just Solved Real-Time AI Interactions!
Thinking Machines just changed the turn-based AI paradigm by introducing a new architecture that tokenizes time into continuous 200ms micro-turns. Let's dive deep into the technical paper to explore their encoder-free early fusion approach and see why it…