More from my site

If you enjoyed this reflection on the state of AI research hardware, you might also be interested in my Research Papers where I share the results of my work, or my Main Blog for deeper technical dives. You can also return to my Homepage.

Learning from First Principles: أحمد فرغل’s رست للغلابة Series

I found Ahmed Farghly’s Arabic Rust course, and it’s been transformative. Each video is ~3 hours of deep systems thinking. I’ve attempted Rust twice before—read the official book years ago—but something didn’t click until now.

What makes this series special: it’s not teaching “Rust syntax.” It’s teaching systems thinking through Rust. Ahmed explains how Operating System concepts—memory layout, alignment, allocators—become visible in Rust code.

Example from video 1: When you define a struct like this:

struct Data {
    a: u8,
    b: i32
}

He doesn’t just say “it compiles.” He shows you:

• Why the order doesn’t matter to Rust (alignment rules)
• How the compiler pads the struct in memory
• Why this is different from C (where order matters)
• What system calls happen under the hood
• How this connects to heap vs stack allocation
• Why Rust has macros at all (they’re solving real compiler problems)

Concepts that seemed abstract before—lifetimes, borrowing, ownership—suddenly make sense because you understand the memory model underneath. That’s the key: Rust semantics emerge from lower-level realities, not arbitrary restrictions.

This series is rare—especially for Arabic speakers. ⭐⭐⭐

References

1. رست للغلابة

Learning Roadmap

Phase 1: Lexical Methods

• Word2vec and n-gram theory
• Stemming and tokenization for Arabic
• TF-IDF and BM25 ranking
• BMX (dense-sparse hybrid from Mixedbread)

Phase 2: Sparse Embeddings

• SPLADE and learned sparse vectors
• Training sparse embeddings with SBERT
• Optimization and interpretability

Phase 3: Dynamic Programming & Alignment (current)

• Minimum edit distance (Levenshtein)
• Word alignment algorithms
• Viterbi and HMMs

Each builds on the last. You can’t understand why sparse embeddings matter without understanding BM25. You can’t implement them efficiently without mastering dynamic programming.

First Chapter: Tokenization, Unicode & Minimum Edit Distance

Chapter 1 of Jurafsky covers:

• Words and tokens: What counts as a “word” when you’re dealing with Arabic morphology?
• Unicode and encoding: How different languages store characters differently (critical for Arabic preprocessing)
• Regular expressions: Practical NLP pattern matching
• Minimum edit distance: The algorithm that powers spell-checking and text similarity

The minimum edit distance is brilliant but requires implementation from scratch. You need to think about:

• State representation (the matrix)
• Recurrence relations (how cells depend on neighbors)
• Backtracking (to recover the actual edits)

I couldn’t implement it cleanly at first. That’s when I realized: my problem-solving muscles are rusty.

Building: Smart Arabic Keyword Checker

The alignment chapter inspired a concrete project: build a spell checker for Arabic that understands context and domain.

Current keyboards (suggest random near-misses). A smarter version would:

1. Technical terminology mapping
   • Suggest Arabic equivalents for ML terms
   • Example: “embedding model” → “نماذج التضمين”
   • Example: “vector database” → “قاعدة بيانات المتجهات”
2. Persona-aware suggestions
   • Different recommendations for engineers, doctors, writers
   • Your vocabulary profile informs what “correct” means
3. Federated learning (like Gboard)
   • Learn from your typing without uploading data
4. Arabic grammar correction
   • Not just spelling; handle diacritics, agreement, tense
5. Adaptive learning
   • Improve suggestions based on what you accept/reject

The tech stack: dynamic programming for edit distance, sparse embeddings for semantic similarity, possibly Rust for performance.

Why Tokenizer Evaluation Matters

After reading the comprehensive guide on GPT tokenizers, I realized that tokenization is often overlooked but critically important. Poor tokenization can lead to:

• Inefficient use of limited context windows
• Higher computational costs (you pay per token!)
• Worse model performance, especially for non-English languages

For Arabic specifically, tokenization is challenging because:

• Arabic has rich morphology with prefixes and suffixes
• Different dialects (Egyptian, Levantine, Gulf) have varying vocabulary
• Diacritical marks (tashkeel) add complexity
• Tasks that need tashkeel (diacritical marks): speech synthesis, grammar correction, poetry analysis

The Evaluation Framework

I built a simple evaluation function to measure tokenizer quality:

def evaluate_tokenizer(text, tokenizer):
    number_of_tokens = len(tokenizer.tokenize(text))
    number_of_bytes = len(text.encode('utf-8'))
    number_of_words = len(text.split(" "))
    fertility = number_of_tokens / number_of_words 
    compression_ratio = number_of_bytes / number_of_tokens 
    return {
        "fertility": fertility,
        "compression_ratio": compression_ratio,
        "total_tokens": number_of_tokens
    }

The Contenders

Test Results

I tested both tokenizers on three different Arabic texts with Encoder only for now more to come later!:

Key Findings

Why AraModernBert Performs Better

Practical Implications

Should You Train Your Own Tokenizer?

After this evaluation, here’s my thinking:

Use AraModernBert if:

• You’re working with Modern Standard Arabic or mixed dialects
• You want state-of-the-art efficiency out of the box
• You don’t have massive compute resources for training

Train your own if:

• You have a very specific domain (medical, legal, etc.)
• You’re working with a specific dialect extensively (pure Egyptian, Levantine, etc.)
• You have unique requirements (handling tashkeel differently, etc.)

For my Egyptian Arabic use case, I’m leaning toward using AraModernBert’s tokenizer as-is, since:

1. It already handles Egyptian dialect reasonably well
2. The 50K vocabulary is large enough to be flexible
3. Training a custom tokenizer requires significant effort and data

Next Steps

1. Test on real data: Evaluate on actual Egyptian Arabic corpus (FineWeb Egyptian)
2. Compare with SentencePiece: Test a SentencePiece tokenizer trained on Arabic
3. Measure downstream performance: Tokenizer efficiency doesn’t always equal better model performance
4. Investigate tashkeel handling: How do these tokenizers handle diacritical marks?

Conclusion

Tokenizer evaluation revealed that AraModernBert significantly outperforms AraBERT v2 in efficiency metrics, with 15-71% fewer tokens needed for the same Arabic text.

This translates to real cost savings and performance improvements in production systems.

The key lesson: don’t assume all tokenizers are equal. A few hours of evaluation can save months of headaches and significant costs down the line.

Code Repository

The complete evaluation code is available as a simple function:

def evaluate_tokenizer(text, tokenizer):
    number_of_tokens = len(tokenizer.tokenize(text))
    number_of_bytes = len(text.encode('utf-8'))
    number_of_words = len(text.split(" "))
    fertility = number_of_tokens / number_of_words 
    compression_ratio = number_of_bytes / number_of_tokens 
    return {
        "fertility": fertility,
        "compression_ratio": compression_ratio,
        "total_tokens": number_of_tokens
    }

this post was built with guide from Amazing Solveit

Internal Resources

If you found these career reflections and thoughts on the AI field useful, you might also be interested in:

• My Research Papers
• Open Source Contributions
• Main Blog Index
• Today I Learned Index

What You’ll Learn

• How PyLate and ColBERT work together for dense retrieval
• Three dataset formats (Triplet, Structured, Passage Ranking)
• Data conversion strategies for large datasets
• PLAID indexing and document processing
• IR metrics (NDCG, MAP, Recall) and evaluation
• Common pitfalls and performance optimization

Core Concepts: PyLate and ColBERT Explained

PyLate is a Python library for vector retrieval and search, specifically designed for ColBERT models. It provides tools for indexing documents, encoding queries, and performing similarity search.

ColBERT (Contextualized Late Interaction over BERT) creates dense vector representations for documents and queries, then uses late interaction (token-level matching) for retrieval instead of single vector similarity. This approach is faster and more efficient than traditional dense retrieval.

Dataset Formats

Data Conversion Strategies

with open('corpus.jsonl', 'w') as f:
    for el in dataset['corpus']:
        if el['corpus-id'] and el['text']:
            json.dump({"_id": el['corpus-id'], "text": el['text']}, f)
            f.write('\n')

documents = [{"id": doc_id, "text": text} for doc_id, text in corpus.items()]
queries = list(queries.values())
# qrels stays as dictionary

Evaluation Approaches

qrels = Qrels(qrels_dict)
run = Run(run_dict)
metrics = evaluate(qrels, run, ["ndcg@5", "map@5"])

Indexing Strategies

Common Pitfalls & Solutions

for row in dataset:
    for neg_doc in row['negative_passages']:
        # process each negative document

Performance Considerations

Evaluation Metrics

What is RAPTOR RAG? Complete Guide to Meta’s RAPTOR Technique

RAPTOR RAG is Meta’s innovative technique designed to improve document retrieval performance. While they claim a 20% improvement, my testing shows mixed results depending on the use case.

DFloat11: Efficient, Lossless Compression for LLMs

DFloat11 (DF11) is a novel, mathematically lossless compression technique that reduces large language model memory usage by about 30% with zero accuracy loss. Unlike traditional quantization methods that can degrade model quality, DF11 uses Huffman coding to compress only the predictable exponent bits of model weights.

Pyrefly: Fast Python Type Checker in Rust

Pyrefly is a blazingly fast Python type checker written in Rust, designed to provide near-instantaneous type checking for large Python codebases.

LlamaIndex RAPTOR Integration

For those using LlamaIndex, RAPTOR RAG can be integrated to improve document retrieval performance. The tree-based approach works particularly well with LlamaIndex’s document processing pipeline.

Meta Pyrefly and Open Source Ecosystem

Meta’s contribution to the Python development ecosystem through Pyrefly demonstrates their commitment to developer productivity. The combination of Rust’s performance with Python’s flexibility creates a powerful tool for modern development workflows.

Conclusion

These three tools represent significant advances in their respective domains:

• RAPTOR RAG offers a novel approach to document retrieval, though with mixed practical benefits
• DFloat11 provides genuine value for LLM deployment with lossless compression
• Pyrefly delivers substantial performance improvements for Python type checking

While RAPTOR RAG’s claims may be overstated, the underlying tree-based approach shows promise for specific use cases like document summarization. DFloat11 and Pyrefly, however, offer clear, measurable benefits that make them valuable additions to any ML or Python development toolkit.

References

1. XMad Team Blog - Meta’s research team insights
2. RAPTOR Paper - Original research paper
3. Pyrefly Official Site - Documentation and installation
4. DFloat11 Research - Compression technique details
5. LlamaIndex Documentation - RAG framework integration

What Can Brave Leo AI Do?

Leo’s got a lot of tricks up its sleeve: - Summarize Stuff Instantly: It can give you quick summaries of webpages, PDFs, Google Docs, Google Sheets, or even YouTube videos by reading their transcripts.

• Answer Questions: Whether it’s about a webpage or just something you’re curious about, Leo can explain things clearly and even offer different perspectives.

• Write and Create: Need an article, email, essay, or some code? Leo can whip it up for you.

• Translate and Code: It can translate text into different languages or help with coding by suggesting or generating code snippets.

• Custom AI Models: With the “Bring Your Own Model” feature, you can plug in your own local or remote AI models for a personalized experience.

Is Brave Leo Safe to Use?

Privacy is Leo’s middle name! Here’s why it’s safe:

• Anonymized Requests: Leo uses a reverse proxy, so Brave can’t tie your requests to your IP address.

• No Chat Storage: Your conversations aren’t saved on Brave’s servers or used to train AI models.

• No Sign-Up Needed: You can use it for free without an account. Even the premium version uses anonymous tokens to keep things private.

• Local Storage: Your chat history stays on your device, and you can clear it anytime through the browser settings.

• Heads-Up on Third-Party Models: If you use external AI models (like Anthropic’s Claude), their data policies might differ (Claude keeps chats for 30 days, for

• example). Always check the privacy terms if you go that route.

What About Your Chat History with Leo?

If you’re using Brave version 1.75 or higher on desktop or Android (not in Incognito or Tor mode), you can keep track of your chats with Leo.

They’re stored locally on your device, not on any server, so you’re in control. You can revisit, continue, or delete them from the Leo full-page view (brave://leo-ai) or the browser’s sidebar.

Just note that clearing your browsing history will also wipe out any webpage-related content in your chats. Easy peasy!

Bring Your Own Model (BYOM) with Brave Leo

With BYOM , you can connect your own AI models to Leo for a custom experience. You can use platforms like vLLM, SGLang, or any inefernce engines with any Hugging Face Transformers model, as long as it follows the OpenAI Chat Protocol.

For example, you can run a model like Qwen2.5-VL-3B-Instruct locally with this command:

python -m sglang.launch_server --port 7501 --model-path Qwen/Qwen2.5-VL-3B-Instruct

This sets up a server for SGLang (or you can use vLLM with a similar command).

Then, in Brave Leo’s BYOM settings, add your model with these details:

Label: Qwen2.5-VL-3B-Instruct

Model Request Name: Qwen2.5

Server Endpoint: http://127.0.0.1:7501/v1/chat/completions

Context Size: 4000

API Key: local

System Prompt: A custom prompt like, “You are Leo, a helpful AI assistant by Brave. Provide clear, concise, polite responses under 80 words. Use a neutral tone, clarify if needed, and ensure accuracy.”

Brave doesn’t proxy these requests, so check the privacy terms of your chosen provider. Once set up, your model integrates with Leo, letting you use it directly in the browser for tailored, private AI chats. It’s like giving Leo your own custom brain!

البداية

لدي من العمر الآن 22 عام لدي الكثير من الندم والحسرة علي العمر الفائت في غير الطاعة والعمل الصالح ووددت لو عندي فرصة في الرجوع للخلف كثيرا والبداية مجددا..لكننا لسنا في فيلم أنمي بل واقع…عزائي إنك إن تستدرك، تستدرك عند رب كريم قادر علي كل شئ. يبدل لك سيئاتك.

أعاني من الكثير من عسر الفهم في كل شئ تقريبا..خصوصا اللغة العربية لدي عسر قراءة و كتابة و الي الصف الثالث الثانوي لدي مشاكل فادحه في الكتابة..الإ أنني كنت احب الرياضيات برغم فشلي فيها ..واحب الإنجيليزي والحال أحسن من العربية بكثير ..احبهم ولكنهم لا يحبوني والحمد لله.

بدات أحاول جاهدا في الصف الثاني الثانوي محاولات تعتبر جاده الي الثالث الثانوي وكنت فاشله وتجاوزت الامتحانات بالغش ودخلت الجامعة ولا استحق هذا المكان ولا ذلك المجموع.

في كل تلك السنوات الفائته كنت تعيش في عالم المقارانات وتنتظر من الأهل و الأصحاب الإعتراف والتقدير ..برغم كل هذا فكنت أتلقي الكثير من الحب والإيمان بي من كل من درس لي تقريبا. فجزاهم الله خيرا.

من داخلي لأ افكر في هذه الأفكار كثيرا..لكني في بداية الجامعة تبدل الحال و أصبح كل وقتي تعلم ودراسة الي أن تخرجت العام الفائت و مررت بكثير من الصعاب وتجاوزتها و الحمد لله و الي الأن أقترب من اللحظة التي سيتغير فيها كل شئ ويتبدل كل الحال..من فرص السفر والمنح و نجاح الأعمال والحرية المادة و الكثير ثم قبلها يوم الحصاد أو يومها مثل أخر مره..أفقد كل هذا و أبدا من جديد..تشعر بالحزن كثيرا..لكن بعد فترة تجد رحمات ربك و تقديره في كل هذا و هناك أوقات لا تعرف ماهي الحكمة لكن تبدا في رؤية تقصيرك و تعيد حسابات وتتوجه الي الله مجددا.

التحرر من الدنس

بعد هذه السنين تنظر لنفسك و تبدا في الغوص في أعماك الذات لماذا أقوم بهذا وماهي الرغبة والهدف !

تجد أنك تم حبسك باستخدام معتقدات و أفكار سامة و فاسدة طيلة حياتك و القليل القلليل من الأفكار الصالحة من التربية و المجتمع…تأتي هذه الأفكار بعد أن تفككت الحياة وتبدلت وهناك الكثير من الموتي في حياتك وأنت مازلت

تعيش في سعيك للاعتراف بك…تعيش أسير للماضي فتظلم بذلك الحاضر والمستقبل.

الأنانية ولذة العلم

كان هناك أوقات يجب أن اختار المال أم العلم!وكنت بكل بساطة العلم ..فالمال سيأتي مستقبلا بعد العلم ..وهذا ليس صحيحا 100% ولكن الواقع مختلف..

لماذا تريد العلم!! أقوم بسرد المائات من الشعارات الجذابة عن العلم والتعلم والهمة ورفع الحرج عن الأمة والثغر وكلام يجعلك تقول ياه بارك الله فيك…

معظمه كلام ..لحظات الحقيقة هي المحن ..اللحظة التي سوف تتخلي عن عمل معين وسعي و لن يذكر اسمك ولكنه سيخرج باسم كلي وليس اسم يخصك فتشعر من داخلك ماذا لو كنت انا من قام به وحدي وتم ذكري اسمي وحدي علي منصاتي سوف أصبح معروفا في مجالي و احصل علي فرصة عمل جيدة….

لن أخبر هذا كيفية عمل النموذج هذا او أن هناك طريقة أسهل الي ان اقوم انا بها حتي يكون لي السبق!

والكثير من لحظات الانانية وحب الذات ولذة العلم تأتي علي حساب الغاية الاولي و مرضاة الله ، و حسن العلم والإخلاص في العمل والكثير من الثمرات التي أضيعها فأصبح عبدا للعبادة (طلب العلم ) ولست عبدا لله.

هل هذا هو الحق؟

الحمد لله الذي أراني بعض من أدناس القلب وفساد النية، هل هذا هو الحق! بالتاكيد لا.

ماهو الحق إذا…هناك الكثير من الإجابات لكن ما يحضر في قلبي حاليا بعد الحصول علي كثير من ما أحب من علم ومال و علاقات و غيره..أجد الإجابة القرانية هي الإجابة الوحيده الشافيه لتعبر عن ما أمر به من أزمة هوية حقيقة.

• { قُلْ هَلْ نُنَبِّئُكُمْ بِالأَخْسَرِينَ أَعْمَالا * الَّذِينَ ضَلَّ سَعْيُهُمْ فِي الْحَيَاةِ الدُّنْيَا وَهُمْ يَحْسَبُونَ أَنَّهُمْ يُحْسِنُونَ صُنْعًا } سورة الكهف.
• سورة الحديد
  أرْوَى. · سورة الحديد - البراء بصفر

وليس عند ما يقال بعد سورة الحديد ومافيها من خطاب يحرك الصخر و يشفي الصدر و يعلي اللهمة ويعطي للعبد النظارة الحقيقة لرؤية الحق.

طلب التوفيق و الصلاح

في الحقيقة كل ما في الأعلي هو تمهيد أن كل أمالي و طموحاتي هي رغبة في المنافسة و الحصول علي التقدير والرفعة في الدينا…وهذا خسران مبين.

أرغب فعلا أن تكون تتحول هذه الرغبة الي رغبة فيما عند الله و محبتة و رضاه فهذا هو التوفيق و أن ترتفع درجاتي في الدنيا بأكبر قدر قبل فوات الأوان عندما تنقلب كل أعمالي الي حسرة و قبل أن يضل سعي و أنا أحسب أني احسن العمل.

بعد كل المصاعب التي تقابلني من حاجة مادة وحاجة للفهم والوقت والتوفيق والعلم و غير هذا..أجد أن الامر كله بيد الله و قدرته.

فلا حول ولا قوة الإبالله و حسبي الله ونعم الوكيل.

يجب علي تقبل المحدودية و الضعف و اللجوء الي من بيده ملكوت السموات و الأرض القادر علي كل شئ. سبحان ربي و تعالي و صلي الله وسلم علي سيدنا محمد.

Managing WHOIS Privacy During Transfer

One critical setting often overlooked: WHOIS Privacy Protection (also called “Domain Privacy” in Vercel context).

Before transferring from Hostinger: 1. Disable Privacy Protection (WHOIS Privacy)—this is required for the transfer 2. Wait for the status to update (usually instant, sometimes 12 hours) 3. Don’t re-enable WHOIS Privacy until after the transfer completes

With Porkbun, you can re-enable Domain Privacy after the transfer is complete for privacy protection.

Why this matters: Vercel users often ask about WHOIS privacy—Vercel doesn’t manage WHOIS directly, but your registrar does. Keep this enabled for privacy unless you need your contact info public.

Moving My Domains

I decided to transfer all 11 of my domains from Hostinger to other providers. I consulted two friends—a team lead in Canada and an SEO specialist in Turkey—and both recommended Porkbun.

I registered with Porkbun and was required to verify my identity. They offered multiple verification options, and I chose to use my government ID. However, I faced issues: I uploaded my ID about 10 times, and it failed every time, even though the verification platform supported my language. I reached out to their help center, and the fastest response came via email.

It took about 12 hours to receive a reply, which felt like a long time. The good news? They instantly verified my ID with no further issues. Later, when I asked another question, their response again took several hours, unlike Hostinger, which typically responds within 1 to 30 minutes (and at most within 2 hours). However, I found Hostinger’s customer service stricter and sometimes less friendly.

I’ll discuss my reasons for moving later, but for now, let’s focus on the process.

Connecting Your Domain: Vercel, Railway, and Cloudflare Setup

Why I Switched from Hostinger to Porkbun

Why use Substack as an AI Researcher?

I used LinkedIn and Medium, and I don’t find them very useful for engagement or building a real network.

My gravity is very tiny in these networks and I can’t get bigger for multiple reasons I will discuss later.

The same is true for X, but it’s a very good place for the ML community.

My reasons:

1. Improve my writing style. I am not good at writing or English, but I am trying to do my best. Communication is a very important skill for my work as a researcher.

2. I build products I want people to learn more about…more later.

3. I have services I want to market…more later.

Did you forget something?

^_^

Die, empty, and publish your knowledge!

Actually, the main reason for me to write is that I feel lonely in my journey of learning about DL and software engineering.

And I feel bored most of the time.

I love the feeling that I am writing and archiving my life. This reminds me that I am doing hard work, learning new things, and my life will not end without doing great things from my point of view.

Recapping what I learned when I write about it makes the concepts stick in your mind, and you get more insights from people’s comments. This is very helpful for me.

For readers, you can see what I do, and you may be interested in it, be inspired, and learn from my mistakes.

Hotel Search with Vectors

Superlinked enhances search by embedding each attribute (text, numbers, categories) into specialized spaces, enabling nuanced, multi-attribute queries.

An LLM interprets user intent, assigning weights to preferences (e.g., price, rating), allowing flexible, business-driven ranking without system redesign.

Hard filters narrow results, while weighted nearest neighbor search ranks them by user preferences.

This unified approach supports multimodal search—combining semantic text, scaled numerical, and categorical data—preserving relationships and preference strengths.

Unlike traditional systems that separate or flatten data, Superlinked enables simultaneous, weighted consideration of all attributes, solving challenges like reconciling

results across types and capturing nuanced user intent.

Reciprocal Rank Rusion (RRF)

Qdrant’s native support for Reciprocal Rank Fusion (RRF) streamlined their retriever implementations, reducing hybrid search code by 80%. The multi-vector capabilities also enabled more sophisticated retrieval methods that better captured semantic relationships.

Qdrant 1.13 GPU Indexing

Here is summary of these new features

Static Embedding with Qdrant and Model2vec

Static embedding from minishLab reduce the model size with 15x reduction and up to 500x speed increase while the maintain more than 85% of the performance levels. it’s like our zaraah model for arabic.

Static embedding are dense embedding so you can also use with qdrant collections. The retrieval is not going to be any faster becuase static embeddings. but the speedup is in creating the vectors from your data and encoding the queries.

If you want to make the retrieval faster use the following: 1. Matryoshka Embeddings 2. Quantization methods like (Scalar and Binary Quantization) ### When to use Static Embeddings ?

1. Mobile applications - although many smartphones have powerful CPUs or even GPUs, the battery life is still a concern, and the static embeddings might be a good compromise between the quality and the power consumption. Moreover, the static embeddings can be used in the applications that require offline mode.

2. Web browser extensions - running a transformer-based model in a web browser is usually not quite an option, but static embeddings might be a good choice, as they have fewer parameters and are faster to encode.

3. Embedded systems - the static embeddings might be a good choice for the devices with limited computational power, such as IoT devices or microcontrollers.

create your own gravitey

sometimes you want to reach and communicate with people in the same space of problems you are solving, search engines are very bad in provide the information you want..this is related to how these engines work and other SEO stuff. but for me, i can’t reach people and help them or get benefit from them because they simple don’t know about me!!

TIL will decreases this spaces and daily TIl about the things i am learning which are alot will start to give me nice SEO and unique because i am talking about things i don’t know and i am interested in and there will be much people in the same boat this will increase my X account and linkedin and this is very useful in the current time.

TILs level up

I also want to think in way to extract more crafted blogs from my TILs. but just start making it habbit and will see who will it comes in the end.

initial thoughts: 1. weekly recap from my TILs and SEO optimization for the keywords that is increasing and i am interested in 2. ML to extract related stuff, for example i will want to take about the folloinwg topics: - Late interaciton (Pylate, Colbertv2, ColPali) - Searching - Model2vec - Visino Language models..etc collecting them and start writing about them and adding internal links for them will be very useful to improve my blog system

Things i learned the last 2 days

What is PALID index?

PLAID index is for indexing very large datasets with Later interaction models like (Colbertv2 & ColPali) IT Solves the storage footprint and allow you to scale to infinity

They swapped the faiss from facebook what is nice thing because i have multiple bad time to install it especially the GPU version.

and used fastkeamns from the amazing @bclaive which i really like his work on embeddings

It’s replacement for Voyager-based HNSW index which was very bad for scaling Late-Interaction retrieval models

Prime Intellect VS Gpuvec

I was creating a website to find and compare the cloud compute instances from all cloud providers in easy and interactive way.

I started to work on it the last month, but stoped for other work.

suddenly i found this website which is called Prime intellect.

And i just want to say,woooooow. it’s a piece of art. The design and information on it and how fast, accurate is very embrassing for my poor gpuvec.com

should i continue improve my website?

Actually yes, we share similar goals but there is multiple chances to compete or even collaborite!! who knows!

They are more than just listing and compare prices, then enable you to use these GPUs from their website.

Also they create decentralized models and have a mutliple expirened engineers.

References

1. Antoine tweet
2. betatim TIL
3. other TILs