Arezoo Samimi Abianeh

Quantization in Large Language Models (LLMs) As LLMs scale, intelligence is no longer the bottleneck — memory, latency, and cost are. Quantization turns out to be one of the most powerful (and misunderstood) techniques enabling real-world deployment of large models. In this blog, I break down: • Why LLMs are surprisingly quantization-friendly • INT8 vs INT4 vs NF4 — what actually works in practice • How modern architectures absorb low-bit noise • Where quantization fails and how to design around it Quantization is no longer an optimization step — it’s foundational infrastructure for production LLM systems. Link: https://lnkd.in/gqGje75G #LLMs #AIEngineering #Quantization #MachineLearning #GenAI #MLOps

9

2 Comments

Maximize data quality and cut annotation time by up to 10× with Argilla Argilla is an open-source data curation platform that helps AI teams build, label, and monitor high-quality datasets for NLP and LLM workflows. Used by teams like Red Cross and Prolific to accelerate labeling and feedback collection - Combines human-in-the-loop and model-assisted labeling (few-shot, active learning, semantic search), slashing manual effort by up to 10× - Easy deployment. - End-to-end MLOps support. - Open and extensible. Try it now → https://lnkd.in/eNbRhGfK #AI #MLOps #Python

1

Another good one from Karpathy’s an 1-hour LLM primer (I missed to watch earlier) covers: • How LLMs are trained (and why the weights matter) • Fine-tuning into helpful assistants • Tool use & multimodality in action • Scaling laws driving progress • Security risks: jailbreaks, prompt injection, data poisoning https://lnkd.in/ev38636z

4

NeurIPS'25 workshop paper Alert: Our paper “Zero Shot Time Series Forecasting: Do Time Series FMs Outperform Cross Modal FMs?” has been accepted  in NeurIPS'25 workshop, BERT2S 🎉. In this work, we systematically compare time series foundation models (TSFMs) such as Chronos, TimesFM, TimeGPT, and MOMENT with adapted cross-modal foundation models like GPT, Whisper, and ViT for zero-shot time series forecasting. Thanks to my co-author Md Younus Ahamed for this collaboration! 🙌

50

4 Comments

Most teams evaluating LLM agents ask the wrong question: “What’s the accuracy?” In a recent workshop, I showed a different approach using a very simple agent: ➡️ Input: product title ➡️ Output: product description No RAG. No tools. Just a prompt. We ran the agent on a small test set, collected expert feedback, and then did something most teams skip: systematic error analysis. Instead of fixing prompts based on intuition, we: • Identified recurring failure modes • Built a lightweight error taxonomy • Tagged new outputs • Used frequency to decide what to fix first The result? Clear, evidence-driven agent improvements that actually generalize. In an upcoming webinar, I’ll walk through this end-to-end using a live notebook and show how this exact method scales to agents, RAG systems, and production workflows. If you’re building with LLMs and want fewer surprises in production, this is for you. Details of upcoming webinar: https://lnkd.in/gFBFkR5h

47

Generative AI is everywhere in the hype cycle right now. But at Walmart, we’re seeing the biggest wins not from chasing the hype, but from hybrid approaches that bring multiple flavors of AI together. • For data science teams, “vibe coding” accelerates exploration and prototyping. • Machine learning still does the heavy lifting on modeling and process automation. • And then we use genBI (generative AI for business intelligence) as the user interface — turning complex model outputs into something business partners can actually use and drill into with follow-up questions. This genAI-as-UI strategy is powerful: it opens up ML insights to a much broader audience without piling extra requests on the data science team. Business partners get answers faster, and our team can stay focused on advancing the models. As this Quartz article (https://lnkd.in/eSPtKMCD) puts it, LLMs aren’t replacing traditional machine learning — they’re enhancing it. That rings true in our experience: the real value comes when each type of AI does what it does best, and we knit them together. The hype is fun, but right now the wins are hybrid.

9

2 Comments

Continuing my series on the basics of text analytics from my last post. Why am I diving into this topic? In our fast-paced world, we are surrounded by AI tools that make life easier, from ChatGPT and Copilot to Gemini, Grok, and Perplexity. Many of us use them daily for writing, research, or problem-solving, but how often do we stop to think about what powers them under the hood? Understanding the fundamentals, like regular expressions for pattern matching, lemmatization for accurate word roots, and stemming for quick simplifications, is key. These are the building blocks of text analytics, which fuel how AI processes and understands language. Without grasping them, we might miss out on using these tools more effectively, like crafting better prompts or spotting biases in outputs. This knowledge empowers you to go beyond basic usage. It helps you debug issues, build custom solutions, or even innovate in your field, whether you are in data science, marketing, or content creation. In an AI-driven era, knowing the basics turns you from a user into a smart creator. What AI tool do you rely on most, and how has learning the basics changed your approach? Let's discuss in the comments! #TextAnalytics #AI #DataScience #NLP

20

👉 “RAG isn’t one thing—it’s evolving…. into 8 powerful architectures that shape how LLMs interact with knowledge.” Here’s the lineup: 1️⃣ Naive RAG – Embed query → retrieve from vector DB → generate. 2️⃣ Multimodal RAG – Expands to text + images + other data sources. 3️⃣ HyDE (Hypothetical Document Embeddings) – Generate a pseudo-answer → embed → retrieve. 4️⃣ Corrective RAG – Cross-checks retrieval, fact-checks via web/external sources. 5️⃣ Graph RAG – Uses structured graph DBs alongside embeddings. 6️⃣ Hybrid RAG – Merges graph DB + vector DB contexts for richer context. 7️⃣ Adaptive RAG – Multi-step reasoning chains that refine queries. 8️⃣ Agentic RAG – Agents with memory, planning, and tool orchestration. Of what i am seeing in the industry, I see organizations moving step-by-step from Naive → Adaptive → Agentic RAG as their systems mature. The big shift? Moving from “retrieving documents” to “reasoning and acting” across data ecosystems. 💬 Which of these RAG architectures do you think will become the enterprise default in the next 12–18 months? #GenerativeAI #RAG #AIArchitecture #LLM #Innovation

10