Chiraag .

This evening in Seattle Union.ai and Weights & Biases are hosting an in-person event with awesome talks, food, and people! We're taking over a venue in South Lake Union and still have some room left! RSVP 👇 Anish Shah from Weights & Biases will discuss the latest in Agentic AI. Agents are messy—multi-step reasoning, tools, memory. Observability is how we keep them reliable. In this talk, we will trace how observability has evolved from classic ML to today’s agents, and share the key tenets for debugging, trust, and scale using Weights & Biases Weave. Ketan Umare, CEO and Founder of Union, will cover the latest in Agent workflows. Agents are transforming modern AI workflows, but they introduce new challenges around fault tolerance, updates, and debugging. In this session, you’ll learn how orchestration has evolved to make AI agents reliable. We’ll highlight the best practices, including automatic caching, versioned deployments, and scaling large fan-outs, to help you build resilient, production-ready agentic systems. We’ll show examples of what this looks like in Flyte 2.0. I hope to see many of you there! RSVP on Weights & Biases: https://lnkd.in/gTc5JPWN

4

2 Comments

🚀 Big news from Intel Labs & Weizmann Institute at ICML 2025! They’ve introduced a breakthrough in speculative decoding that lets any small draft model accelerate any large LLM — regardless of vocabularies or ecosystems. The result? Up to 2.8× faster inference, vendor-agnostic, and already integrated into Hugging Face Transformers. This is a huge step toward faster, cheaper, and more interoperable AI. Excited about how we can leverage this in production to speed up inference, reduce serving costs, and scale flexibly from cloud to edge. Definitely something to watch closely. 🔍 What is Speculative Decoding? Large Language Models (LLMs) normally generate text token by token in an autoregressive way. This is accurate but slow. Speculative decoding accelerates this by pairing: ✅ Draft model – a smaller, faster model that proposes several next tokens at once. ✅ Target model – the large accurate model that verifies or corrects those tokens. Workflow: Drafting – The small model “guesses ahead” multiple tokens. Verification – The large model checks those guesses in fewer forward passes. Acceptance/Rejection – If correct, the draft tokens are accepted; if not, they’re replaced. This reduces the number of computations per token while preserving accuracy. ⚡ Benefits 🚀 2–4× faster inference (recent Intel–Weizmann, Snowflake, vLLM updates). 💸 Lower compute costs for serving models in production. 🔄 Works across ecosystems (now vendor-agnostic thanks to ICML 2025 Intel paper). 📱 Enables deployment from cloud to memory-constrained edge devices. #AI #LLM #SpeculativeDecoding #MachineLearning #HuggingFace #Intel

6

🚀Excited to be part of the team at Nand.ai as we launch Magus — an AI system built to help enterprise sales teams handle complex workflows like RFPs, security questionnaires, and technical proposals. We’ve focused on building systems that can process and organize large volumes of dense content, reliably and at scale. Curious how it works? Happy to connect for a quick walkthrough. #AI #EnterpriseSoftware #RFP #NandAI #Launch

23

Understanding of Generative AI, LLMs & Real-World Use Cases Gained a strong foundation in how Generative AI and Large Language Models (LLMs) work, including prompt–response structure, model capabilities, limitations, and ethical considerations. Explored real-world applications such as customer support automation, content generation, code assistants, QA automation, business intelligence, and task optimization. ✅ Advanced Prompt Engineering Techniques — (Elaborated) Developed hands-on expertise in crafting high-quality prompts for accurate and context-aware outputs. Mastered techniques like: Role prompting Chain-of-thought prompting Multi-step workflow prompts Data extraction prompts Evaluation and refinement loops Learned how to structure prompts for precision, creativity, automation, and decision-making tasks. --- ✅ AI-Powered Research, Content Generation & Data Analysis — (Elaborated) Learned to use AI tools to: Summarize long documents & datasets Generate insights from data Create structured reports Draft test scenarios, emails, and technical documentation Conduct competitive analysis & research efficiently This boosted my ability to quickly gather information, reduce manual effort, and produce high-quality content. --- ✅ Using AI for Problem-Solving & Decision-Making — (Elaborated) Gained experience using AI as a thought partner to break down complex problems into actionable steps. Applied prompt-based frameworks for: Root cause analysis Bug impact assessment Scenario evaluation Risk-based prioritization Decision-making with pros/cons and data-supported reasoning This helps in taking more informed and efficient decisions in professional tasks. --- ✅ Generative AI for Automation: Improving Productivity & Accuracy — (Elaborated) Learned how to automate repetitive daily tasks using AI, including: Test case drafting Requirement understanding Ticket analysis Document formatting Email generation Data cleanup This allows faster output with fewer human errors and improved consistency in tasks. --- ✅ Creating Workflows that Combine AI + QA for Smarter Testing — (Elaborated) Gained practical understanding of integrating AI into QA workflows, such as: Auto-generating test scenarios from user stories AI-assisted bug replication Faster documentation validation AI-powered regression support AI-guided exploratory testing suggestions This improves test coverage, reduces testing time, and enhances the QA professional’s productivity. --- ✅ Enhancing Reporting & Documentation — (Elaborated) Improved skills in developing polished, clear, and structured documentation using AI, including: Test reports Sprint summaries Defect logs ---------------------------- Happy weekend✨ "Grateful for the learning" Thanks upGrad ✨✅ #learning #GenerativeAI #AI #microsoft

21

2 Comments

Day 13/30 - Building a Streaming Storytelling Agent with OpenAI Agents SDK Today in my 30-day AI Learning Journey on my own platform, BCA PATHSHALA by Websyro Academy, I explored a live interaction concept that transforms how users experience AI responses: Streaming Storytelling Agent using the OpenAI Agents SDK Unlike regular agents that send a final response after thinking, a streaming agent sends partial results as it thinks, making the user experience feel real-time, dynamic, and interactive. What is a Streaming Storytelling Agent? This is an agent designed to tell long-form stories in real time by streaming content back to the user while it’s being generated, like watching a story unfold live. Instead of waiting for the final answer, users see the narrative build piece by piece. This is great for: ✔ Interactive storytelling apps ✔ Live chat experiences ✔ Voice assistants & gaming interactions Thanks to the OpenAI Agents SDK, you don’t need complex custom implementations to achieve this. Learn more and follow the full hands-on guide here: https://lnkd.in/gkWiwDnF 🚀 Day 14 coming next! #AI #OpenAIAgentsSDK #StreamingAI #InteractiveAI #StorytellingAgent #GenAI #WebsyroAcademy #BCAPatshala #LLM #AgenticAI #30DaysOfAI #LearningJourney #RealTimeAI

18

Just read an interesting article, "In the AI era, Wikipedia has never been more valuable" and it immediately reminded me of something Andrej Karpathy said on Dwarkesh Patel's podcast. He made a striking point: a huge amount of internet data is essentially garbage. As we chase bigger datasets for AI training, we're often just adding noise. The counterintuitive insight? The path to better AI might actually be through smaller models trained on high-quality, human-curated sources—and Wikipedia is the perfect example. What makes Wikipedia special:  • Every article is collaboratively edited and verified by humans  • Information is structured, not just scraped  • Quality control happens through community consensus  • It's comprehensive without being cluttered with SEO spam, clickbait, or low-effort content We've been so focused on scale that we've undervalued curation. But training on meticulously maintained knowledge bases like Wikipedia could give us more accurate, efficient models than training on terabytes of unfiltered web content. Twenty years ago, people questioned if Wikipedia could work. Today, it might be one of our most valuable resources for building better AI. Sometimes the best path forward isn't more—it's better. #ArtificialIntelligence #MachineLearning#DataQuality #Wikipedia

12

#MCP is an open protocol that abstracts the complexity of tool integration for LLM-based applications. It establishes a standardized communication layer, eliminating the need for bespoke code when connecting to external data sources. By conforming to this open standard, developers can seamlessly provide context to an LLM by enabling programmatic access to diverse tools, thereby simplifying how they provide context to their LLM-based applications. For example, you can give context via third-party tools that let your LLM make API calls to search the web, access data from local docs, retrieve code from a GitHub repo, and so on. #MCP #RichContext #AI #Anthropic #LLM #GENAI

5

I do not really post a lot about tech stuff, but maybe I can start. Here is a paper I read recently. ROSE: Robust caches for Amazon product search [https://lnkd.in/gPhbgHHQ] Thanks to Arpit Bhayani's paper shelf recommendations. The insights from this are really fascinating for the search world and how Amazon always knows how to find what you are looking for even if you have clumsy fingers like mine. For example if you are looking for shoes, you might type something like: “nike shoe,” “nike shoes,” “nikes shooes.” A normal cache treats these as different, so it gets big and slow, and you miss the cache a lot. Amazon’s ROSE fixes this by caching the intent, not the exact spelling. What ROSE essentially is, is a typo/variant-tolerant cache for product search that groups similar queries into the same “bucket,” so most lookups hit fast without growing memory. It’s deployed in Amazon Search and improved both latency (single-digit ms for most traffic) and business metrics. How this works is even more fascinating. 1. Locality-Sensitive Hashing: Similar queries collide in the same bucket. • Lexical-preserving hashing (character n-grams + minhash) for typos/variants. • Product-type-preserving hashing (weighted minhash) to keep the category intent (e.g., “dishwasher” vs. “dishwasher parts”). 2. Reservoir sampling: caps each bucket so memory stays constant even as queries grow. (Super convenient, and one of the many engineering gotcha's people might miss) 3. Count-based k-selection: avoids pairwise similarity inside buckets; just count collisions across hash tables → near constant-time retrieval. Where we can apply some of these concepts with Gen-AI (which is in many flavors essentially a search problem with extra steps)- Semantic prompt cache: Group paraphrased prompts to reuse an answer/logits/KV cache even when text isn’t identical. Can be helpful in saving your tokens and $. RAG request cache: map similar user questions to one retrieval plan to cut vector searches + reranking cost. [I remember Azure AI search having a RAG cache, but I am not sure how that worked] Tool/agent call cache: Dedupe near-duplicate function/tool invocations (e.g., same API call phrased differently). Eval & feedback loops: Bucket similar generations or errors to reuse critiques/patches. Support/search front-ends to LLMs: Typo-tolerant, intent-stable pre-rewrites before hitting the model, ensure we only use tokens when absolutely needed.

174

6 Comments

I finished reading the paper on MemGPT. I really like the concept of treating the context window as RAM. Many analogies are drawn from operating system concepts such as virtual memory, paging, and I/O control. This year, Andrej Karpathy described LLMs as operating systems with different sets of components. The paper explains how we can create an infinite context window by managing context effectively in real time giving the illusion of infinite context. The concept is inspired by virtual memory. Just like how an OS uses external storage such as a disk to manage multiple processes through paging, MemGPT uses similar principles for managing context. The more accurate the context, the higher the quality of responses. Context engineering has become far more important today. https://lnkd.in/dqjUb3df

16

AI as ghosts Andrej Karpathy' succinctly distills the major takeaways from the AI space in 2025. some important notes from his article below 1. LLMs are "spirits/ghosts" that you invoke, not animals that grow/evolve 2. LLMs have jagged intelligence - they can be absolutely dumb in some areas and wonderfully brilliant in others 3. "Cursor for X use case" has become a new kind of startup pitch, where you create your own LLM app with GUI and specific prompts/fine tuning for the use case. it's almost like the era of websites and apps all over again there's lot more interesting stuff, check it out! https://lnkd.in/gq4CHD6T #karpathy #llmin2025

3

I was reading a very interesting paper on data poisoning by Anthropic and thought of jotting down the key takeaways in a short blog. It’s surprising how a handful of poisoned samples can impact large language models. Read it here: (Blog link)[https://lnkd.in/gd9uKNsP]

6

Day 22/60: Measuring the Infinite (Find Length of Loop) Targeted by: Amazon, Microsoft, Goldman Sachs, Qualcomm. Yesterday, I built the Linked List. Today, I had to break it or rather, deal with a list that’s broken into a cycle. The challenge: Find the exact number of nodes inside a loop. It’s one thing to know if a loop exists, but calculating its length requires a shift from simple detection to precise tracking. 1. The Scout and the Sprinter (Floyd’s Cycle-Finding) To find a loop, we use the classic Slow and Fast Pointer approach (The Tortoise and the Hare). Slow moves 1 step. Fast moves 2 steps. If they meet at the same node, a loop is confirmed. But once they meet, how do we know how big the "track" is? 2. Freezing the Moment Once the pointers meet, we stop the "Fast" pointer in its tracks. It acts as a stationary landmark. We then take the "Slow" pointer and let it continue its journey, one step at a time, counting each node it visits until it makes a full lap and hits the stationary "Fast" pointer again. 3. The Math of the Lap The logic is foolproof: Meet: Find any point inside the cycle. Count: Start a counter at 1. Traverse: Move the temp pointer to temp.next. Repeat: Keep moving and incrementing until temp == meetingNode. The final count is the exact number of nodes trapped in the cycle. 4. Why This is Elegant Space Complexity: O(1). We don't need a Hash Set to remember visited nodes. We use the structure of the list itself. Time Complexity: O(N). We traverse the list at most twice. This pattern is vital for Deadlock Detection in operating systems. When resources are pointing to each other in a cycle, the system needs to know the "size" of the deadlock to resolve it efficiently. 5. Final Reflection The "Length of Loop" problem is a reminder that even when you feel like you're going in circles, you can still measure your progress. You just need to find a fixed point to reference. Day 22/60 done. ✅ When life puts you in a loop, stop running for a second. Find your landmark, count the steps, and master the cycle. #60DaysOfCode 👨‍💻 #DataStructures #LinkedList #CodingLife #DSA Anchal Sharma Ikshit ..

6

Excited to share my review of research on Schelling coordination in LLMs! During the Apart Research Fellowship, I conducted a review of existing work exploring whether language models can coordinate on focal points without communication. This is an important question for multi-agent AI systems. Put simply, Schelling coordination is about finding common ground without talking. Classic example: if you need to meet someone in NYC tomorrow but can't communicate, where do you go? Many people converge on "Grand Central Station at noon" because it's a natural focal point. The question is whether LLMs can do this too, and what the implications are for deploying multiple AI agents that might coordinate in unexpected ways. I hope this blog post will be useful for others working on multi-agent AI alignment and coordination problems. Read the full post here: https://lnkd.in/edvcTMt5 I’m grateful to Apart Research for the opportunity and my supervisor Jacob Haimes for his mentorship.

31

1 Comment

Day 51 of #100DaysOfGenAI Today’s session with Siddhant Goswami was a deep dive into the nuanced world of LLM Workflows vs. Agents, and it left me with many thoughts on where the future of GenAI is headed. 🧩 We explored the advantages and limitations of both paradigms: LLM Workflows are structured, predictable, and easier to debug Agents bring reasoning, flexibility, and adaptability, but also added complexity And it's not a black-or-white switch. It’s more like a spectrum — 🌀 On a scale of 1 to 10, how "agentic" is your system? 💡 We also brainstormed some interesting ideas for implementation, including things like: A "Truth Terminal" Twitter bot Using advanced tools offered by LLM providers like OpenAI: 🔍 WebSearch 📂 File Store 🔗 Remote MCP 💻 Computer Use 🧮 Code Interpreter I’ve used some of these before — especially for RAG apps and MCP-style tools on Claude — but the Computer Use tool caught my attention. Seeing how OpenAI integrates all these tools into its new Agent architecture is exciting — the dots are starting to connect. 🌐 The possibilities feel much broader than just RAG — and way more dynamic. Can’t wait to experiment with some of these new ideas in the coming days! #0to100xEngineers

6

Is the golden era of pure AI research at Meta officially ending? The departure of VP Jitendra Malik to Amazon Robotics signals a massive shift in the AI landscape. For years, FAIR (Fundamental AI Research) was the gold standard for academic freedom within Big Tech. But as the focus shifts aggressively toward commercial products, the culture is changing. Here is what this exodus tells us about the current state of AI: 📍 The Pivot is Real Meta is moving from "what is possible?" to "what is profitable?" The aggressive push for commercial AI products is reportedly leaving pure researchers feeling sidelined. 📍 The Talent War is Shifting Malik isn't just leaving; he is joining Amazon’s robotics efforts. This suggests the next frontier isn't just LLMs, but embodied AI and physical automation. 📍 The "Brain Drain" Effect With reports citing former employees describing the division as "dying a slow death" and following other high-profile exits, retaining top-tier research talent requires more than just high salaries—it requires a mission they believe in. Think of this like a university lab being turned into a factory floor. While the factory produces immediate value, the explorers who thrive on discovery often pack their bags when the goal becomes pure efficiency. The lesson for tech leaders and developers? We are entering the era of Applied AI. The companies that win won't just be the ones doing the best research; they will be the ones who can successfully transition that research into tangible products without alienating the minds that built it. If you are in AI, now is the time to bridge the gap between theory and deployment. Do you think Big Tech can still sustain pure research labs, or is everything destined for productization? Let me know in the comments. 👇 #AI #TechNews #Robotics #Leadership #Meta #Amazon

3

Why LLMs May Struggle in India Large Language Models are built on a simple assumption: the user will ask a question. But in India, asking questions doesn’t always come naturally. As children, we’re endlessly curious. Yet as we grow, we’re taught to stay quiet—don’t challenge authority, don’t upset the status quo, don’t risk offending someone. Over time, we stop asking even the most basic questions: Why are there potholes on the road? Where is my tax money going? In many settings, asking feels uncomfortable—or even unsafe. This cultural habit creates a barrier. No matter how affordable AI tools become, adoption will be limited if the experience depends on people openly questioning. For LLMs to truly work here, they may need to evolve from being passive “answer machines” into proactive companions that guide, prompt, and anticipate needs—so people don’t have to break through cultural conditioning just to use them. What do you think—can AI overcome cultural barriers, or will culture always shape how we use technology? PS: Yeah I did use an LLM to polish this up 😅

31

1 Comment

Really resonate with this perspective. The future of trustworthy AI systems won’t just be defined by smarter models, but by how observable and accountable we make them.

14