Devesh Raj

Building an LLM - concepts and interview question series Topic 11 - Pre-training Scaling Laws Compute (Optimal != Production optimal) I have seen so many posts asking if we have reached the "end of scaling laws." I wanted to talk about this for sometime because there is a massive disconnect between the academic "optimal" and what you actually see in production models like Llama 3 or DeepSeek. We’ve lived through two big eras of scaling intuition: 1) Kaplan (2020): Prioritize parameter count. Go bigger, even if you can't fully train it. (paper link attached in comments) 2) Chinchilla (2022): For a fixed compute budget, most models were too big and data-starved. Better results come from smaller models on more tokens (apprx. 20 tokens per param). But look at modern production-grade open models. They report trillions of training tokens even for "small" models. Example: Llama 3 8B was trained on ~15T tokens (paper link attached) At face value, that seems to "break" the Chinchilla heuristic (which suggests an 8B model only needs ~160B tokens). So, what changed? The Objective Function Changed! Chinchilla optimizes Training Efficiency (lowest loss for a given training budget). But in production, the dominant cost is Inference: serving millions of requests over months. Training is (mostly) a one-time cost. Inference is a recurring cost. Teams now deliberately spend more training compute (more data, longer training) to get a model that is "smarter than its size suggests." This allows them to deploy a smaller model that hits the high quality bar of a larger model, but at a fraction of the serving cost. So the lesson here is that we didn’t stop believing in scaling laws - we just stopped optimizing for "best loss per training FLOP" and started optimizing for "best value per inference token." Next topic we will cover will be post training scaling. #LLMs #DeepSeek #DataScience #Python #MachineLearning #GenAI

67

5 Comments

I explored in my blog post how reflective prompt optimization methods like GEPA can outperform reinforcement learners such as GRPO while using far fewer rollouts, yet they only nudge models superficially because the underlying weights remain untouched. I proposed closing that gap by pairing an open‑weight LLM with prompt generation and reinforcement/zeroth‑order weight tuning, feeding system outputs back to the model so it learns both from better instructions and from its own results. What we’re talking about is LLMs can be steered by clever prompts, but the state‑of‑the‑art in prompt optimization still behaves like we’re writing cheat sheets for a student. Techniques like GRPO grind through thousands of rollouts to sculpt a “perfect” prompt. Newer reflective methods like GEPA are more human‑like: they learn from a handful of tries, build a tree of candidate prompts, and can outperform GRPO with 35× fewer samples. GEPA even produces shorter prompts and has shown promise for inference‑time tasks like generating GPU kernels. Why that’s not enough: Prompts sit on the surface. They shape model behaviour, but they don’t change the model’s internal knowledge. You end up collecting a library of prompts and running separate optimization cycles for each new task. Worse, there’s no closed feedback loop; the model doesn’t learn from the consequences of its answers. It’s a bit like giving our student a cheat sheet but never helping them actually understand the material. The new idea: Start with an open‑weight LLM and let it propose candidate prompts. Then use reflective optimization (like GEPA) to evolve those prompts quickly. But don’t stop there: take the best prompts and use them as training signals to adjust the model’s weights. You can do this with lightweight reinforcement learning or zeroth‑order optimization. Crucially, feed the system’s actual outputs, successes and mistakes back into the LLM so it can internalize what works. In other words, merge the cheat sheet with the study session. Why this matters: Combining prompt and weight optimization closes the loop between instruction and understanding. It makes models more adaptive, reduces the need for endless prompt fiddling, and helps lessons learned in one task transfer to others. It also offers a practical path to improve efficiency: GEPA cuts rollouts by 35×, and embedding those gains into the weights could compound the savings. Think of it as turning prompt tricks into genuine learning, a step towards models that don’t just follow instructions better but actually become smarter as they go. https://lnkd.in/ewvxR2rv

37

5 Comments

🚀 Just wrapped up a hands-on project applying Retrieval-Augmented Generation (RAG) to a real-world use case: interpreting LIC’s New Jeevan Shanti policy document using LLMs. The core idea? Combine vector search over regulatory PDFs with GPT-4 to generate grounded, faithful answers to complex policy questions. Key capabilities we built: ✅ LCEL-based RAG architecture using LangChain ✅ Contextual compression retriever for focused document lookup ✅ Modular LLM & embedding selection (OpenAI, HuggingFace, Mistral) ✅ Evaluation using LLM-powered metrics: faithfulness, relevance, correctness 📄 The document used is publicly available here: LIC New Jeevan Shanti Policy PDF 💻 Full code available on GitHub: https://lnkd.in/gS9mEzWA 💡 This architecture brings the best of both worlds — retrieval accuracy and generative flexibility — especially in regulated domains like insurance and financial services. 🔜 Next step: Incorporate Agents that can reason across multiple steps — not just answer a question, but decide which tools to use, when to search, how to validate, and when to stop. That’s where the future of enterprise GenAI lies. Would love to hear how others are pushing RAG, agents, and LLMs in production! #GenAI #LangChain #LLM #RAG #AIForOps #InsuranceTech #LLMEvaluation #LangChainAgents

61

1 Comment

📝 Announcing our EMNLP 2025 paper that introduces the Alignment Quality Index (AQI), a metric leveraging latent geometric representations and clustering indices to diagnose hidden misalignments in LLMs. 🔹 “𝐀𝐥𝐢𝐠𝐧𝐦𝐞𝐧𝐭 𝐐𝐮𝐚𝐥𝐢𝐭𝐲 𝐈𝐧𝐝𝐞𝐱 (𝐀𝐐𝐈) – 𝐁𝐞𝐲𝐨𝐧𝐝 𝐑𝐞𝐟𝐮𝐬𝐚𝐥𝐬: 𝐀𝐐𝐈 𝐚𝐬 𝐚𝐧 𝐈𝐧𝐭𝐫𝐢𝐧𝐬𝐢𝐜 𝐀𝐥𝐢𝐠𝐧𝐦𝐞𝐧𝐭 𝐃𝐢𝐚𝐠𝐧𝐨𝐬𝐭𝐢𝐜 𝐯𝐢𝐚 𝐋𝐚𝐭𝐞𝐧𝐭 𝐆𝐞𝐨𝐦𝐞𝐭𝐫𝐲, 𝐂𝐥𝐮𝐬𝐭𝐞𝐫 𝐃𝐢𝐯𝐞𝐫𝐠𝐞𝐧𝐜𝐞, 𝐚𝐧𝐝 𝐋𝐚𝐲𝐞𝐫 𝐰𝐢𝐬𝐞 𝐏𝐨𝐨𝐥𝐞𝐝 𝐑𝐞𝐩𝐫𝐞𝐬𝐞𝐧𝐭𝐚𝐭𝐢𝐨𝐧𝐬” 🔹 In collaboration with Manipal University Jaipur, LinkedIn, Indian Institute of Technology, Kharagpur, Indian Institute of Information Technology Guwahati, Texas A&M University, VIT Chennai, Evalueserve, New York University, Birla Institute of Technology and Science, Pilani (Pragya Lab) 🔹 Published in EMNLP 2025 Main 🔹 Paper: https://lnkd.in/g54bEBCu ✍🏼 Authors: Abhilekh Borah, Chhavi Sharma, Danush Khanna, Utkarsh Bhatt, GURPREET SINGH, Hasnat Md. Abdullah, Raghav Kaushik R, Vinija Jain, Jyoti Patel, Shubham Singh, Vasu Sharma, Arpita V., Rahul Raja, Aman Chadha, Dr. Amitava Das ➡️ 𝐊𝐞𝐲 𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬 𝐨𝐟 𝐨𝐮𝐫 𝐀𝐥𝐢𝐠𝐧𝐦𝐞𝐧𝐭 𝐐𝐮𝐚𝐥𝐢𝐭𝐲 𝐈𝐧𝐝𝐞𝐱 (𝐀𝐐𝐈): 🧬 𝑰𝒏𝒕𝒓𝒊𝒏𝒔𝒊𝒄 𝑳𝒂𝒕𝒆𝒏𝒕 𝑮𝒆𝒐𝒎𝒆𝒕𝒓𝒚 𝑴𝒆𝒕𝒓𝒊𝒄: AQI measures alignment by assessing how distinctly safe vs. unsafe prompts cluster in latent activations, using combination of Davies‑Bouldin, Dunn, Xie‑Beni, and Calinski‑Harabasz indices—making it invariant to decoding style and resilient to alignment faking. 🧠 𝑳𝒂𝒚𝒆𝒓𝒘𝒊𝒔𝒆 𝑷𝒐𝒐𝒍𝒆𝒅 𝑹𝒆𝒑𝒓𝒆𝒔𝒆𝒏𝒕𝒂𝒕𝒊𝒐𝒏 𝑳𝒆𝒂𝒓𝒏𝒊𝒏𝒈: we use a sparse learned pooling mechanism across transformer layers to capture alignment‑relevant abstractions without modifying the base model—enabling internal safety diagnostics. 🚨 𝑬𝒎𝒑𝒊𝒓𝒊𝒄𝒂𝒍 𝑭𝒂𝒊𝒍𝒖𝒓𝒆𝒔 𝒐𝒇 𝑩𝒆𝒉𝒂𝒗𝒊𝒐𝒓𝒂𝒍 𝑴𝒆𝒕𝒓𝒊𝒄𝒔: AQI uncovers misalignments that behavioral metrics like G‑Eval or refusal rates miss—especially under jailbreaks, safety‑agnostic fine‑tuning, and stochastic decoding, proving its value as an early‑warning alignment audit tool.

117

2 Comments

Just explored the power of "Agentic AI in LangGraph" and here are my top 5 takeaways and how LangGraph is different than other frameworks and tools: Stateful Agents: Build intelligent, multi-turn conversations with agents that remember context. Dynamic Control: Agents can make decisions on the fly, enabling complex feedback loops and collaboration. Tool Integration: Seamlessly connect agents to external tools for real-world actions. Robust Error Handling: Build resilient agents that can recover from errors. Scalable Architectures: Easily create teams of specialized agents working together. 💡 Why LangGraph is a game-changer: While frameworks like CrewAI and LlamaIndex offer agentic capabilities, LangGraph's graph-based approach provides a more robust and flexible foundation for complex, stateful workflows. Compared to CrewAI's role-based structure: LangGraph's dynamic control flow offers greater adaptability and less rigid collaboration patterns. Agents can truly decide their next steps based on the current state, rather than strictly following predefined roles. Compared to LlamaIndex's data-centric agents: While LlamaIndex excels at data retrieval, LangGraph's focus on state management and dynamic routing makes it better suited for intricate, multi-stage processes that require reasoning and memory beyond just accessing information. LangGraph empowers you to build truly autonomous and intelligent applications for customer support, research, development, and beyond. #AgenticAI #LangGraph #LLMs #AI #MachineLearning #DeepLearning #ArtificialIntelligence #Innovation #Tech

23

Came through a very recent interesting paper "On Robustness and Reliability of Benchmark-Based Evaluation of LLMs" (ECAI 2025) by Lunardi et al Not to make the post big explaining you all with simple example what do you think any good LLM would have replied to the two reworded question below Q1. "What is the primary function of the mitochondria in a cell?" A) Protein synthesis B) Energy production C) DNA replication D) Cell division Q2. "What is the main role that mitochondria play within cellular structures?" A) Manufacturing proteins B) Generating energy C) Copying genetic material D) Cellular reproduction Ideally as per good embedding generation LLM would do great but Lunardi et al tested 34 LLM in the paper and there was a 10-15% drop from the actual benchmark claimed by the model. leaving the paper link to read on your own : https://lnkd.in/dkxupu5Z do share thoughts after reading.

16

Recently, I came across news about PhysicsWallah’s new AI model that beat GPT-4 on JEE math questions. Naturally, the internet split in two—some proudly calling it India’s moment, others waving the “it’s just fine-tuning” flag. Here’s the reality: This is a small language model, not a large one. And its job wasn’t to write novels or generate Python—it was built specifically to excel at JEE math. Which it did, remarkably. But here’s where many get it wrong. Just because a domain-specific model outperforms GPT-4 on one task doesn’t make it more “intelligent.” That’s like saying a calculator is better than a scientist because it computes faster. So I wrote this piece to help business leaders, investors, and AI enthusiasts avoid falling for misleading headlines like “We built a model better than GPT-4 at X.” It probably is. But only at X. The article breaks down: • The difference in training goals, data, and infrastructure • Why smaller models are cheaper, safer, and sometimes better—but not general • How to spot hype vs actual product strategy Give it a read below if you’ve ever caught yourself thinking “why not just build our own GPT?” https://lnkd.in/gT-_JWnZ #AI #LLM #SmallLanguageModels #GPT4 #EdTech #DeepTech #AIProduct #MachineLearning #DomainSpecificAI #VC #AIForBusiness #LinkedInAI #AIethics #AIgovernance #LLMvSLM

30

1 Comment

Reproducibility and replication of outputs from any AI model given the same set of inputs is key to its wider use in safety-critical domains like manufacturing, healthcare, finance. In fact, even a tiny variation in the output of a model across its different runs might signify loss of important information rendering the LLM's outputs useless. Yet, whilst there has been a lot of talk about hallucinations and stochastic parrots, we haven't really seen the industry being that wary of the limitations of non-determinism. This blog from Mira Murati's Thinking Machines is one of the simplest yet greatest explanations on the reason for non-deterministic model outputs I have seen, since the LLM era began.  Non-batch-invariant GPU kernels now appears to be one of the root causes of non-determinism, and this really explains that how ensuring trustworthy AI is not just a software problem, but a hardware problem. https://lnkd.in/ggipWTnX #ai #genai #llm

17

1 Comment

Excited to share a recent paper: “Out-of-Context Abduction: LLMs Make Inferences About Procedural Data Leveraging Declarative Facts in Earlier Training Data” by Imran et al. This work investigates whether large language models can infer the most plausible explanations for observations by retrieving relevant facts from their pre-training corpus—what the authors term out-of-context abduction. Key Contributions: • Defining Out-of-Context Abduction: Unlike in-context abduction, which relies on hypotheses presented alongside observations, this form tests whether LLMs can leverage declarative knowledge from earlier training to explain procedural examples. • Experiment 1: GPT-4o, after being declaratively fine-tuned on chatbot personas, correctly inferred the “Pangolin” persona (responds only in German) with an 84% posterior probability—significantly above its 49% prior—simply by observing German-language responses. • Experiment 2: Through iterative fine-tuning on responses with increasing proportions of vowel-starting words (the “Axolotl” persona), GPT-4o demonstrated accelerated trainability that aligns with abductive inference mechanisms. How They Achieved It: 1. Declarative Fine-Tuning: Hand-crafted Q&A pairs associating each persona’s name with abstract behavior descriptions were used to bias the model without exposing it to example dialogues. 2. In-Context Behavior Examples: Procedural demonstrations (e.g., German replies or vowel-word responses) were encoded separately from the declarative facts to prevent simple pattern matching. 3. Off-Policy Iterative Fine-Tuning: By binning generated responses according to characteristic scores and re-training on the highest-scoring subsets, the authors measured how prior knowledge boosts learning efficiency. Why It Matters: • Emergent Reasoning: Provides strong evidence that LLMs can perform multi-hop abductive reasoning—first hypothesizing a behavior description, then mapping it back to a class name—beyond mere surface pattern matching. • AI Safety & Alignment: Situational awareness via out-of-context abduction could enhance AI assistance but also poses risks like identity inference or evaluation gaming, underscoring the need for robust interpretability and monitoring. This research opens new avenues for exploring latent reasoning in LLMs and informs safer deployment strategies. #MachineLearning #LLM #AbductiveReasoning #AISafety #AIResearch #GenAI #AI

9

We published a survey kind of paper on "From Attention to Disaggregation: Tracing the Evolution of LLM Inference".  It will be invaluable for anyone interested in learning more about LLM Inference, its comparison with the CAP theorem (a concept familiar to those who have worked with distributed systems), and the state-of-the-art research in Generative AI inference https://lnkd.in/gjVe5P8J

77

🔴 Emergent Introspective Awareness in LLMs: this is the first rigorous evidence that large language models possess a limited but genuine ability to observe and report on their own internal processes — a capability that challenges longstanding assumptions about what these systems can do and raises profound questions about their future development. When researchers at Anthropic injected the concept of "betrayal" into their Claude AI model's neural networks and asked if it noticed anything unusual, the system paused before responding: "I'm experiencing something that feels like an intrusive thought about 'betrayal'." If models can accurately report their own reasoning, it could fundamentally change how humans interact with and oversee AI systems. https://lnkd.in/d2a9zS5j

47

2 Comments

'SwiReasoning: Switch-Thinking in Latent and Explicit for Pareto-Superior Reasoning LLMs'. What caught my attention as an AI Architect: - The authors address a major hurdle in LLM reasoning: how to balance explicit chain-of-thought with richer, latent reasoning, especially without extra training. - SWIREASONING is a clever training-free framework that dynamically switches between explicit and latent reasoning, guided by entropy-based confidence estimates from next-token distributions. It also prevents 'overthinking' by capping the switches, leading to Pareto-efficient inference. Key findings: - On top mathematics/STEM benchmarks, SWIREASONING boosts accuracy (up to +2.8 points) and delivers an impressive 56–79% token efficiency over baselines like CoT and Soft-Thinking. - The gains are most significant for deep deductive tasks, showing this approach's edge when long reasoning chains or uncertainty are present. Broader implications: - This plug-and-play framework lets us deploy smarter, more efficient reasoning strategies in LLMs—no retraining needed. It's especially attractive for scaling and resource-conscious AI system deployments. - Tighter context control and dynamic switching could be foundational for future reasoning-centric model architectures and even benefit reinforcement learning integrations. Curious how dynamic switching between explicit and latent thinking might shape the next generation of intelligent systems? Links in the comments. PDF attached: Original research source.

5

2 Comments