Litan Kumar Mohanta

🚨 Meta joins Google and Microsoft in issuing an ultimatum to employees: use AI tools, or risk being left behind 🚨 📊 Meta’s internal dashboards now track adoption across divisions, and its “Level Up” program with the Metamate chatbot rewards employees with badges as they hit usage milestones. Reality Labs alone jumped from just 30% AI adoption in June to 70% today — with a new target of 75% by year-end. 💻 They aren’t alone. Microsoft’s Julia Liuson has directed managers to treat AI use in performance reviews as essential as collaboration. Google says over 30% of its code is now AI-generated. Mark Zuckerberg has gone further, projecting that by late 2025 Meta’s AI will perform at the level of a mid-level engineer, with coding agents handling major R&D work by mid-2026. The message across Big Tech is clear: AI use is no longer optional. ⚡ But here’s the friction. Adoption isn’t the same as habit. Dashboards and mandates can drive usage spikes, but if AI isn’t woven into daily rituals of work, with role-specific value and leadership modeling, it stalls. True adoption comes when people use these tools naturally because they help, not because a dashboard tells them to. ❓ So the question is: will mandates and gamification be enough, or will companies need to focus on building behavior, change, new habit formations that actually stick? #AI #LeadershipInTheLoop #DigitalLabor #StudioCX #AIAdoption #MarkZuckerberg

3

🧠🆔 What if your recommender's memory of you were written in plain language? What if your item IDs carried meaning? Both are in production at DoorDash. #Recsperts Episode No. 34 on Generative #RecSys for Quick Commerce with Raghav Saboo from DoorDash is now live. 🎉 Raghav is a Staff Machine Learning Engineer at DoorDash and Tech Lead for Personalization and Search for New Verticals which comprises groceries, convenience, retail, alcohol, and more. Together, we explore what personalized recommendation, ranking and search look like when the catalog is groceries and the delivery window is minutes. We start with the three-sided marketplace and the three pillars his team organizes the consumer side around: familiarity, affordability and novelty. From there we get into #LLM-based consumer profiles: memory blocks, structured natural-language representations of a consumer organized by semantic domain, which feed LLM-generated collections that are resolved into real items through embedding-based retrieval. The second half is about #SemanticIDs — hierarchical product identifiers built by recursively clustering item content embeddings, so the taxonomy is learned rather than hand-built. Two applications run in production: the ID prefixes serve as aggregation keys for dense ranking features — alongside the hand-built taxonomy, not replacing it — cutting popularity bias in the top slot by 18% while adding 8% add-to-cart at position 1; and query reformulation in search, traversing the learned hierarchy to refine or diversify a query. We close that section on agentic shopping versus conversational RecSys. 👨‍🏫 🤝 And the conversation doesn't end here! Raghav and I are very much looking forward to taking these and many more topics to the stage with all of you — our tutorial "Recommender Systems in Delivery Platforms: Challenges, Solutions and Learnings" at #RecSys next week in Minneapolis, joined by our colleague Paavo Leonhard C. from Wolt. Come by and bring your questions. 🙌 Listen to this new episode of #Recsperts - Recommender Systems Experts! ✅ Subscribe on your favorite podcast player ⭐ Leave a rating, review or comment ✍ Share your thoughts on Generative RecSys and Semantic IDs 🎧 Website: https://lnkd.in/gKucy3Vh 🎧 Spotify: https://lnkd.in/g4qYHUs9 🎧 Apple: https://lnkd.in/gsRKsZM9

42

Sarvam AI just dropped two very interesting open-weight reasoning models. While everyone is waiting for DeepSeek V4, India just introduced two serious contenders: Sarvam 30B and Sarvam 105B. As a Data Scientist working on building AI agents, I’m less interested in hype and more interested in the engineering decisions behind these models. What stands out here isn’t just scaling - it’s thoughtful architectural optimization The Architectural Highlights 1️⃣ Multi-Head Latent Attention (MLA): The 105B model adopts MLA (popularized by DeepSeek-V2), which dramatically reduces the KV-cache bottleneck. Despite being nearly 4× larger than the 30B model, it maintains a surprisingly efficient memory footprint 2️⃣ Tokenizer Built for India: Sarvam trained a tokenizer from scratch, optimized for Indian languages, achieving ~4× higher token efficiency. In production systems, this directly translates to lower latency and lower API cost 3️⃣ Kernel-Level Optimizations: Beyond architecture, Sarvam implemented low-level kernel and code optimizations, reportedly delivering 20–40% higher throughput compared to comparable models like Qwen-3 Built for Agents: While many models focus on coding benchmarks, Sarvam performs surprisingly well on agentic reasoning benchmarks like Tau2, suggesting these models may be particularly useful for AI agents and task-completion systems, not just chat Deep Dive References: For the fellow nerds wanting to see the "why" behind the "what," check out these foundational papers: 1️⃣ DeepSeek-V2 (MLA Architecture): The blueprint for the 105B’s memory efficiency. (https://lnkd.in/g6txGckZ) 2️⃣ The GQA Paper: The logic behind the 30B’s balanced performance. (https://lnkd.in/g_XTk9FZ) 3️⃣ The Tau Bench: Why "agentic reasoning" is the new benchmark that actually matters. (https://lnkd.in/g627Qc5H) India is no longer just a consumer of AI - we are architecting its future #GenerativeAI #LLM #SarvamAI #AIArchitecture #DeepLearning #TechInnovation #IndiaAI

10

Skills That Matter for Building LLM, RAG, and Agentic AI Applications Building LLM-, RAG-, and agentic AI- applications is less about any single discipline and more about the amalgamation of skills across engineering, data science, and product thinking. Teams achieve this amalgamation not through deliberate pairing, shared ownership of system behavior, and cross-functional reviews. Here’s how I think about the skills that need to come together - and how teams can intentionally blend them. Core skills that cut across LLM, RAG, and agentic systems • problem framing and task decomposition • prompt and instruction design as an interface, not magic • retrieval design: what to fetch, when, and why • evaluation thinking beyond accuracy (failure modes, drift, edge cases) • understanding probabilistic outputs vs deterministic expectations • designing guardrails, constraints, and human-in-the-loop flows These systems succeed or fail at the system level, not the model level. System engineering skills that become critical System engineers already bring strengths in scalability, reliability, and integration. What’s new is learning to operate in a stochastic world. Key shifts: • deterministic → probabilistic behavior • idempotent logic → best-effort reasoning • fixed outputs → distributions and confidence • debugging code → diagnosing behavior What system engineers need to add: • a working understanding of how models learn from data • why LLMs can sound confident and still be wrong • how retrieval, memory, and tools shape outcomes • how to design for uncertainty, not eliminate it You don’t need to become a data scientist - but you do need to understand the underbelly of AI: what’s real signal, what’s pattern matching, and where illusions come from. Data science skills that complement this shift Data scientists bring deep strengths in: • statistical thinking • uncertainty, bias, and variance • model evaluation and validation What many now need to add: • software engineering discipline (APIs, services, observability) • system-level thinking beyond isolated models • product and workflow awareness • operating constraints: latency, cost, reliability, governance Beyond concepts, some concrete skills matter: • designing retrieval pipelines and embeddings • building evaluation harnesses for LLM behavior • understanding orchestration frameworks and tool use • measuring performance at the application level, not just model metrics The real shift LLM, RAG, and agentic applications sit between disciplines. They require: • engineering judgment • statistical thinking • product sense • and systems design The teams that succeed won’t be the ones with the “right titles.” They’ll be the ones that deliberately blend these skills — and know where human judgment still belongs. #AI #EnterpriseAI #LLM #RAG #AgenticAI #AIEngineering #ResponsibleAI #TYS #RayAI #AILiteracy #IntelligenceInBits

9

1 Comment

#Sarvam solves multi-lingual problems for India, then why Mātr? Because we are the “efficiency layer” for similar native models. Hear me out. Sarvam, Jais and similar "native-first" initiatives are vertical solutions. They optimize the model architecture, tokenization, and training corpus for a specific, high-value set of languages (e.g., Indian languages, Arabic etc).  #Mātṛ is a horizontal layer. The core difference lies in how we solve the problem: Native Models (e.g., Sarvam): They solve the tokenization penalty by re-training or adapting the base model to be more efficient with specific script structures. This is highly effective but expensive and geographically bounded - you cannot feasibly re-train a model for every one of the world's 7,000+ languages, can you? Mātṛ (Universal Interpretation): We are solving the mathematics of divergence. By modeling how languages deviate from one another rather than just training on a dataset, Mātṛ can be applied to any existing model - even one like Sarvam. If a "native" model encounters a dialect or a linguistic structure outside its specific training optimization, it will still suffer from token bloat and hallucination. Mātṛ acts as the "safety net" and "efficiency layer" for the long tail of global languages that native models cannot feasibly cover. More in comments. #nativeAimodels #sovereignAi #linguisticEquity #Universalinterpretation

15

3 Comments

Continuing from Post 6 as to how do we actually align raw LLMs? 💡 Mini-Series: Decoding LLMs (Post 7) Last week, we saw why pre-training alone gives us a fluent but unhelpful model. The first fix is Supervised Fine-Tuning (SFT). Take GPT-3 (2020). It could generate essays and poems, but if you asked: “Write a SQL query for the top 5 customers by revenue” it often gave a response with a paragraph about what SQL is. After OpenAI fine-tuned GPT-3 on instruction response pairs, the resulting model InstructGPT (Ouyang et al., 2022) produced the actual query: SELECT customer_id, SUM(revenue) as total_revenue FROM sales GROUP BY customer_id ORDER BY total_revenue DESC LIMIT 5; That’s the shift SFT creates: from fluent but unhelpful text to direct, useful answers. 🔍 So here is how Supervised Fine Tuning works !! SFT is simply supervised learning applied to a pre-trained model. ✔️ Dataset → curated instruction response pairs (often 1k–100k examples). Example: “Explain quantum entanglement to a 10-year-old.” → “Imagine two teddy bears that are magic twins…” ✔️ Training objective → same cross-entropy loss as pre-training, but applied only to the demonstration outputs. ✔️ Optimization loop → forward pass → compute loss on target tokens → backpropagation → update weights. After enough examples, the model shifts from just predicting words to acting like an assistant. ⚠️ Limitations SFT makes models useful, but it’s still mimicry. They copy examples without knowing why one answer is better. Performance is capped by dataset quality and diversity, and hallucinations remain fluent, but not always truthful. That’s why SFT is only the first step in alignment, but full fine-tuning usually means updating billions of parameters which is far too expensive. LoRA (Low-Rank Adaptation) solves this by keeping the main model frozen and only training a few small “adapter” weights. It’s like adding knobs to a giant machine instead of rebuilding it. The result? Even a 7B model can be finetuned on a single GPU, and you can swap adapters for different domains like coding, medicine, or law. Takeaway: SFT is the bridge from capability to utility. It transforms a pre-trained model into an instruction follower. And with LoRA, fine-tuning becomes practical at scale. 👉 Next up: a closer look at LoRA how parameter-efficient fine-tuning made LLM adaptation accessible, before we dive into advanced post-training strategies #AI #LLM #SFT #LoRA #DeepLearning #GenerativeAI #MachineLearning #NLP #Alignment #DataScience

17

Wrote an article on Byte Pair Encoding (subword Tokenization) and how it works which was potentially used in the the earlier versions of GPT / traditional Language Learning model paradigms. Took a sample of Hindi Language corpus text data to showcase the same. Article below - https://lnkd.in/gCu7Bgpf

33