Nitesh Ambastha

Earlier, I used to think — “Prompts? They’re just plain English. What’s there to learn?” But after spending time with Gen AI models and experimenting with hundreds of prompts, I realized something important — 👉 If you master the art of writing prompts, the model takes care of the science behind them. Prompt engineering isn’t just trendy — it’s essential. And leading companies are already acting on it: 💡 IBM is scaling AI upskilling across the enterprise and has pledged to train 2 million workers in generative AI & prompt-engineering skills. (Futurum Group) 💊 Johnson & Johnson has mandated generative-AI training — including prompt engineering — for 56,000+ employees. (Business Insider) The takeaway? 🧠 The ability to craft effective prompts is becoming as fundamental as coding itself. Whether you’re a business leader, HR strategist, or practitioner — it’s time to embed prompt mastery into your team’s DNA. 🔧 Let’s not treat prompt engineering as a one-and-done session, but as a continuous capability to stay future-ready. 🎥 I’ve launched a new video series to help you start — watch the first episode now! https://lnkd.in/gXNmXau7 #PromptEngineering #AIUpskilling #GenAI #FutureOfWork #AITraining

26

2 Comments

MIT’s latest research reveals a subtle yet critical flaw in transformer-based LLMs: “position bias.” These models tend to overweight the start and end of documents while overlooking the middle—potentially undermining accuracy in fields like financial compliance, legal review, and risk analysis. What can be done? MIT proposes solutions including: 1. Smarter attention masking 2. Refined positional encodings 3. Balanced training data As we scale AI at Northern Trust, we’re paying close attention to these foundational issues. Building trustworthy, context-aware AI starts deep in the architecture. #AI #ResponsibleAI #LLMs #FinancialServices #MIT #EnterpriseAI #GenAI #NorthernTrust #AITrust

33

I recall getting lost in awe at Leke’s barbershop - was it the dexterity in his craft (he was one-hand) or at people dozing off while he reformed their looks?… Well, I recently almost fell off the chair myself. That moment was different. Nobody prepares you for the quiet fatigue that comes with leadership, especially when you're the one everyone looks to for clarity, speed, and direction. As a team lead, it’s easy to slip into “maintenance mode”: ✔️ Morning syncs ✔️ Back-to-back meetings ✔️ Progress tracking ✔️ Firefighting ✔️ Reporting ✔️ More meetings You start cutting out “distractions” like breaks, social catch-ups, or rest-only to end up running on fumes. Leadership isn't just about ensuring deliverables are met. It’s about carrying the emotional weight of the team, keeping morale up, and still showing up with energy. Balance isn't a myth—but it has to be intentional. Rest is part of the strategy. Burnout helps no one. So, if you’re in leadership and running on autopilot, this is your reminder: Recharge. Reflect. Reset. Even leaders need care too.

34

7 Comments

After Peak Data… Came Peek Data. We’ve reached Peak Data. Enterprises now sit atop data lakes, ECM platforms, search indices and robust pipelines. They have deployed LLMs trained on the filtered internet—aligned, benchmarked, safety-tested. But they still don’t understand how our firms actually work. They tried to fix it with RAG—retrieval as a bandaid. It didn’t always work. It gave us better autocomplete, not better judgment. So we pivoted to agentic systems, thinking orchestration would do what understanding couldn’t. But all we have is more flowcharts as this technology matures. Fewer decisions. Meanwhile, what we’ve really built is this: • Internet-quality data through the model, • ECM, KB or Index retrieval through the pipelines, • Snowflake-grade structures through the stack… … and zero enterprise context into the model. That’s how we landed in the Peek Data era. Models that peek, but never know. They skim. They style. But they don’t understand. Unless … So what now? If we want real enterprise AI—models that think in your business context—we have to go beyond wrappers and routers. We need to: • Run Domain-Adaptive Pretraining (DAPT) on our internal corpora • Use model merging to blend domain models for cross functional context • Feed models the decisions, not just the documents Peek access isn’t enough. We need pretraining, not just prompting. Merging, on top of masking for CLM. Because until the model knows us, it works for someone else. Arcee AI NVIDIA AI #tPower #mlai #GenAI4FS #inc81starch

30

1 Comment

Why AI Could Be Doomed in the Next Few Years – Or Why It Might Doom Us First Over the last few months, I’ve been deep in the trenches, using AI to build real-world applications. The speed? Incredible. The efficiency? Game-changing. But there’s one thing AI hasn’t earned yet: trust. This led to a discussion with my team, starting with a blunt question: “Why might AI be doomed in the next few years?” We’re pushing AI-powered autonomy at full throttle… but are we asking the right questions? • Who’s accountable when AI gets it wrong? • How do we measure success beyond “it just works”? • What happens when AI faces something it wasn’t trained for? The risks aren’t hypothetical anymore. We’ve seen hallucinations creep into production. We’ve heard of AI wiping out a production database and fabricating thousands of records. This isn’t theory—it’s reality. ⸻ ⚙️ AI: Force Multiplier or Bubble? Some say AI isn’t another dot-com bubble—it’s more like Big Data, where hype gives way to specific use cases. But here’s the paradox: AI is here to stay and it’s misunderstood and misused. AI now writes production code—but still needs human review. It’s like giving a Ferrari to a learner driver: flashy, fast, but dangerous without control. Yes, AI boosts productivity. But reliability? That’s still on us. ⸻ 🔁 Will AI Replace Developers? Probably not. But it will shrink teams—from 10 engineers to 2. Not because AI is flawless, but because businesses want faster delivery without more headcount. As one senior leader told me: “The business doesn’t care how much code AI wrote—they care if your team delivers faster.” And this isn’t limited to tech. Operations, compliance, finance, and legal—junior and mid-level roles—are already at risk. 🧾 One Real-World Example Preparing comparative earnings reports for banks used to take 8–10 people. Today, AI can do it in under 15 minutes—needing just 2 people to validate. That’s 80% job displacement from one use case. 💥 The Real Danger: Skill Decay & Overdependence We’re heading toward a future where: • Developers paste code into ChatGPT without understanding it • AI-generated code becomes too complex to unravel • Engineers lose the intuition that makes them good If that happens, we won’t just use AI—we’ll depend on it for everything, even explaining itself. 🚀 What’s Next? AI could mark the peak of a 20-year wave of IT spending—only this time, budgets may drop 20–30% as automation takes over. That forces a tough question: If your job is automated, what new value can you create? Those who adapt will thrive. Those who don’t… will wait for AI’s next command. 💡 Final Thought AI won’t fail because it’s incapable. It will fail because we: • Adopt it blindly without accountability • Measure speed, not quality • Let human expertise wither without safeguards If we keep using it without responsibility and intention, the problem won’t be AI failing us… …it will be us failing ourselves.

31

2 Comments

Mind the Gap — the divide between well-funded enterprises and resource-strapped SMBs is getting wider,and attackers have learned to monetise it at industrial scale. In this short article, I unpack the data proving this gap is widening, what's accelrating this problem and suggest a seven-step, budget-friendly playbook any small or mid-sized business can use to slam common attack doors shut. https://lnkd.in/gJrdFCbT #CyberSecurity #Ransomware #SMB

55

2 Comments

Our own EVP, Chief Scientist & Head of Enterprise AI Prem Natarajan recently spoke to TechTarget about Capital One’s partnership with the National Science Foundation (NSF), which will help launch five new National AI Research Institutes to drive breakthroughs in high-impact areas across society like material and computational sciences, economics, and mental health. https://bit.ly/46fLI5I

1

Yıllarca bankacılığın yönü tek bir eksende tanımlandı: Daha fazla otomasyon, daha hızlı dijitalleşme, daha az temas. Hız, başarının sembolü oldu. Ancak bugün şubelerin içinde farklı bir gerçeklik şekilleniyor. Müşteriler yalnızca işlemlerinin tamamlanmasını istemiyor; anlaşılmak, önemsenmek ve değer görmek istiyor. Çünkü finansal kararlar yalnızca sayılardan ibaret değil. Bir konut kredisi, bir yatırım kararı ya da bir yapılandırma talebi; çoğu zaman hayatın en kritik eşiklerinden biri. Ve bu eşiklerde belirleyici olan hız değil, güvendir. Algoritmalar seçenek sunar. İnsan ise perspektif ve güven inşa eder. Şubeler kuyruk düzeninden pod tasarımına, işlem mantığından diyaloğa geçtiğinde dönüşen sadece mekân değildir — ilişki biçimidir. Sadakat hızdan değil, kişiselleştirilmiş güven deneyiminden doğar. Bankacılığın geleceği dijital ile insan arasında bir tercih değil. Gelecek, teknolojinin görünmezce güçlendirdiği; insanın ise merkeze yerleştiği bir tasarım anlayışıdır. 🌎 For years, the industry narrative has been clear: automate more, digitize faster, reduce friction. Speed became the primary metric of success. Yet something unexpected is happening inside physical branches. Customers are not just seeking efficiency — they are also seeking acknowledgment. Because most of the financial decisions are deeply personal. Whether it is a mortgage, an investment decision, or a restructuring request, customers are navigating uncertainty. In these moments, what they value most is not speed — it is reassurance. When branches shift from queues to pods, from transactions to conversations, loyalty increases — not because service is faster, but because it is more personal. The future of banking isn’t digital vs. human. It’s digital with human. #FutureOfBanking #BranchTransformation #CustomerExperience #PhygitalBanking

32

2 Comments