Asif Habib

Brilliant piece by Will Larson - https://lnkd.in/gkyRE-Q8 I find myself frequently describing LLMs as just another component in our toolkit, and agents as just standard software wrapped around API calls to LLMs that may or may not have access to “tools” - yup, more standard software. Of course, as this is all just software, you can combine it to your heart’s content, creating larger software systems including ones with an agent as the main orchestrator. The compounding non-determinism of systems like these will either feel like magic or a nightmare. Once you see it that way, it’s much easier to reason about.

12

Gemini 3 is here! It was incredible reading about Google's new model. It's feels far more than a simple performance bump—it signifies a real shift where AI moves into being a genuine “thought partner”. Few features that stood out: • State-of-the-Art Reasoning: I'm excited that this model is built to genuinely grasp depth and nuance, which means it's phenomenal at "reading the room" for complex problems and is setting a new standard for intelligence. • Advanced Agentic Power: What really caught my attention is Google's claim that this is their most powerful agentic and vibe-coding model yet. It enables planning and executing complex, multi-step workflows, which is a massive leap forward for productivity. • Broad Multimodality: I was amazed to read it's claimed as the best model in the world for multimodal understanding. It can seamlessly synthesize information—vision, code, and text—helping us learn, build, and plan complex tasks in totally new ways. Looking at these features, we are entering a new era of intelligence. I absolutely cannot wait to see what everyone builds with this! #Gemini3 #GoogleAI #AIInnovation #FutureOfWork #LLM #AI https://lnkd.in/guMDTxAU

114

4 Comments

As generative AI moves from experimentation to production, the conversation is still heavily skewed toward capability. What can the model do? How fast can it deliver results? How does it compare on benchmarks? Those are important questions. However, they are incomplete. When AI systems enter real workflows, such as healthcare, finance, customer service, HR, the central issue shifts. The question is no longer just technical performance; it's operational risk. I've found it useful to think about AI productionization across three dimensions: 1️⃣ Technical Capability Model performance, reliability, latency, scalability. This is where most investment and executive attention goes (understandably so). If the system can't perform, nothing else matters. 🤷‍♀️ But beyond benchmarks: can we actually measure whether the model's outputs are valid in production? Not just accurate in testing, but defensible in the real workflow it's embedded in. 2️⃣ Governance & Accountability Clear ownership of outcomes. Defined escalation paths when something goes wrong. Auditability that a legal team (and a regulator) can actually work with. And critically: a plan for what happens when the system produces harm. (not if, but when) This layer also includes asking whether a foundation model is even the right solution. There are problems where traditional ML, rule-based logic, or a simpler heuristic is more auditable, more explainable, and frankly more appropriate. Defaulting to LLMs is not a strategy. Without this layer, capability becomes liability. 3️⃣ Institutional Legitimacy Why should customers, regulators, and employees trust this system? Is its use transparent? Are its trade-offs acknowledged? This third dimension is where many AI initiatives stall out. Not because the model isn't impressive, but because the organization can't defend the system under real scrutiny. AI failures don't just create technical debt. They create trust debt. (and it compounds!) Responsible innovation, in this context, is not about slowing progress. It's about building systems that can survive contact with reality, i.e. procurement reviews, legal review, public attention, and edge-case failure. The organizations that win in AI will not be the ones with the most impressive demos. They will be the ones that integrate capability, accountability, and legitimacy into a coherent operating model ➡️ before customers force the issue for them. Which dimension is your organization pretending doesn't exist? #AIProductStrategy #ResponsibleAI #AILeadership #AIGovernance #TrustInAI #EnterpriseAI #AIProductionization Sources: AI Now Institute, AI-Generated Business AI Now Institute, Artificial Power: Executive Summary

24

4 Comments

SFTP integrations have been the standard for years, but they're holding back modern employment data operations. The core issues with file-based transfers: ▪️No real-time data access, limiting use cases and forcing batch processing ▪️Complex data structures require manual transformation and increase error risk ▪️Onboarding takes months of back-and-forth setup, mapping, and testing ▪️Scaling to multiple providers becomes unmanageable as each requires custom logic 97% of HR decision-makers say employment integrations are important to them. They expect real-time access, not weekly batch files. Our whitepaper breaks down when SFTP still makes sense and why APIs have become the preferred method for employment data. 🔗 Download it here: https://lnkd.in/e8Wt9Mfy

19

🧠'Agent Learning via Early Experience' As an AI Architect, this work stands out for its pioneering approach to training language agents—moving beyond passive imitation learning and bridging toward a true era of experience. Instead of relying on expert demonstrations or reward-based RL, the authors introduce a scalable middle-ground: agents learn by interacting with environments, then using those resulting states as direct supervision, without reward signals. 🔍 Key insights: • Motivation: Real-world agents often face hard-to-scale expert data and sparse—or absent—rewards. This paradigm lets agents learn from their own actions, scaling up diversity and generalization. • Methods: Two main strategies—implicit world modeling (internalizing environment dynamics) and self-reflection (learning from 'what went wrong'). Both use agents' own experiences to improve reasoning and robustness. • Results: Evaluated across 8 environments and major model families (Llama, Qwen), these approaches consistently outperform standard imitation learning, especially in complex and out-of-domain tasks. Early experience even boosts downstream RL, serving as a strong warm start for further finetuning. • Implications: Practical signal for agentic AI system design—build architectures that leverage interaction-derived learning, not just rewards or pre-collected demos. This can make agents more resilient, adaptive, and capable in open-ended domains (web, tools, APIs). Curious how 'early experience' could transform future agent frameworks and LLM deployments? 🔗 Attaching the original research PDF for those who want to dive deeper! #AI #LLM #AgenticAI #ReinforcementLearning #LanguageAgents #AIArchitecture

10

𝗧𝗵𝗲 𝗵𝗶𝗱𝗱𝗲𝗻 𝗰𝗼𝘀𝘁 𝗼𝗳 𝗔𝗜 𝗮𝘁 𝘀𝗰𝗮𝗹𝗲. He was the architect of their intelligent document processing platform. Built on Vertex AI. Fine-tuned transformers. Auto-scaling via Cloud Run. 2 million documents a month across 5 languages. It worked like magic. The Tech-Director called it a breakthrough. The sales deck showcased it on every slide. Customers loved it. Until the CFO saw the cloud bill. ₹𝟭.𝟮 𝗖𝗿𝗼𝗿𝗲𝘀/𝗺𝗼𝗻𝘁𝗵 𝗶𝗻 𝗚𝗖𝗣 𝘀𝗽𝗲𝗻𝗱. → ₹85L on inference → ₹20L on pipelines → ₹15L on storage, logging, and APIs 𝗥𝗲𝘃𝗲𝗻𝘂𝗲 𝗳𝗿𝗼𝗺 𝘁𝗵𝗲 𝗔𝗜 𝗳𝗲𝗮𝘁𝘂𝗿𝗲: ₹𝟭𝟭 𝗟𝗮𝗸𝗵𝘀/𝗺𝗼𝗻𝘁𝗵. Suddenly, the ask wasn’t “How can we make this better?” It was: “Can we afford to keep this alive?” So began his second act - not as an innovator, but as a surgeon.  • Rewrote pipelines in Cloud Functions to cut Vertex AI usage  • Switched from online to batch inference  • Distilled the model to a smaller size  • Lowered OCR quality settings to save credits  • Tradeoffs everywhere - accuracy, latency, user experience It worked. The cloud bill dropped by 78%. The feature survived - barely. But the “magic” slowly disappeared. And here’s the hard truth: In cloud AI, performance is rented. Intelligence is cheap to prototype, but expensive to scale. And the best engineers? Not just the ones who build powerful systems - but the ones who know what to sacrifice when power becomes too costly. How do you write “𝘴𝘢𝘤𝘳𝘪𝘧𝘪𝘤𝘦𝘥 𝘢𝘤𝘤𝘶𝘳𝘢𝘤𝘺 𝘵𝘰 𝘴𝘵𝘢𝘺 𝘢𝘧𝘭𝘰𝘢𝘵” on your resume? In the GenAI boom, let’s not forget: Innovation is exciting but 𝘴𝘶𝘴𝘵𝘢𝘪𝘯𝘢𝘣𝘪𝘭𝘪𝘵𝘺 𝘪𝘴 𝘴𝘶𝘳𝘷𝘪𝘷𝘢𝘭. #AI #CloudComputing #GenAI #MLOps #VertexAI #AIEngineering #TechLeadership #FinOps #ProductReality #CloudCosts

23

5 Comments

FINOS events always have great sessions and this topic is very timely. AI needs guardrails, and Neurosymbolic AI may be the perfect pattern for more consistent, explainable, and safer AI, which is critical for financial services.

2

We recently used our syftr multi-objective optimization framework to compare NVIDIA's Nemotron 49B v1.5 against few other larger models on FinanceBench. The smaller model, when optimized properly, delivered results comparable to much larger and more expensive alternatives. This matters for real-world AI deployment where cost and latency constraints are just as important as accuracy. Full details in our blog: https://lnkd.in/giAus2wb

21

By outsourcing core AI to Gemini, Apple signals that internal vertical integration of generative AI is no longer an option. Apple's decision to use Google Gemini's next-generation AI models is a total cede of AI to a competitor. Rather than competing in the AI space, Apple has given up and pulled in what is currently rated the top model provider. This one move brings with it several implications: Google is primed to overtake OpenAI in total userbase. OpenAI cannot compete with the combined Android and Apple ecosystem userbases. Apple's AI team will likely dissolve and go away as there is no reason to pay for top talent if you are going to just buy off the shelf AI. OpenAI will likely make some big announcement to try and increase their relevance in what has been a rough end of 2025. The continue to find themselves under increasing pressure to compete with incumbents that have substantial user bases with an app that is easy to leave.

44

12 Comments