Diego Granados

Intuit Mailchimp just rolled out one of our biggest updates ever—and it’s a game changer for #ecommerce. 🚀 By combining unified data with powerful automation, Mailchimp is helping brands deliver higher ROI for less cost. Check out the details here: https://lnkd.in/gMyJGQ4S

3

I just finished watching all of the top AI-related sessions from the World Economic Forum 2026, with leaders like Jensen Huang, Satya Nadella, and Dario Amodei. The consensus? The constraints that restricted our abilities to drive product innovations years ago have largely evaporated. Here are 5 "Impossible" Product Ideas from two years ago that are now engineering realities: 1. The "Self-Coding" Feature Suite 💻 • The 2026 Reality: The loop is closing. Anthropic predicts that within months, models will perform most maybe all of what software engineers do end-to-end. We are moving from "Human in the loop" to "Human in the lead," meaning products can now autonomously maintain and upgrade themselves. 2. The "Universal Shopper" That Actually Buys 🛒 • The 2026 Reality: "Agentic Commerce" is here. Visa has rolled out protocols where agents carry attribute payloads (spending limits, preferences) to buy natively on platforms. The buy button is moving from the merchant's site directly into the AI agent. 3. The "Physical" Health & Care Assistant 🏥 • The 2026 Reality: AI has moved to "Physical Intelligence." It now understands the laws of physics and biology. With humanoid robots slated for public sale by next year, the hardware finally exists to deploy the software we dreamed of. 4. The "Brute Force" Intelligence Engine 🧠 • The 2026 Reality: The cost of intelligence has collapsed. Inference costs dropped 99% from GPT-4 to GPT-5 mini. We can now afford to throw massive test time compute at problems, allowing the model to think for minutes before answering, drastically reducing hallucination. 5. The "Sovereign" Enterprise Brain 🔐 • The 2026 Reality: Sovereignty is the new standard. The focus has shifted to "Sovereignty of the Firm", distilling your own private data into a model you control, ensuring your IP doesn't leak into a generic API. The agents are ready to work. Time to find that dry-erase marker!

2

𝗧𝗵𝗲 𝗩𝗲𝗿𝗯𝗼𝘀𝗶𝘁𝘆 𝗣𝗿𝗼𝗯𝗹𝗲𝗺: 𝗪𝗵𝘆 𝗬𝗼𝘂𝗿 𝗔𝗜 𝗢𝘂𝘁𝗽𝘂𝘁𝘀 𝗧𝗼𝗼 𝗠𝘂𝗰𝗵 (𝗔𝗻𝗱 𝗛𝗼𝘄 𝘁𝗼 𝗙𝗶𝘅 𝗜𝘁) We've trained ourselves into a bad habit with LLMs: dump context, get a wall of text, copy what we need. But here's the irony—the verbose output that feels "thorough" often means the model is compensating for ambiguity by covering all it’s bases. 𝗧𝗵𝗲 𝗺𝗲𝗰𝗵𝗮𝗻𝗶𝗰𝘀 𝗺𝗮𝘁𝘁𝗲𝗿 LLMs don't "understand"—they predict the next token based on everything that came before. Tokenization fragments your intent into pieces the model reassembles probabilistically. Long, ambiguous prompts create more room for drift. Each completion builds on the last, compounding small misalignments into large ones. RLHF reasoning models feel smarter. They are. But even with emergent chain-of-thought processing, you may not see their reasoning—and when you do, it may not align with yours. The model optimized for "helpful" isn't the same as the model optimized for "correct given your specific goal." 𝗧𝗵𝗲 𝗳𝗶𝘅: 𝗣𝗿𝗼𝗴𝗿𝗲𝘀𝘀𝗶𝘃𝗲 𝗜𝗻𝘁𝗲𝗻𝘁 𝗗𝗲𝘁𝗲𝗿𝗺𝗶𝗻𝗮𝘁𝗶𝗼𝗻 Two instructions that change everything: 𝟭. "𝗕𝗲𝗳𝗼𝗿𝗲 𝗿𝗲𝘀𝗽𝗼𝗻𝗱𝗶𝗻𝗴, 𝗮𝘀𝗸 𝗰𝗹𝗮𝗿𝗶𝗳𝘆𝗶𝗻𝗴 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻𝘀." 𝟮. "𝗟𝗶𝘀𝘁 𝘆𝗼𝘂𝗿 𝗮𝘀𝘀𝘂𝗺𝗽𝘁𝗶𝗼𝗻𝘀 𝗯𝗲𝗳𝗼𝗿𝗲 𝗽𝗿𝗼𝗰𝗲𝗲𝗱𝗶𝗻𝗴." This surfaces the gap between what 𝘺𝘰𝘶 meant and what the 𝘮𝘰𝘥𝘦𝘭 inferred. Now you can align it with your actual objective before it generates 2,000 words in the wrong direction. Why this works: → 𝗥𝗲𝗱𝘂𝗰𝗲𝘀 𝗼𝘂𝘁𝗽𝘂𝘁 𝗯𝗹𝗼𝗮𝘁 — The model doesn't hedge when it understands your actual goal → 𝗔𝘃𝗼𝗶𝗱𝘀 𝘁𝗵𝗲 𝗨-𝘀𝗵𝗮𝗽𝗲𝗱 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 𝗽𝗿𝗼𝗯𝗹𝗲𝗺 — Critical details buried in long prompts get lost; shorter exchanges keep everything salient → 𝗜𝗺𝗽𝗿𝗼𝘃𝗲𝘀 𝘆𝗼𝘂𝗿 𝗼𝘄𝗻 𝘁𝗵𝗶𝗻𝗸𝗶𝗻𝗴 — Seeing the model's assumptions forces you to clarify your own 𝗧𝗵𝗲 𝗵𝗮𝗿𝗱 𝗽𝗮𝗿𝘁: 𝗬𝗼𝘂 𝗵𝗮𝘃𝗲 𝘁𝗼 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝗿𝗲𝗮𝗱 𝘁𝗵𝗲 𝗼𝘂𝘁𝗽𝘂𝘁. Human validation isn't optional. If you're skimming and accepting, you're not collaborating—you're gambling. And if the session has drifted too far? Don't try to correct mid-stream. Start fresh. A new conversation with clear intent beats a long thread full of compounding misalignment. 𝗪𝗵𝘆 𝘁𝗵𝗶𝘀 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 𝗳𝗼𝗿 𝗮𝗴𝗲𝗻𝘁𝘀 When the model is just generating text, misalignment costs you time. When the model has tools—when it can send emails, modify files, update databases, provision resources—the intent gap has real consequences. An agent that misunderstands your goal doesn't just give you a bad answer. It takes bad actions. At scale. Automatically. This is why agentic systems require more than better prompts. They require intent verification, human-in-the-loop checkpoints, and architectures that make assumptions visible before tools execute. What assumptions is your model making that you've never seen? #AI #LLM #Agents #EnterpriseAI #PromptEngineering

5