Let's Talk
← Back to Pipeline podcast

ai-realist-2026-brief.md

ai-realist-2026-brief.md

# What You Can Realistically Expect from AI in 2026 **Runtime:** ~18 minutes **Tenet:** AI + Clarity ## Cold Open "Everyone's talking about what AI *will* do. Today I'm going to tell you what it *can't* do — and that's where the actual opportunity is." ## Segments ### 1. The Setup (3 min): The Hype vs Reality Gap AI is everywhere. CEOs are demanding "AI strategies." But here's the dirty secret: most AI projects fail because people don't understand the hard limits. Today we're going to look at the actual research — not the marketing — on what AI can and cannot do. This isn't pessimism. It's clarity. And clarity is where you find the edge. ### 2. The Deep Dive — What AI Can't Do (8 min) **The Reasoning Illusion** Apple researchers proved this in 2024. Take a simple maths problem — "Sophie has 5 apples, buys 3 more." The AI gets it right. Now add: "Sophie also has a red shirt." Suddenly the AI fails 65% of the time. Why? It's pattern-matching, not reasoning. It can't tell what's relevant. Implication for you: Don't trust AI for novel situations. Use it for known patterns. **Model Collapse** Here's the thermodynamic truth: AI trained on AI output dies. Within a few generations, the output becomes gibberish. Human creativity is the irreplaceable fuel. Implication: Your people matter more, not less. **The Creativity Trap** Wharton study: AI-assisted groups produced ideas that were only 6% unique. Everyone converged on "Build-a-Breeze Castle." The AI is a normalisation engine, not an innovation engine. Implication: Use AI to raise the floor, not the ceiling. Breakthroughs still require humans. ### 3. How to Actually Build (5 min) If you're building with AI — or hiring people who are — here's what separates the amateurs from the professionals: **Research → Plan → Implement** Never let AI start coding without research first. The biggest failure mode is "vibe coding" — just asking AI to build something and hoping it works. **Context Engineering** The AI's memory is precious. Don't dump everything in. Select, compress, isolate. **Test-Driven Development** Write the test first. Make the AI pass the test. This is how you get reliable outputs from probabilistic systems. ### 4. Outro (2 min) The opportunity isn't in pretending AI can do everything. It's in knowing exactly where it fails — and building humans and processes around those gaps. That's the Clarity play. That's where we win. Next episode: How to orchestrate multiple AI agents without losing your mind. ## Quotable Moments - "AI doesn't ask for clarification. It amplifies the ambiguity at scale." - "Model collapse is thermodynamic. You can't train AI on AI forever — the entropy wins." - "The AI is a normalisation engine, not an innovation engine. Use it to raise the floor, not the ceiling." ## Links for Show Notes - Apple GSM-Symbolic research: arxiv.org/abs/2410.05229 - Model Collapse study: Nature 2024 - Wharton Creativity study: knowledge.wharton.upenn.edu