Insights Gained From AI Experiments

Everyone's talking about AI-powered Deep Research—so I tested all three: ChatGPT (yes, available last week to $20/month Plus users!), Gemini, and Perplexity. My top observations are that AI research tools are evolving fast—but not without their trade-offs. We all have those P1 and P2 tasks sitting on our to-do lists because they require too much research time: ▪️ Simulating different wealth planning scenarios—stock research, tax strategies, and how lifestyle choices impact long-term financial health. ▪️Catching up on AI trends like agentic workflows and how researchers and leaders are reacting to DeepSeek. Instead of manually searching, opening 20+ tabs, and filtering noise, I tested all three LLMs to see how well they could develop a deep research plan and synthesize insights. In about 10 minutes, I had three structured reports—each compiled from 150+ sources. The Delighters: 🚀 Speed & Effectiveness – I learned more about personal finance strategies in 30 minutes than I would have in an entire weekend of Googling! 🔍 Expanding My Curiosity – It encouraged me to explore topics I would’ve ignored and helped me ask better questions in follow-ups with human experts. 🌍 Diverse Perspectives – It pulled insights from sources I might never have found on my own. But is it perfect? Not quite. Ben Thompson called Deep Research “a pretty decent research assistant”—great at surfacing existing knowledge but not generating new insights. Benedict Evans was more critical: he found that some numbers were just wrong or based on weak sources. His take? AI research tools are like “infinite interns”—useful, but requiring oversight. My Take: 🔹 I still verify the data, and I’m skeptical of some sources. But the time savings make it worth it. 🔹 If I need 100% accuracy, I double-check. But for fast learning and directional insights? It’s a no-brainer. The Bigger Shift: AI research is moving beyond just saving time. It’s forcing us to rethink who does research and how it’s valued. Case in point: Washington’s General Services Administration (GSA) just ordered federal agencies to terminate contracts with the top 10 consulting firms. If AI can synthesize insights faster and cheaper, what does that mean for industries built on deep research? But I’m optimistic—I see AI research tools empowering professionals to focus on higher-value thinking, rather than replacing them. How are you using AI for deep research?

Johnson & Johnson is zeroing in on GenAI use cases that it sees a strong ROI for and shutting down pilot projects for which it doesn't - and there are some powerful lessons here for all of us: 𝐖𝐡𝐚𝐭'𝐬 𝐭𝐡𝐢𝐬 𝐚𝐛𝐨𝐮𝐭?  - J&J initially encouraged employees across the company to experiment with AI, resulting in ~900 GenAI experiments across R&D, commercial, HR, and supply chain. - After reviewing, only 10–15% of these delivered 80% of the business value. - Now they're prioritizing and only scaling the high-value use cases and axing the rest. 𝐎𝐭𝐡𝐞𝐫 𝐢𝐧𝐭𝐞𝐫𝐞𝐬𝐭𝐢𝐧𝐠 𝐜𝐡𝐚𝐧𝐠𝐞𝐬:  - They're moving away from a centralized GenAI governance board. - And letting each business unit own their own AI agenda. - While setting up AI and Data Councils to ensure ethical use and scalability of AI tools. 📘 𝐇𝐚𝐯𝐢𝐧𝐠 𝐬𝐩𝐞𝐧𝐭 𝐲𝐞𝐚𝐫𝐬 𝐢𝐧 𝐂𝐨𝐫𝐩𝐨𝐫𝐚𝐭𝐞 𝐈𝐧𝐧𝐨𝐯𝐚𝐭𝐢𝐨𝐧, 𝐭𝐡𝐞𝐫𝐞'𝐬 𝐬𝐨𝐦𝐞 𝐭𝐞𝐱𝐭𝐛𝐨𝐨𝐤 𝐢𝐧𝐧𝐨𝐯𝐚𝐭𝐢𝐨𝐧 𝐩𝐥𝐚𝐲𝐛𝐨𝐨𝐤 𝐦𝐨𝐯𝐞𝐬 𝐢𝐧 𝐡𝐞𝐫𝐞:  - Start broad. Learn fast. Then double down on what works. - Some people are calling this a failure. I completely disagree with that. This is what smart scaling looks like. - J&J mastered the Experiment → Validate → Prioritize → Operationalize cycle and built the real execution muscle around this. - Experimentation isn't just about wins; It's about building your path to value. 👉 𝐈𝐟 𝐲𝐨𝐮'𝐫𝐞 𝐢𝐦𝐩𝐥𝐞𝐦𝐞𝐧𝐭𝐢𝐧𝐠 𝐀𝐈 𝐢𝐧 𝐲𝐨𝐮𝐫 𝐨𝐫𝐠𝐚𝐧𝐢𝐳𝐚𝐭𝐢𝐨𝐧, 𝐮𝐬𝐞 𝐭𝐡𝐞𝐬𝐞 𝐥𝐞𝐬𝐬𝐨𝐧𝐬 𝐢𝐧 𝐲𝐨𝐮𝐫 𝐬𝐭𝐫𝐚𝐭𝐞𝐠𝐲 𝐚𝐧𝐝 𝐫𝐨𝐚𝐝𝐦𝐚𝐩: start broad, find what works, then scale proven value. ♻️ Share this with someone who needs to know this playbook. ➕ Follow me Heena Purohit for more AI news, insights, and real talk. 👉 Over to you: What aspects of this story stood out to you? --- 🔗 Full article: https://lnkd.in/dY_Mb4uE #EnterpriseAI #ArtificialIntelligence #AIforBusiness #GenerativeAI #AIRealTalk

OpenAI built their agent by training on what people actually do. Not math problems. Not coding challenges. Not abstract benchmarks. Real human tasks. Here's the breakthrough insight from OpenAI's Deep Research team: Instead of hoping models would generalize from academic tasks to real-world use, they took a different tact. They trained directly on browsing behavior. They studied how humans research online. They captured the messy, multi-step process of finding and synthesizing information. The result? An AI that can spend 30 minutes researching what would take humans hours. It searches multiple sources. Reasons through conflicting information. Synthesizes findings into analyst-level reports. Provides precise citations. The key insight: Training on real human workflows beats hoping for generalization. This changes how we should think about AI development: - Start with the end user behavior - Train on actual tasks, not proxies - Make the AI learn human reasoning patterns One example: The team leader used it to find Korean night markets within 15 minutes of her location, cross-referencing Reddit and Korean sources she couldn't read herself. It delivered specific recommendations with exact store ratings. That's the power of training on real human needs. — What real-world task would you want AI to master next? (Insights from Isa Fulford's presentation at Sequoia AI Summit) — Enjoyed this? 2 quick things: - Follow me for more AI automation insights - Share this a with teammate