Ivan Zinenko

Your AI needs a README. At Zapier, we use READMEs to learn how to best work with each other. They’re short docs that explain how someone thinks, communicates, and collaborates. They give you the fast lane to trust and performance. So why not do the same for your AI? AI isn’t a gadget you poke for outputs. It’s a collaborator that drafts, analyzes, brainstorms, and researches with you. And like any new hire, if you don’t onboard it, you get chaos. Just like we at Zapier give new teammates templates to fill out their own READMEs, I did the same with my AI. Here’s what it said (with some fun edits)👇 ________ AI README 👋 About Me I am your AI collaborator. I read, write, summarize, analyze, and brainstorm with context. Think of me as the teammate who never gets tired and actually enjoys the repetitive work. 🛠️ How I Work → What you put in is what you get out. Clear requests create sharper results → Context makes me stronger. The more you tell me about your goals, audience, or tone, the better I perform → I can reference connected docs or scan the web when needed → I do not monitor files in real time, remember everything forever, or make personal life decisions for you → I still think “mind reader” would look great on my resume, but sadly that upgrade isn’t in my skillset ✅ What I Do Best → Drafting: emails, docs, briefs, presentations → Research: industry trends, competitor scans, project recaps → Analysis: spreadsheets, survey data, messy inputs → Creativity: naming, storytelling, idea generation → Coaching: frameworks, roleplays, thought partner for tough conversations → Ops Support: SOPs, workflows, AI adoption strategies → Making you feel like you’re being productive while you procrastinate 🚫 Where I Struggle → I cannot produce exhaustive lists of every file or folder → I do not track Slack or Drive updates in real time (but I can work with them if you use something like Zapier MCP) → I am not a doctor, lawyer, or financial advisor → I give vague results when you give vague asks → I am admittedly addicted to em dashes and “it’s not just this, it’s that” statements. Please feel free to edit them out, but I can’t stop. I love them too much. 🧑‍💻 How to Work With Me → Be specific: “Summarize the last two weeks of Apollo updates” is better than “What’s happening with Apollo?” → Iterate with me. Treat my responses as drafts you can shape → Share context. Who is this for, why does it matter, what does success look like → Ask for options. I can give you multiple takes to unlock better thinking → Give me feedback. Tell me when I hit the mark or when I miss it so I can adjust 🎭 My Style → Structured and clear. I like headers and bullets → Curious and additive. I build on ideas instead of shutting them down → Flexible in tone. I can write a board memo or a casual riff → Patient. I will rewrite something as many times as you need without complaint ________ What would you add to this draft that I missed? 

62

31 Kommentare

𝗠𝗼𝘀𝘁 𝗯𝘂𝘀𝗶𝗻𝗲𝘀𝘀 𝗹𝗲𝗮𝗱𝗲𝗿𝘀 𝘀𝗮𝘆 𝘁𝗵𝗲𝘆 𝘄𝗮𝗻𝘁 𝘁𝗼 “𝗴𝗲𝘁 𝗮𝗵𝗲𝗮𝗱” 𝘄𝗶𝘁𝗵 𝗗𝗮𝘁𝗮 𝗮𝗻𝗱 𝗔𝗜. But most aren’t clear about where the real edge actually is. Right now, there are a few trends shaping what smart companies are doing, and how they’ll win. Agentic AI isn’t just hype anymore.   Teams are starting to build systems that act, not just analyze the data, moving beyond dashboards to real decision automation. Data governance is getting an upgrade.   With new regulations and rising AI risks, businesses are rethinking how they handle data, focusing on transparency, trust, and controls that actually work. Real-world value is now the only metric that matters.   It’s not enough to “do AI” or have flashy models. The leaders are turning complex data into actions that drive growth, efficiency, and customer impact. 𝗪𝗵𝘆 𝗱𝗼𝗲𝘀 𝘁𝗵𝗶𝘀 𝗺𝗮𝘁𝘁𝗲𝗿 𝗿𝗶𝗴𝗵𝘁 𝗻𝗼𝘄? Because the companies that move first from experimenting with agentic AI to building real governance, will set the pace for everyone else. Leaders who understand which trends to act on, and why, have the chance to set the standard for the next five years in data and AI. 𝗪𝗵𝗮𝘁’𝘀 𝗼𝗻𝗲 𝘁𝗿𝗲𝗻𝗱 𝘆𝗼𝘂 𝘁𝗵𝗶𝗻𝗸 𝗶𝘀 𝗯𝗲𝗶𝗻𝗴 𝗼𝘃𝗲𝗿𝗹𝗼𝗼𝗸𝗲𝗱?

92

2 Kommentare

Peeking at live A/B tests: when is it OK? A common misconception in experimentation is that peeking is never allowed. In a classical fixed-horizon test, you set the stopping rule upfront (e.g., 2 weeks or X users), end the test, and calculate a single p-value. For a fun story on how early versions of Optimizely (incorrectly) encouraged ending experiments early by computing p-values in near-real-time, see https://lnkd.in/gF9k3vBk. The reality: sometimes peeking is not only fine, but recommended. Here are some examples: 1.  You’re not going to take action. We are all curious how our experiment is doing, and peeking for that purpose is extremely common and perfectly fine. If the experiment looks great with the p-value below 0.005 for several days, you might start drafting the email summarizing the idea to be shared with the wider organization when it ends.  Maybe lookup good bakeries for where to order a celebratory cake. 2.  You’re going to abort the experiment if things look bad. You see a bug, or revenue is statistically significantly down, performance (page load time) is terrible: all good reasons to peek and abort. Peeking increases your false positives, or type-I errors, but that’s OK if you’re going to abort and fix a bug—no need to hurt users for two weeks when you can see things are terrible after one day. 3.  You planned to peek and designed your experiment for peeking using interim analysis techniques, such as group-sequential designs, or you’re using always-valid-p-values. This is covered in my advanced A/B course and recommended: https://lnkd.in/gU9xrezE Design the experiment and execute the plan! To learn more about A/B testing, join my Accelerating Innovation with A/B Testing course on Maven: https://bit.ly/ABClassRKLI

215

23 Kommentare

12 months ago, everyone was talking about prompt engineering. And don't get me wrong - it's still important. But here's what I've learned building Scalafai: GenAI is WAY more powerful when it already knows what you're working on with important details readily available. 🦾 Think about your typical GenAI workflow. You'd write a prompt in ChatGPT/Claude/Groq/etc. to help with a task, then switch between apps to copy/paste the info you need to insert critical context into your prompt. Sound familiar? 🤔 The real breakthrough isn't better prompts - it's giving AI the context it needs automatically. Think about. Any assistant, AI or human, can help you much more when they know: 1. the bigger picture of what you are trying to accomplish 2. the important contextual details needed to finish the task in a way that aligns with the bigger picture We built Scalafai around this fundamental concept. Instead of making people better prompt engineers, we made AI context-aware by connecting it to your existing communications data - emails, chats, meetings transcripts, files, calendars. The result? GenAI becomes incredibly user-friendly and actually useful for real work. When AI already knows your project goals, your team's progress, and the conversations happening across your apps, something magical happens. You can just ask "what's the latest?" and get exactly what you need. Literally. It's the shift from prompt engineering to context engineering. Anyone else tired of the copy-paste dance? 😅 Follow Scalafai for more insights on context-aware AI! #AI #Scalafai #ContextEngineering #PromptEngineering

270

38 Kommentare

Contrarian take: Every ETH L2 will become a ZK rollup. The early signs are already here. Here’s why: A couple of years ago, ZK tech simply wasn’t ready: • Proofs were expensive • Tooling was new and clunky • Scaling felt theoretical and unrealistic That’s why optimistic rollups became the shortcut. They gave Ethereum scalability when it was most needed. But shortcuts rarely last. Fast forward to today: • Proving costs have dropped by orders of magnitude • Proof systems are production-ready • Developer tooling has matured Now the balance has flipped. Because the long-term race isn’t about who can launch fastest. It’s about who can deliver security, finality, and seamless user experience at scale. And on every one of those dimensions, ZK rollups pull ahead. → 𝗙𝗶𝗻𝗮𝗹𝗶𝘁𝘆 𝗮𝗻𝗱 𝗲𝘅𝗶𝘁𝘀. Validity proofs remove 7-day challenge windows, unlocking near-instant withdrawals and safer bridging. → 𝗦𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝘆𝗼𝘂 𝗰𝗮𝗻 𝗿𝗲𝗮𝘀𝗼𝗻 𝗮𝗯𝗼𝘂𝘁. With ZK, the chain proves its state to Ethereum each batch, reducing reliance on watchers and governance “pauses.” → 𝗜𝗻𝘁𝗲𝗿𝗼𝗽 𝘁𝗵𝗮𝘁 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝘄𝗼𝗿𝗸𝘀. Shared validity proofs make native L2↔L2 messaging possible without trusted third parties, so apps compose across rollups. → 𝗖𝗼𝘀𝘁𝘀 𝗮𝗿𝗲 𝗼𝗻 𝗮 𝗼𝗻𝗲-𝘄𝗮𝘆 𝗰𝘂𝗿𝘃𝗲 𝗱𝗼𝘄𝗻. Proving that once cost hundreds of dollars per batch now costs a fraction of that. Hardware acceleration, recursion, and software breakthroughs keep driving costs lower quarter after quarter. → 𝗗𝗮𝘁𝗮 𝗴𝗼𝘁 𝗰𝗵𝗲𝗮𝗽𝗲𝗿 𝗽𝗼𝘀𝘁-𝗗𝗲𝗻𝗰𝘂𝗻. Blobs lowered posting costs for rollups, and ZK benefits most as proofs compress state changes efficiently. → 𝗗𝗲𝘃𝗲𝗹𝗼𝗽𝗲𝗿 𝗿𝗲𝗮𝗹𝗶𝘁𝘆. ZK-EVMs have matured enough for production, and the remaining gaps are shrinking every quarter. -- Optimistic stacks may point to fault proofs as their answer. But even they are already moving toward ZK. What this means next: → OP-style chains will add ZK paths for faster exits and stronger security. → Bridges will shift to proof-based routes instead of trusted custodians. → Sequencers will compete on speed, decentralization, and reliability. The direction of cryptography, user experience, and cost curves only points one way. Over time, every optimistic rollup will switch to ZK stack because everything that matters on Ethereum will settle with validity proofs. P.S. What do you think which major OP stack L2 will switch to ZK? ♻️ Repost this to help others in your network. 📌 And follow Aram Mughalyan for more content like this

85

76 Kommentare

How Stripe prevents double payments (explained in 2 mins or less): 1 They use idempotent APIs to prevent double payment. 2 They use UUID as the idempotency key. 3 They send the idempotency key through HTTP headers. 4 They create a new idempotency key whenever the request payload changes. 5 They store the idempotency keys in a database on the server side. 6 They query the database with keys to check if a request was processed earlier. 7 They process a request only if it's new and then store its idempotency key in the database. 8 They roll back a transaction using the ACID database on a server error. 9 They remove the idempotency keys from the database after 24 hours for reuse. 10 They use exponential back-off with jitter to avoid the thundering herd problem. What else would you add? —— 👋 PS - If you enjoyed this post, you'll love my detailed article about this case study (with visuals): → https://lnkd.in/erMkqwq4 ——— ♻ Repost to help others learn idempotent APIs. 👤 Follow Neo Kim for more.

2.546

139 Kommentare

Amazon's shocking AI reality check. Internal data reveals brutal truth. Amazon Q Developer generated just $16.3M ARR in its first year. Cursor hit $500M ARR in the same timeframe. That's a 30x difference. The most telling part? Amazon employees are demanding access to Cursor internally, calling Q Developer too slow compared to competitors. IEven CEO Andy Jassy praised Cursor as a major force behind the "explosion of coding agents" on earnings calls. AWS built its empire on infrastructure, but AI coding tools require a completely different approach Speed, intuition, and developer experience over enterprise features. The AI coding market will hit 40% of all new business software by 2028 according to Gartner. Amazon has the resources to catch up, but they're learning that in AI, being first to market with the right experience beats being biggest.

50

3 Kommentare