Max Meiden Dasuki

AI is getting smarter. But most teams are still missing the point. When I was at ByteDance, we built internal AI tools to move faster: Auto-generate copy. Draft strategy. Predict winning creatives. It worked.  +70% NPS $6.4M saved in operational costs. But here’s one thing. You can build the best AI tools, and still miss what people actually need. Because AI can map behavior, but it can’t feel hesitation. It can scan data, but it doesn’t know why someone smiled at an ad.  It can write a love letter but it has never had its heart broken. So we had to shift focus. Not just for speed, but to sense. Because productivity doesn’t always mean progress. Features don’t always equal understanding. And outputs don’t equal insight. Even now at Atria, we use AI to help brands create faster than ever. But the judgement call? The taste? The “this feels right”? That still needs a human. And follow me Ray Jang for daily posts as I scale Atria, an AI platform for marketers to create killer ads with ease.

150

142 Comments

Stripe is building a way for developers to completely automate billing for tokens: 1/ pick model (e.g. openai o3, claude sonnet 4, gemini 1.5 flash, etc) 2/ set margin (e.g. charge +10% over inference) 3/ add two lines of code to your app It works great with subscription models too: charge X dollars per month for N tokens, then, after that, pass along token cost with a Y% markup on overage. “Intelligence-based pricing” Want to try it? [email protected]

168

10 Comments

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 Comments

When we started the Frameworks for Growth interview series, I wanted to have real conversations with the people building and backing today’s fastest-growing companies—and extract the frameworks that make growth repeatable, not lucky. Over the first five episodes, I got to learn from: * Garry Tan, who reminded us that persistence is underrated (especially in an era of swooping revenue growth curves?) * Daniela Amodei, who shared how trust can coexist with velocity and what it takes to build enduring culture when the world is watching every move you make * Amjad Masad who built Replit by betting that the next billion developers won’t look like the last * Bret Taylor, who explained why AI is still in its Yahoo! Directory phase * Steffen Tjerrild, who gave the case for building a world-class AI company in Europe Building is always hard, AI is changing everything, and the frameworks for how we decide, prioritize, and scale matter more than ever. Huge thanks to these founders for their candor, and to everyone who’s listened, shared, and debated along the way.

143

7 Comments

ByteDance just open-sourced a 36B parameter model with 512K context. Everyone's talking about the technical specs. They're missing the real story. This isn't about model performance. It's about geopolitical AI supply chains. During a private dinner last week, a private equity investor shared something that changed how I think about enterprise AI procurement. "We can't build our portfolio companies' entire AI strategy on models controlled by three American companies. Our LPs are asking hard questions about vendor concentration risk." ByteDance's move solves a problem most AI founders don't even know exists. Enterprise buyers are terrified of AI vendor lock-in. Not because of costs. Because of control. Here's what this release actually signals: Chinese AI companies are building enterprise trust through radical transparency. Apache-2.0 licensing eliminates the biggest enterprise objection: "What happens if they change the terms?" 512K context windows make enterprise document processing actually viable. Zero API fees remove budget approval friction entirely. Most American AI companies are still selling access. Smart enterprises want ownership. When you can download, modify, and deploy a model that rivals GPT-4 performance without ongoing licensing fees, the procurement conversation completely changes. The enterprise AI market just shifted. It's no longer about who has the best model. It's about who offers the most strategic flexibility. Three questions every AI founder should ask: Are you selling technology or selling control? If ByteDance offers free commercial licensing, how does your pricing model compete? What happens to your enterprise pipeline when procurement teams discover they don't need API dependencies? Enterprise deals stalling because buyers fear vendor lock-in? StratNorth helps AI companies position for procurement reality. Let's connect.

174

74 Comments

BREAKING: 86.8% SWE-bench Verified, Biggest Leap Ever. "SWE Bench verified has been the Super Bowl of AI coding benchmarks for the past two years. The last time the coding benchmarks had a double digit leap was March of 2024. This ushered in a huge era for AI coding. The AI agent was here." -Brian Elliott We just proved the AI pundits wrong at Google's Sunnyvale campus. While everyone's saying "AI acceleration is plateauing**,** we took a different approach at Blitzy. The results? 86.8% on SWE-Bench verified - the industry's premier AI coding benchmark that tests systems on real-world GitHub issues - delivering a 10% lift over the current leader and the first double-digit leap since March 2024. Here's what changed the game: --> System 2 AI: Trading time for quality. Our platform spends hours mapping every line of code relationally and hierarchically summarizing entire codebases (100M+ lines). --> Multi-agent orchestration: Extended inference time compute with validation across multiple models with built-in prompt engineering. As I told the audience: "The next frontier of progress is not just pre-training models. By driving more inference time compute into your systems, you are able to drive up the quality of output dramatically." Link to white paper in comments. #blitzy #ainativesdlc #autonomoussoftwaredevelopment #aicoding #nextfrontier

262

39 Comments

Oracle just joined NVIDIA, Google, and Airbnb in switching to front-loaded vesting. Is this the new standard? This is one of the biggest changes to tech compensation in decades. Here's what's happening: 1/ More companies are front-loading equity → Oracle, NVIDIA, Google, Airbnb, DoorDash, Uber, Pinterest → Instead of 25/25/25/25, it's now 40/30/20/10 or similar splits 2/ Why now? → Cuts long-term equity costs → Makes financials more attractive to investors → Gives candidates big Year 1 perceived value → Ties refreshers to performance 3/ What it means for you: → Don't just look at Year 1 ask about refresher grants → Compare steady-state vesting (Years 2–4), not just the front-load → Be ready for more performance-driven equity culture Front-loaded vesting is spreading fast across major tech companies. Expect mid-to-large tech companies to adopt it next. Would you rather have more equity upfront or stability spread evenly across 4 years?

88

11 Comments