Andy Ye @creatstar

Still thinking about our PhoenixAI happy hour at Ula 🐦‍🔥 Thank you to everyone who came out for great data + AI conversations, new connections, and Phoenix Margaritas made just for the event 🍹 Congrats again to our AirPods raffle winners 🎧 See you at the next one!

keyboard_arrow_left Previous keyboard_arrow_right Next

Sharing a few recent thoughts on AI. 1. At this stage, what AI brings to people is increasingly anxiety rather than excitement. The pace of progress is simply too fast! What changes in a single week today might have taken a whole year in the past. Imagine suddenly having a team of highly capable, tireless members who can work 24 hours a day. You would immediately start worrying about how to make this team do more. Because if you don’t, someone else will. The FOMO only grows stronger. A friend of mine told me that if he doesn’t assign his AI agent a “big task” before going to bed, he can hardly fall asleep. 2. AI is forcing people to do what they least want to do: deep thinking. Execution work is no longer a competitive advantage. AI can do it faster and better than you. You can no longer make a living by execution alone. Instead, you must think deeply about how to direct AI toward more meaningful work. But here comes the paradox: many forms of deep thinking depend on firsthand execution experience. If you have never done the work yourself, how do you gain the intuition and judgment required for deep thinking? How can a programmer who has never written functions truly design modules, let alone an entire system? 3. Looking back at the technological transformations from my childhood to now, I feel both excited and deeply shaken. AI will dramatically widen the gap between people. Steve Jobs once said that in the software industry, one outstanding person can be worth a hundred average ones. AI may expand that gap by another hundred times, meaning that one person who can truly master AI could be equivalent to ten thousand ordinary people. And this gap will not be limited to the software industry. As a father, I have very mixed emotions about my children being born into such an era. “It was the best of times, it was the worst of times.”

I just finished reading the “State of AI” report jointly published by a16z and OpenRouter, which is based on an analysis of over 100 trillion tokens of real-world interaction data. It’s extremely impressive. Below are some key takeaways. 1. The Industry Turning Point & The Rise of "Reasoning" * The o1 Release: The release of OpenAI's o1 reasoning model on December 5, 2024, is viewed as a pivotal moment. It marked a shift in the field from "single-pass pattern generation" to "multi-step deliberation" inference. * Agentic Inference: Users are increasingly utilizing models as components within larger automated systems (Agents). 2. Open Source vs. Closed Source Models * The Rise of Open Source: While closed (proprietary) models still dominate, Open Source (OSS) usage has grown steadily, accounting for approximately 1/3 of total token volume by late 2025. * Contribution from Chinese Models: Open-source models developed in China (e.g., DeepSeek, Qwen, Moonshot/Kimi) have seen significant growth and now occupy a substantial share of the OSS ecosystem. * Dynamic Competition: There is no single hegemon in the open-source space. The landscape is highly dynamic; no single model consistently holds more than 20-25% of the share. New releases (like DeepSeek V3, GPT OSS) rapidly capture market share. 3. Evolution of Model Sizes: "Medium" is the New Small * Decline of "Small" Models: Models with fewer than 15B parameters are seeing their market share decline despite their high quantity. * Rise of "Medium" Models: The 15B-70B parameter range (e.g., Qwen2.5 Coder 32B, Mistral Small 3) has found "Model-Market Fit," striking a balance between performance and efficiency. * Diversification of "Large" Models: The 70B+ category shows a pluralistic competitive landscape where users do not converge on a single model. 4. Key Use Cases * Programming: This is the primary driver of prompt token growth. Coding tasks often involve extremely long contexts (frequently exceeding 20k tokens) for code understanding and debugging. * Roleplay: This remains a stronghold for open-source models, which often outperform more restricted closed-source models in creative tasks. 5. Other Key Findings * Global Usage: Over 50% of usage on OpenRouter originates from outside the United States, highlighting the globalization of AI adoption. * Retention (The "Cinderella Effect"): The report identifies a "Cinderella 'Glass Slipper' effect." This refers to early user cohorts who found a perfect fit between the model and their specific needs; these users exhibit much higher long-term retention compared to later users. Original report: openrouter.ai/state-of-ai

🎥 Watch on-demand: youtu.be/tktN3gjDoYk?si… Missed our “CelerData Cloud BYOC: New Features, Smarter Scaling for Real-time Analytics” webinar — or want to give it another watch? The full recording is now available!

How do you power dashboards that join over 100 tables and scan 20B rows—yet still respond in under two seconds? That was the challenge facing Celonis. See how they re-engineered their analytics architecture to achieve sub-second process mining performance. hubs.la/Q03Wqkmp0

CelerData Newsletter of November 2025! linkedin.com/pulse/november…

A German 🇩🇪 philosopher once said, “We learn from history that we learn nothing from history.” I’m often reminded of this line when I see senior leaders on social media sharing insights that I already figured out 15 years ago as a junior engineer.

🚀 Turning the Promise of the Data Lakehouse into Reality — Together with Hewlett Packard Enterprise Many organizations adopt data lakehouses to unify their analytics, but they often struggle with complexity — from fragmented data integration and inconsistent performance to governance and operational challenges. That’s why CelerData and HPE have partnered to deliver a truly modern, high-performance lakehouse architecture. 💡 HPE Alletra Storage MP X10000 provides cloud-native, all-flash storage with massive scalability and enterprise-grade reliability — the foundation for real-time analytics at any scale. ⚡ CelerData Enterprise, powered by StarRocks, brings blazing-fast query performance, elastic scalability, and seamless integration with open formats like Apache Iceberg — enabling complex analytics directly on large datasets, with no extra pipelines or overhead. Together, HPE and CelerData simplify the path to real-time, unified insights, helping organizations make faster, smarter decisions with greater confidence and lower cost. Please visit this link to learn more: lnkd.in/gb3D7Saf hashtag#DataLakehouse hashtag#RealTimeAnalytics hashtag#HPE hashtag#CelerData hashtag#StarRocks hashtag#ApacheIceberg hashtag#AIAnalytics

🌍 That’s a wrap on #StarRocksSummit2025! Thank you to everyone who registered, tuned in, and supported us — you made this milestone possible! Session replays are starting to roll out! Catch (or rewatch) them now — and check back soon for more: 🔗 app.events.ringcentral.com/events/starroc…

🌟 Speaker Highlight: Join Nicholas Reich of Demandbase at #StarRocksSummit2025 (Sept 10, free & virtual) as he shares how his team eliminated denormalization with StarRocks’ high-performance joins to power real-time analytics! 👉 Agenda + free pass: hubs.la/Q03GCVVj0

🌞Countdown check — 8 days left! At #StarRocksSummit2025, our customer keynote features Intuit — hear how they run real-time analytics with StarRocks at the core. 👉See the full agenda + grab your free pass if you haven’t already: hubs.la/Q03Gzd1p0

🔥 The wait is over! The full agenda for #StarRocksSummit2025 is live — Sept 10 | Free & Virtual 🔗 summit.starrocks.io/2025/Twitter 👉 20+ engineer-led talks on what actually works at scale in production: real-time customer-facing analytics, the lakehouse, and AI. ✅ RSVP today!

Demandbase, a B2B GTM leader, faced JOIN inefficiencies & scaling issues with ClickHouse. By switching to StarRocks-powered CelerData Cloud, they cut storage costs by 90%, reduced hardware by 60%, and optimized ETL pipelines for faster analytics.👇 hubs.la/Q03C8pvf0

Recently, some friends from Japan have been asking me questions about StarRocks. In fact, our Japanese documentation has been live for two months! 🇯🇵 For our friends in Japan, you can now read StarRocks docs directly in Japanese: 🔗 docs.starrocks.io/ja/docs/introd… We’d love to hear your feedback on the Japanese docs—your input will help us make them even better! 皆様、いつもありがとうございます！

StarRocks reached 10K ⭐ on GitHub! 🥳 From everyone building, using, and championing StarRocks: THANK YOU 💙 One milestone down—many more ahead!

⚡ Supercharging Customer-Facing Analytics with Delta Kernel + @StarRocksLabs Delta Lake’s Delta Kernel offers a robust set of Engine APIs to make data access faster and more efficient. StarRocks taps into these APIs to power smart caching techniques that: 🔹 Eliminate redundant reads of Parquet and JSON files 🔹 Slash query latency for real-time user-facing dashboards 🔹 Improve resource usage with intelligent cache control With tools like 𝗖𝗮𝗰𝗵𝗲𝗱𝗣𝗮𝗿𝗾𝘂𝗲𝘁𝗛𝗮𝗻𝗱𝗹𝗲𝗿 and 𝗖𝗮𝗰𝗵𝗲𝗱𝗝𝘀𝗼𝗻𝗛𝗮𝗻𝗱𝗹𝗲𝗿, StarRocks boosts performance and delivers snappy queries—perfect for high-concurrency scenarios. 👉 Dive into the blog for the full breakdown: delta.io/blog/starrocks… #deltalake #opensource #starrocks #linuxfoundation #APIs #analytics #dataengineering

Check out this behind-the-scenes look at how Zepto—one of India’s fastest-growing quick commerce companies—scaled its real-time brand analytics with StarRocks, evolving from a lean MVP to powering external dashboards with over 300 million rows! hubs.la/Q03v5slR0

🎟️lu.ma/95a5qys1 - The Iceberg Meetup lands in NYC on July 10, featuring 15+ talks from engineers, architects, and open-source contributors shaping the future of the lakehouse. Don’t miss our session on achieving sub-second latency on Iceberg with materialized views!

keyboard_arrow_left Previous keyboard_arrow_right Next