Dinghao Li

𝗟𝗶𝘃𝗲𝗢𝗽𝘀 𝗶𝘀𝗻’𝘁 𝗮 𝗰𝗮𝗹𝗲𝗻𝗱𝗮𝗿. 𝗜𝘁’𝘀 𝘁𝗵𝗲 𝗼𝗽𝗲𝗿𝗮𝘁𝗶𝗻𝗴 𝘀𝘆𝘀𝘁𝗲𝗺 𝗼𝗳 𝗴𝗿𝗼𝘄𝘁𝗵. 𝗣𝗮𝗿𝘁 𝟭: 𝗧𝗵𝗲 𝗙𝗿𝗮𝗺𝗲𝘄𝗼𝗿𝗸 Most studios treat LiveOps like seasonal wallpaper. 🎃 in October. ❄️ in December. Maybe a summer sale. That’s decoration. Not growth. 𝗪𝗵𝗮𝘁 𝗶𝘁 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝗶𝘀: infrastructure that wires 𝗔𝗦𝗢 → 𝗨𝗔 → 𝗣𝗿𝗼𝗱𝘂𝗰𝘁 → 𝗥𝗲𝘃𝗲𝗻𝘂𝗲 into one loop. 𝗪𝗵𝗮𝘁 𝘁𝗵𝗲 𝗶𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲 𝗹𝗼𝗼𝗸𝘀 𝗹𝗶𝗸𝗲 (𝗻𝗼𝘁 𝘁𝗵𝗲𝗼𝗿𝘆): 𝟭) 𝗦𝗵𝗮𝗿𝗲𝗱 𝘀𝗰𝗼𝗿𝗲𝗯𝗼𝗮𝗿𝗱 (𝗼𝗻𝗲 𝘀𝗲𝘁 𝗼𝗳 𝗞𝗣𝗜𝘀): - Leading: Event impressions (store), CPP/CSL CVR, ad CTR/IPM, session starts - Lagging: D1/D7 retention deltas, ARPPU, ROAS/LTV velocity by cohort, revenue/DAU. If teams don’t stare at the same numbers, they won’t pull the same way. 𝟮) 𝗢𝗻𝗲 𝘀𝗶𝗻𝗴𝗹𝗲-𝘁𝗵𝗿𝗲𝗮𝗱𝗲𝗱 𝗼𝘄𝗻𝗲𝗿 (𝗚𝗿𝗼𝘄𝘁𝗵 𝗣𝗠): Decision rights over the 𝗲𝘃𝗲𝗻𝘁 → 𝘀𝘁𝗼𝗿𝗲 → 𝗨𝗔 chain. No tug-of-war. No “please” meetings. 𝟯) 𝗧𝗵𝗲 𝘁𝗵𝗿𝗲𝗲 𝗰𝗮𝗱𝗲𝗻𝗰𝗲𝘀:   • 𝗪𝗲𝗲𝗸𝗹𝘆 “𝗚𝗿𝗼𝘄𝘁𝗵 𝗛𝗲𝗮𝗿𝘁𝗯𝗲𝗮𝘁” (𝟯𝟬 𝗺𝗶𝗻): next event, CPP/CSL mapping, creative angles, status.   • 𝗠𝗼𝗻𝘁𝗵𝗹𝘆 𝗿𝗲𝘁𝗿𝗼 (𝟲𝟬 𝗺𝗶𝗻): cohort outcomes, what bent the curve, what didn’t.   • 𝗤𝘂𝗮𝗿𝘁𝗲𝗿𝗹𝘆 𝗿𝗼𝗮𝗱𝗺𝗮𝗽 (𝟵𝟬 𝗺𝗶𝗻): themes, offer slots, tech asks, kill/scale calls. 𝟰) 𝗧𝗵𝗲 𝗲𝘃𝗲𝗻𝘁 𝗯𝗿𝗶𝗲𝗳 (𝗼𝗻𝗲 𝗽𝗮𝗴𝗲, 𝗻𝗼 𝗻𝗼𝘃𝗲𝗹𝘀):   • 𝗣𝗿𝗼𝗺𝗶𝘀𝗲: what players get (theme/value).   • 𝗖𝗼𝗵𝗼𝗿𝘁𝘀: who we target (new vs returning vs payer).   • 𝗦𝘁𝗼𝗿𝗲: CPP/CSL IDs, assets, copy, locales.   • 𝗨𝗔: angles to ship, by network.   • 𝗢𝗳𝗳𝗲𝗿: what’s claimable 𝗻𝗼𝘄, what’s gated for later.   • 𝗠𝗲𝗮𝘀𝘂𝗿𝗲𝗺𝗲𝗻𝘁: success threshold + rollback rule. If the brief can’t fit on one page, you’re building a feature, not running LiveOps. 𝟱) 𝗠𝗶𝗻𝗶𝗺𝗮𝗹 𝗽𝗹𝘂𝗺𝗯𝗶𝗻𝗴 (𝘁𝗵𝗲 “𝗵𝗼𝘄” 𝗲𝘃𝗲𝗿𝘆𝗼𝗻𝗲 𝘀𝗸𝗶𝗽𝘀):   • 𝗗𝗲𝗲𝗽 𝗹𝗶𝗻𝗸𝘀/routing so cohorts land in the right welcome state.   • 𝗢𝗳𝗳𝗲𝗿 𝗳𝗹𝗮𝗴𝘀 in the client (toggle without a build).   • 𝗔𝘁𝘁𝗿𝗶𝗯𝘂𝘁𝗶𝗼𝗻 𝘁𝗮𝗴𝘀 that survive to day-7 (creative → CPP/CSL → event). Yes, this is a real engineering lift. But it's a one-time cost to build a permanent capability. You're building switches you can flip forever. ——— If your LiveOps plan is still pumpkins and Santa hats, you don’t have a growth system. You have arts & crafts with analytics. This is the “what”. Next post, I’ll share the "how".

5

3 Comments

20,000+ purchases in 90 days 📈💸 For NetEase’s Fragpunk launch, our goal wasn’t just installs. It was purchase velocity. We built a high-intent funnel and optimized every step until purchases scaled predictably. Here’s what actually moved the needle on Fragpunk’s launch: 🔹 We followed buying signals, not clicks. Mapped high-intent behaviors, built regional cohorts, and shifted budget to the audiences most likely to purchase. 🔹 We tested fast, killed fast. Weekly promote / iterate / kill loops on trailers, creator cuts, and statics, judged on purchase rate. 🔹 We met players where they are. Platform-specific flows for PC, Xbox and PlayStation, syncing creator traffic with paid to shorten time to purchase. 🔹 We wired the data end to end. Tracking with the client, steam funnel tuned, live dashboards to keep spend inside cpp/roas guardrails. 🔹 We kept it in-house. One team owning brand safety, fraud controls, and transparent reporting. Curious which lever would move purchase velocity for your next launch? Drop your biggest challenge below and we’ll share the playbook we used. Read the full case study:  https://lnkd.in/ddJgnqeX #Gamebassadors #GameLaunch #UserAcquisition #GrowthMarketing #InfluencerMarketing #MobileGames #CaseStudy

7

That's awful: NetEase will stop funding Nagoshi Studio in May 2026, after learning they need at least $44M of additional funding to finish Gang of Dragon: "Nagoshi is trying to find new sponsors without success so far". Source: https://lnkd.in/edN9xr7D

132

27 Comments

𝗜’𝗺 𝗿𝗲𝗹𝗲𝗮𝘀𝗶𝗻𝗴 𝘁𝗵𝗲 𝗳𝗶𝗿𝘀𝘁 𝗵𝗶𝗴𝗵𝗹𝘆-𝗰𝘂𝗿𝗮𝘁𝗲𝗱, 𝗵𝗶𝗴𝗵-𝗾𝘂𝗮𝗹𝗶𝘁𝘆 𝘀𝘆𝗻𝘁𝗵𝗲𝘁𝗶𝗰 𝘀𝘂𝗽𝗲𝗿𝘃𝗶𝘀𝗲𝗱 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗱𝗮𝘁𝗮𝘀𝗲𝘁𝘀 𝗳𝗼𝗿 𝗥𝗼𝗯𝗹𝗼𝘅, 𝗮𝗻𝗱 𝗶𝘁’𝘀 𝗼𝗽𝗲𝗻-𝘀𝗼𝘂𝗿𝗰𝗲𝗱. 𝗜𝘁 𝗰𝗼𝘂𝗹𝗱 𝗱𝗿𝗮𝘀𝘁𝗶𝗰𝗮𝗹𝗹𝘆 𝗶𝗺𝗽𝗿𝗼𝘃𝗲 𝗳𝗿𝗼𝗻𝘁𝗶𝗲𝗿 𝗺𝗼𝗱𝗲𝗹𝘀 𝗯𝘆 𝟭𝟬𝟬𝗫. For months, I’ve been building SuperbulletAI, and I was able to understand how top models behave inside Roblox Studio. And honestly? I found a major problem: ❗ Even frontier models (GPT-5, Claude, etc.) are inconsistent when generating Roblox code compared to JavaScript/TypeScript. They hallucinate API usage, miss edge cases, break Studio workflows, and often misuse core Luau patterns. Why? Because they were trained mostly on public code, old repos, and outdated examples without any supervision. That’s not enough anymore. So I’m releasing something new: 𝗦𝗬𝗡𝗧𝗛𝗘𝗧𝗜𝗖 𝗦𝗨𝗣𝗘𝗥𝗩𝗜𝗦𝗘𝗗 𝗗𝗔𝗧𝗔𝗦𝗘𝗧𝗦 𝗙𝗢𝗖𝗨𝗦𝗘𝗗 𝗢𝗡 𝗥𝗢𝗕𝗟𝗢𝗫’𝗦 𝗜𝗡𝗧𝗥𝗜𝗖𝗔𝗧𝗘 𝗔𝗣𝗜 𝗦𝗘𝗥𝗩𝗜𝗖𝗘𝗦 + 𝗖𝗟𝗔𝗦𝗦𝗘𝗦 Think of it as a fully structured, deeply supervised, massively scalable dataset series built to improve: - Frontier LLMs - Open-source coding models - Roblox’s own AI assistants - And even SuperbulletFrameworkV1’s ecosystem This dataset will be categorized in two clean ways: ✔ By API service (CollectionService, PathfindingService, RunService, ContextActionService, etc.) ✔ By instance class (Humanoid, Players, RemoteEvents, BaseParts, Constraints, etc.) Every example is synthetic, uses modern patterns, and reflects the real quirks and bugs of Roblox Studio so models learn how to work around them, just like top scripters do. Only ~20% is human-reviewed, because the pipeline is automated. So I’m inviting top Roblox scripters, engineers, and technical creators in my network to help refine and review them. 𝗧𝗵𝗶𝘀 𝗶𝘀 𝗼𝗻𝗹𝘆 𝘁𝗵𝗲 𝗯𝗲𝗴𝗶𝗻𝗻𝗶𝗻𝗴. The first release: ~5M tokens The long-term roadmap: billions of tokens, fully supervised, continuously updated. It will also include: - A Python repo that compiles everything for easy training - A user-friendly fast Dataset Viewer for quick inspection - Modern Luau practices (no wait(), no old pairs/ipairs, no deprecated patterns) - A highly organized project structure built for clarity, scalability, and easy modification, letting you update or expand specific training datasets by API service or instance class Improving it doesn’t require magic, it requires better data. This is my contribution to push the next generation of AI coding models forward as the founder of the first and strongest Roblox AI game builder. Once I launched this, I'll be open-sourcing the heavily modernized expansion of the outdated 25M-token Roblox in-house official Luau-Corpus training datasets. Roblox’s ecosystem is massive, fast-moving, and painfully under-represented in modern LLM training. 𝗧𝗵𝗶𝘀 𝗶𝘀 𝗵𝗼𝘄 𝘄𝗲 𝗽𝘂𝘀𝗵 𝗥𝗼𝗯𝗹𝗼𝘅 + 𝗔𝗜 𝘁𝗼 𝗶𝘁𝘀 𝗻𝗲𝘅𝘁 𝗲𝘃𝗼𝗹𝘂𝘁𝗶𝗼𝗻.

14

3 Comments

Microsoft should spin off Xbox—before it’s too late. If Bloomberg reporting is accurate about the fourth round of layoffs in the last 18 months, its time to face reality, Xbox has lost this generation: Game Pass failed to hit scale. Hardware share is collapsing. Software share is worse. And the strategic value to Microsoft? Almost nonexistent. Let’s break it down: 🎯Game Pass missed the mark. Still under 40M subs—nowhere near the 100M target. Subscriptions don’t fit gaming like they do TV. Most gamers play a few select titles. Heavy users subscribe, then buy fewer games. That’s not sustainable. 🎯Hardware share is cratering. By end of 2025, Xbox is projected to hold <30% of the console market vs. Sony. And by cycle end, Sony will have sold 4x as many consoles. 🎯 Software is the real red flag. Microsoft’s share of console software vs Sony is expected to fall to just 14% in 2025—thanks to subscription cannibalization and declining 3rd-party sales. That’s disastrous, given the margins on software royalties. 🎯Cloud gaming? It’s a feature on the box, not a strategy. Not expanding the market. Not removing friction. 🎯Cross-platform promise broken. Xbox on mobile makes for nice PR—but it’s not what gamers actually care about. 🎯AI won’t save Xbox. Gaming doesn’t move Microsoft’s AI agenda. And the suggestion of using Xbox player data to train models is a privacy minefield—not to mention, there’s no evidence they’re doing it. Satya Nadella has likely already asked the question: Why are we still in gaming? The solution is clear: Spin off Xbox. 🎯 Let it become an independent company built to compete head-on with Sony and Nintendo. 🔄 Sunset Game Pass in favor of a lighter, online-focused sub. 🎮 Recommit to exclusives for majority of franchises. Offer differentiated Xbox experiences—especially for big brands like Call of Duty. 🔧 Kill Series S. Focus on efficient, scalable first party hardware and license to 3rd parties. 📈 Go public. A standalone Xbox would be another major, pure-play gaming company on public markets in the West. 🧠 Bring in real leadership. People like Frank Gibeau or Owen Mahoney who know how to run big public companies and know gaming. Spinning off Xbox frees it from Microsoft’s broader tech and AI agenda—enabling faster, more gaming focused decision-making. It could even pursue bold M&A moves. The alternative—keeping Xbox at Microsoft—avoids the short-term pain of a major reorg and strategic pivot, but it also guarantees a slow, quiet decline and a slow destruction of teams and franchises. It trades disruption for complacency, and that path leads straight to death by a thousand cuts. #Xbox #GamingStrategy #Microsoft #SpinOff #GamingIndustry #MergersAndAcquisitions #Innovation

244

94 Comments

Want to learn more about Netflix culture? See this fantastic episode of Pragmatic Engineer podcast, featuring Netflix CTO Elizabeth Stone. The episode covers key pillars of Netflix engineering culture, breadth of Tech work happening at Netflix and how we learn and iterate as a growing company. Especially excited about seeing Live used to demonstrate culture of rapid innovation, self-reflection and distributed ownership within Netflix engineering. Thanks Gergely Orosz for another great episode! #netflix #engineering #innovation #live

69

2 Comments