Hear the trend before it has a name.

Your mission: spot what players are starting to talk about.

Diimsum reads what players say on QQ groups, Discord, Bilibili, Weibo, Steam, TapTap and Bahamut, and turns it into insight for game studios. The question every studio asks us is simple: what's changing in our community right now? This puzzle asks you to build something that answers it.

The catch is that nobody knows the topics in advance. There's no fixed list to sort messages into, and the thing that matters most next month usually doesn't have a name yet. Players invent the words as they go: 坐牢, 刮痧, a nickname for a character, an in-joke that started in one group chat.

So your solution has to invent its own categories as the conversation unfolds, then tell a studio which ones are growing, which are a one-day storm, and which aren't real at all.

This is a puzzle about understanding players, not about statistics. The data has no likes, scores or ratings, only what people wrote. If you know how players think, how gacha economies feel from the inside, why a patch day sounds different from a slow-building grievance, or what a group chat sounds like when it's joking, you'll get further than someone with a bigger model.

An example, from a different game.

Here's the kind of result we're hoping for, using an imaginary farming game called 《云海牧歌》. The puzzle data hides different stories, but the shape of a good answer is the same.

Messages like these trickle in…

There's no shared keyword. "肝", "手都酸了", "打工鱼" and "累死" are all the same complaint: a daily chore that's turned into a job. A game player recognises that straight away. Your code has to as well. The 「确实」 only counts because of what it's replying to.

…your solution groups them into categories it invented…

For example: Daily grind and chores → Fishing minigame too demanding. A good set of categories reflects how players actually think about a game (economy, fairness, effort, social life, the characters they love), not just which words appear.

…and notices how it's moving.

Compare a one-off, like a server outage that's loud for two days and then gone, or something that isn't real, like fifty brand-new accounts posting the same praise at the same minute.

Your run.py writes it out, using only what it could see on that day:

{
  "as_of": "2026-03-12",
  "trends": [{
    "title": "钓鱼小游戏太肝",
    "title_en": "Fishing minigame feels like a chore",
    "category": "Daily grind & chores",
    "status": "growing",
    "first_seen": "2026-03-02",
    "voice_ids": ["0196f3a2-…", "0196f3b8-…", "…"],
    "why": "Started in TapTap reviews, reached QQ and Weibo within a week; players now joke about it (打工鱼), which usually means it's sticking."
  }]
}

status is one of new, growing, one-off, fading or not real. Everything else (titles, categories, your reasoning) is yours to decide.

And a person could act on it:

What to send us

1. Working code. We'll run your solution like this, at several dates we won't tell you in advance:

   python run.py --input voices.jsonl --as-of 2026-07-24 --out trends.json

   It may only look at messages created on or before --as-of. That's how we know your solution would have spotted things at the time, not with hindsight. Any language model, library or approach is welcome. If it calls an LLM, it can use OPENAI_API_KEY or TYPESAFE_API_KEY (Jev), and we'll supply our own when we run it.
2. The output above: up to 8 trends per run. We don't need to share your category names. We read the voice_ids, so pick the messages that best show each trend.
3. A short README (English or Chinese) explaining your idea: how your solution decides what belongs together, how it tells something new from something growing, a one-off or something fake, and what you tried that didn't work. Then tell the studio, in a few paragraphs, what's going on in 《星环守望》 by the end of August.
4. Warm-up. Explain these to a colleague who doesn't read Chinese: what each one literally says, what it actually means, and how the player feels.
   策划是懂玩家的，懂得怎么悄悄改概率
   这波属于是逼氪了，懂的都懂
   艾琳=铁锅，新人别懵
   白嫖党表示毫无压力 [doge]
   非酋报道，160抽一个金都没有 绷不住了

No LLM access? You don't need your own. The keys.env file attached to your invitation has capped OpenAI and Jev keys for this puzzle, plus a quick guide to calling Jev. They stop working after 20 October. If a key stops working, email careers@diimsum.com for a test key of your own.

Not everything that gets louder is a trend.

There's no single right method. A simple idea that really understands players will beat a complicated one that doesn't. Surprise us.

Players who think hard.

Open to students and graduates of any year and any subject. You'll need native-level Mandarin and enough coding to make an idea run. Knowing games, especially Chinese gacha and live-service games, from the inside is a real advantage here.

Rules

• Submissions close 20 October 2026, 23:59 Hong Kong time.
• Solo entries only. Use any tools you like, including AI assistants, and tell us how in your README.
• Send a zip or GitHub link with your code, a requirements file, your README and the warm-up.
• Please don't publish your solution while the puzzle is open.