gg2

How gg2 works

Last reviewed September 2026

Every gg2 card is written by AI from reporting published that day. This page is the full account of how that happens, including the parts where the method is weak. We think AI writing is only defensible if it shows its work, so nothing here is held back.

Step by step

  1. Picking the stories. gg2 watches what is rising across Google Trends, Reddit, Hacker News and major outlets. Stories are not chosen editorially, and gg2 commissions no reporting of its own.
  2. Finding same-day material. For each story, gg2 searches reporting published that same day. When a story arrives with a known source article and the search happens to miss it, gg2 pulls that article’s text directly so the original account is in the material.
  3. Building a fact list before anything is written. A model reads the retrieved material and first produces a structured fact list: a dated timeline, who said what, background, whatever the sources themselves flag as unverified, and related events the sources explicitly mention. Every line carries the numbered source it came from. The writing step never sees raw search results without this pass.
  4. Checking the material is actually on topic. The same step judges whether the sources are about this story at all. If every one of them turns out to be about something else, gg2 treats the search as having found nothing, rather than writing a card from material that does not match.
  5. Writing the card. A second pass turns the fact list into the card you read: the one-line gist, the timeline, what different sides are saying, what is still unverified, and the sources.
  6. Checking the finished card back against the material. A separate verification pass reads each claim in the card against the retrieved material and marks it supported, partial, or absent. Claims marked absent, meaning no basis at all, are struck out. The check is deliberately conservative: when it cannot tell, it keeps the claim, so it only removes things with zero support. If the one-line gist itself comes back unsupported, only that line is rewritten rather than regenerating the whole card, so the rest of the card does not drift.
  7. Safety. Three layers, in order: a keyword prefilter, the model’s own refusal on topics it should not decode, and a content-safety check on the finished text. In production a hit blocks the card, and a failure of the safety service is treated as a block rather than as a pass.
  8. Pictures. When a card carries an image it comes from the story’s own coverage and is credited. Images go through the same safety check. gg2 does not use AI-generated images, because a synthetic picture of a real event is the one thing a reader has no way to check.

When the material is thin

If the search comes back with nothing relevant, gg2 publishes a reduced card rather than filling the gap from what a model happens to remember. A thinner card is the honest outcome of a thin search. This is also why some cards carry fewer sections than others.

Which stories get a public page

Every card is readable in the app. A story only gets a page listed in our sitemap and pushed to search engines when it clears a quality bar: it needs a gist, at least three timeline entries, and at least two source links. Cards that do not clear it stay in the app and out of the index. We would rather publish fewer pages than pad an index with thin ones.

Corrections

If a card gets something wrong, email support@gg2.guru with the link.

Live cards are regenerated periodically as a story develops, so the card you report may change on its own. Tell us anyway. A recurring error usually means the instructions or one of the guards needs fixing, which is a different repair from correcting a single card. Pages for stories that have rotated out are kept as they were.

Where this method is weak

Three limits worth knowing, stated plainly:

What gg2 is not

gg2 is not original reporting, and it is not a replacement for the outlets it reads. It is a layer on top of their work that answers the same questions every time. The sources are listed on every card precisely so you can leave gg2 and go read them.

gg2 是怎么做的

最后核对:2026 年 9 月

gg2 的每一张卡都由 AI 依据当天发出的报道写成。这一页把这件事是怎么发生的完整交代出来,包括方法本身薄弱的地方。我们认为 AI 写作只有把过程摊开才站得住,所以这里没有保留。

一步一步

  1. 选题。gg2 盯着 Google Trends、Reddit、Hacker News 与主要媒体上正在走高的话题。选题不由编辑决定,gg2 自己也不采写任何报道。
  2. 找当天的材料。对每个话题,gg2 去搜当天发出的报道。如果这条话题带着一篇已知的源文而搜索恰好漏掉了它,gg2 会直接把那篇的正文取回来,保证一手叙述在材料里。
  3. 先列事实清单,再动笔。模型先读完检索到的材料,产出一份结构化的事实清单:带日期的时间线、谁说了什么、背景、来源自己标注为未证实的说法、以及来源明确提到的相关事件。每一行都标着它出自第几条来源。写卡那一步永远不会绕过这一遍、直接看原始搜索结果。
  4. 核对材料是不是真的在说这件事。同一步还会判断这些来源与话题是否相关。如果每一条都在讲别的事,gg2 按「什么都没搜到」处理,而不是拿不对版的材料硬写。
  5. 写卡。第二遍把事实清单加工成你看到的卡:一句话看懂、来龙去脉、各方怎么说、待核实、来源。
  6. 把写完的卡拿回材料里核一遍。一个独立的核验环节逐条把卡里的说法对回检索材料,判为有据、部分有据、或全无依据。判为全无依据的直接剔除。这个核验刻意偏保守:拿不准就保留,只打击零依据的那些。如果连「一句话看懂」本身都被判无据,只重写那一行,而不是整张卡重掷,这样卡的其余部分不会跟着漂移。
  7. 安全。三层,依次是:关键词粗筛、模型自身对不该解读的话题的拒答、以及对成稿文本的内容安全检查。生产环境下命中即拦截;安全服务本身故障时按拦截处理,而不是放行。
  8. 配图。卡上如果有图,它取自这件事自己的报道并标注出处,并且经过同一道安全检查。gg2 不使用 AI 生成的图片,因为一张真实事件的合成图恰恰是读者最无从核实的东西。

材料不足时

如果搜索没拿回相关材料,gg2 会发一张内容更少的卡,而不是用模型记忆去填补空缺。材料薄,卡就该薄,这是诚实的结果。这也是有些卡比别的卡少几块内容的原因。

哪些话题会有公开页面

每张卡在 App 里都能读。但一个话题只有过了质量门才会进我们的 sitemap、被推给搜索引擎:需要有一句话看懂、至少三条时间线、至少两条来源外链。过不了的留在 App 里,不进索引。我们宁可少发页面,也不拿薄页面把索引撑起来。

纠错

如果某张卡写错了,把链接发到 support@gg2.guru

现役卡会随事件发展定期重新生成,所以你报的那张卡可能自己就变了。还是请告诉我们。反复出现的错误通常意味着要修的是指令或某一道守卫,那和改一张卡是两回事。已经轮出的话题,其页面保持原样不再改动。

这套方法薄弱在哪

三条限制,直说:

gg2 不是什么

gg2 不是原创报道,也不是它所阅读的那些媒体的替代品。它是架在那些工作之上的一层,每次回答同样的几个问题。每张卡都列出来源,正是为了让你能离开 gg2 去读它们。