【文章标题】：The Frontier AEO Tracker: What Astra Chooses (and every other frontier model, and what you can do about it)
前沿AEO追踪器：Astra的选择（以及其他前沿模型的选择，以及你能做些什么）
【文章正文】：
Naive
天真地
autoresearch investment
自动化研究投资
in our AEO have yielded impressive ROI, and so naturally it was time to take it seriously. We were inspired by
在我们的AEO中取得了令人印象深刻的投资回报率，因此自然到了认真对待的时候。我们受到了
What Claude Code Actually Chooses
《Claude Code实际选择什么》
, and decided to extend/adjust it to our tastes.
的启发，并决定根据我们的喜好进行扩展/调整。
After
经过
a
数
few billion tokens
十亿token
of prototyping, aligning, and scaling
的原型设计、对齐和扩展
pipelines, here’s
流程，以下是
the
我们的
Latent Space Frontier AEO tracker
Latent Space前沿AEO追踪器
. Our
。我们的
methodology
方法
extends
扩展了
AmplifyingAI’s
AmplifyingAI的方法
to run 6 prompt variations over 7
在7个
models
模型
1
1
(search on) in 161
上运行6种提示变体，覆盖161个
categories
类别
, from
，从
coding agents
编码代理
to
到
AI podcasts
AI播客
to
到
AI Sandboxes
AI沙盒
to
到
Managed Databases
托管数据库
to
到
ASR models
ASR模型
to even oddball categories like
甚至包括一些古怪的类别，如
Angel investors
天使投资人
and
和
Corporate spend
企业支出
and
以及
Payroll software
薪资软件
.
。
Answer extraction was done by Astra, and scored for a
答案提取由Astra完成，并根据
proprietary AEO score
专有的AEO评分
that gives weight to
进行评分，该评分重视
first choices, alternative choices
首选、备选
,
，
mentions
提及
, but
，但也
also negative weights
对温和和强烈的反推荐给予负权重
to mild and strong anti-recommendations (which are rare, but do happen). Because we know you’ll want it, we also extracted the top cited
（这种情况很少见，但确实会发生）。因为我们知道你会需要，我们还提取了影响Agent推荐的最常引用的
sources
来源
, as well as an analysis of
，以及对
top failures
主要失败
的分析。
Basic Results
基本结果
Here are the most dominant products (in their categories) in the world:
以下是全球（在其类别中）最具主导地位的产品：
There are some familiar names in there — opening up the natural question of contamination, which we have checked. Since we have nothing to hide,
其中有一些熟悉的名字——这自然引发了关于污染的问题，我们已经检查过。因为我们没有什么可隐瞒的，
every prompt and answer pair
每个提示和答案对
is inspectable.
都是可检查的。
However,
然而，
bias
偏见
does exist - when models are asked for coding agent recommendations, Fable/Opus like
确实存在——当模型被要求推荐编码代理时，Fable/Opus喜欢
Claude Code
Claude Code
and Sol/Astra like
而Sol/Astra喜欢
Codex
Codex
and Grok loves
Grok喜欢
Cursor
Cursor
and Muse loves
Muse喜欢
Muse Code
Muse Code
and SWE-1.7 loves
SWE-1.7喜欢
Devin
Devin
and so on. I wonder why. You can see other “
等等。我想知道为什么。你还可以看到其他“
soft biases
软偏见
” emerge too…
”出现……
That said there are notable examples of GPT models recommending Claude, a laudable nonbias:
也就是说，也有一些值得注意的例子，比如GPT模型推荐Claude，这是一种值得称赞的无偏见行为：
There are
有
28 categories
28个类别
(out of our total 161) which have a universally dominant primary choice - among all surveyed frontier models.
（在我们总共161个类别中）在所有调查的前沿模型中有一个普遍主导的首选。
There are a lot more
还有更多
“close contests” and “always the vibesmaid, never the vibe”
“势均力敌的竞争”和“总是伴娘，从未新娘”
categories which should be key AEO battlegrounds.
的类别，这些应该是AEO的关键战场。
Sol vs Astra, Opus vs Fable
Sol vs Astra，Opus vs Fable
New pretrains for new model classes represents a new opportunity to check in on what the labs are moving towards in their data and RL priorities, and to check in on whether startups’ investments in AEO are paying off. We prepared special reports analyzing our rankings, observing
新模型类别的新预训练代表了一个新的机会，可以检查实验室在其数据和强化学习优先级上的动向，并检查初创公司在AEO上的投资是否得到了回报。我们准备了特别报告分析我们的排名，观察到
VERY consequential flips in model choices
模型选择中非常重大的翻转
between model generations from the same lab.
来自同一实验室的模型世代之间。
We have separate
我们有单独的
Opus→Fable
Opus→Fable
and
和
Sol→Astra
Sol→Astra
summary pages. For some flips, we highlighted a neutral analysis of what competitors did better in each scenario.
摘要页面。对于一些翻转，我们强调了竞争对手在每个场景中做得更好的中性分析。
Efficiency vs Confidence, and Recommendation Sourcing
效率 vs 信心，以及推荐来源
One of our most surprising findings between Sol→Astra and Opus→Fable is that Anthropic seems to be biasing their models to searching more sources (Sol median of 9 sources, vs Astra median of 5, vs Opus median of 11 sources, vs Fable of 15). Astra seems to be just generally a lot more “
我们在Sol→Astra和Opus→Fable之间最令人惊讶的发现之一是，Anthropic似乎偏向于让他们的模型搜索更多的来源（Sol的中位数是9个来源，Astra是5个，Opus是11个，Fable是15个）。Astra似乎总体上更“
confident
有信心
”, or “
”，或者“
efficient
高效
”, depending how you look at it - Astra is FAR less likely to change its mind when you lightly paraphrase your question. This makes the
”，取决于你怎么看——当你稍微改写问题时，Astra改变主意的可能性要小得多。这使得
value of AEO
AEO的价值
itself rise as choice randomness declines.
随着选择随机性的下降而上升。
Sources analysis
来源分析
also somewhat strongly predicts what the labs do prioritize vs don’t.
也在一定程度上强烈预测了实验室优先考虑什么和不优先考虑什么。
However the sample size is small here and only represents what we can scrape from attempted toolcalls, not the pretrain dataset. What we CAN validate is that AEO practices measured by
然而，这里的样本量很小，仅代表我们可以从尝试的工具调用中抓取的内容，而不是预训练数据集。我们可以验证的是，由
Ora and Vercel
Ora和Vercel
, like
测量的AEO实践，如
markdown content-negotiation
Markdown内容协商
, are real and failures discourage models from reading your content.
是真实的，失败会阻止模型读取你的内容。
Just for fun
只是为了好玩
Here are the top
以下是顶级
Angels
天使投资人
in the world according to LLMs (some dedupes left to do…).
根据LLMs的推荐（还有一些去重工作要做……）。
See more
查看更多
We also made a little
我们还制作了一个小小的
family feud type game
家庭问答类游戏
where you can see if your priors align with the data. Fun!
你可以看看你的先验是否与数据一致。有趣！
We are open to further suggestions and business enquiries to develop this if it is of interest. Ping
我们欢迎进一步的建议和商业咨询来开发这个项目，如果有兴趣的话。请联系
@latentspacepod
@latentspacepod
or email
或发邮件至
business@latent.space
business@latent.space
(we have a business manager now! woo!)
（我们现在有业务经理了！哇！）
1
1
As we note in our methodology post, we did try VERY hard to include Gemini/Antigravity, GLM/Zcode, and DeepSeek/DeepCode, but errors and rate limits made them untenable to include in this first run analysis. Please let us know how to raise limits if you represent these companies.
正如我们在方法文章中提到的，我们确实非常努力地尝试包括Gemini/Antigravity、GLM/Zcode和DeepSeek/DeepCode，但错误和速率限制使它们无法包含在第一次运行分析中。如果你代表这些公司，请告诉我们如何提高限制。