Claude 4.5 Is Here, But I Care More About What They Didn't Say

Yesterday morning I woke up and checked my phone to see news of Claude 4.5’s release.

Honestly, my first reaction wasn’t “finally” but rather “what are they hiding this time.”

Anthropic’s official blog is beautifully written — 30% performance improvement, faster inference, 15% cost reduction. The data looks great, the comparison charts are professional. But after reading the entire article, you’ll notice an interesting phenomenon: it hardly mentions any technical details.

Where does the training data come from? What’s the parameter scale? What new architecture designs were used? None of these core details were mentioned.

My personal feeling is that this “results without process” release style has become standard for closed-source large models. OpenAI does this, Anthropic does this, Google’s Gemini does this — they give you a bunch of benchmark data, but never tell you “how they did it.”

This reminds me of a conversation last week with a friend doing AI research. He said something quite sharp: “Closed-source large models are like black box magic. You see the rabbit come out of the hat, but never know what mechanism is inside the hat.”

But to be fair, there are reasons for being closed-source.

From a business perspective, technical details are core competitive advantages. If Anthropic公开训练数据和架构设计，竞争对手（包括开源社区）就能快速复现，那他们的商业护城河就没了。这在商业上是合理的。

From a technical evolution perspective, however, this “black box competition” is actually quite wasteful.

Everyone is exploring similar directions, everyone is stepping on similar pitfalls, but because of mutual secrecy, the entire industry’s technical accumulation efficiency is low. It’s like 5 people solving the same math problem simultaneously, each deriving from scratch instead of sharing insights.

This reminds me of the period after the 2017 Transformer paper was published — Google开源了整个架构，结果整个AI社区在3年内实现了爆发式增长。BERT、GPT、T5……所有这些突破，都建立在Transformer这个公开的基础之上。

如果当时Google把Transformer闭源了，现在的AI会是什么样？可能GPT还在实验室里憋着，BERT根本不会出现。

所以我对Claude 4.5的态度是——性能提升是好事，但别把”黑箱竞争”当成理所当然。

闭源可以理解，但至少应该在技术路线图、安全性评估、社会影响这些层面更透明一点。不是所有东西都要开源，但至少要让用户和研究者知道”你用了什么数据””你的能力边界在哪””你可能存在什么风险”。

说真的，如果让我选，我更愿意用那些技术细节公开的模型——哪怕性能稍微差一点。因为至少我知道它是什么，能做什么，不能做什么。这种”可控性”，在长期使用中比单纯的性能更重要。

当然，这只是我个人的倾向。对于普通用户来说，可能只关心”好不好用”，不关心”怎么做到的”。

最后留个问题：你觉得大模型厂商应该公开核心技术细节，还是保持闭源保护商业利益？如果你是Anthropic的技术负责人，你会怎么选？

我猜很多人会说”当然开源好”，但如果你站在商业公司的立场，答案可能没那么简单。