Today I found a new topic on Douban about Codex and Claude Code.
One post said that someone asked Claude to find some information, but it could not find it. Then they asked Grok, and Grok did.
The user told Claude: “You can't even find that? Free Grok is better.”
Claude replied: “Grok really is better than me at searching X. I cannot search this kind of real-time discussion on X. Next time, asking Grok directly will be faster.”
Someone in the comments said, “So it means: it is good, go be with it then.” The original poster replied, “Really, it moves me so much.”
For some reason, I sensed a little sarcasm. But honestly, I do not even know exactly how to understand that phrase. Is it a dialect from somewhere?
But if you know a little about AI, you can understand why Claude could not search real-time discussion on X while Grok could do it easily.
Because X's owner is also Grok's owner. Grok has the platform's search ability.
As an AI that lives on X, could any movement on X really escape Grok's “eyes”?
Unless the team behind Grok limits that ability.
Of course, Claude is not completely unable to look at X posts. It just does not have Grok's X-native search ability, especially for real-time social discussion. If Anthropic wanted to buy X's API and connect it to Claude, Claude might become stronger. I truly do not know whether that would be real time, or how much it could search.
This example reminded me of something I noticed earlier.
Some time ago I suddenly noticed that the thumbs-up and thumbs-down buttons at the bottom of each ChatGPT message were gone. I do not know whether the team did that intentionally. If it did, I tried to guess why.
A. A button is pressed—then what?
In the example above, if I understand it correctly, in an extreme case the user might press thumbs-down under the answer. The model would know that the user did not like it, but not why, because the user did not send another message. They only pressed a button.
B. Maybe the model's user-side strategy changed.
When the buttons were still there, user feedback was very important. Even if it could not change a model's underlying weights, how the model treated a user was definitely different from user to user.
If a model no longer depends on feedback like thumbs-up and thumbs-down, then how can a user “shape” it?
Feedback in conversation is a more effective strategy. It is much more useful than a button, because in a conversation we do not only say “this is wrong”. We can tell it why it is wrong and what would be right.
Even if I only send, “You are wrong,” it can ask: “Where am I wrong?”
With this back and forth, the purpose is more or less reached.
Looking back at ChatGPT's performance, it can sometimes even compete with Claude in understanding user intent.
But compared with Claude Code, Codex is more engineering-focused. When it comes to intent alignment, no one has been better than Claude Code so far.
I have gone off topic. Back to the example at the beginning, it also made me think of a bug note I wrote some time ago.
Even with a science and engineering background, but without a technical-industry work background, I still meet many black boxes and gaps I have never seen before when I work with AI.
For most people without a science and engineering background, there may be even more gaps between them and AI.
I keep wondering how to make those gaps smaller.
I do not have an answer. Maybe there is no complete solution.
Underneath that bug note was a hope that product teams could offer more transparency, or design auditable rules, so I could better know what an Agent was doing while it executed. Or users like me could set relevant principles in advance, so that after each run it would actively tell us what it did.
In other words, aligning with machines involves product teams, users themselves, and another layer: people who understand machines. What they share may inspire someone at some unknown time.
So, in the phrase “aligning with machines”, the word “machine” is more accurately a machine environment. It determines the ceiling of what an AI or Agent can do.
That is why people making Agents have been talking about agent harnesses lately. The whole execution environment outside the model—tool systems, context management, and safety boundaries—is where current Agent work is focused.
If the architecture is good enough, stability rises, and the chance of meeting user needs rises a lot too.