Changkun Ou

Researcher. Engineer. HCI/CG/ML/PL. | |

[Changkun Ou @ouchangkun]

AI acceleration makes human-in-the-loop more important, not less. The bottleneck shifts from generation to verification: can humans understand, trust, and responsibly act on what the system produces?

[Anthropic @AnthropicAI]

Our internal data shows Claude is accelerating AI development—a possible path to recursive self-improvement, or AI autonomously building a more capable successor.

It’s happening faster than we thought, and the implications deserve greater attention. https://t.co/OVVPJO7VQx

[Anthropic @AnthropicAI]

Our internal data shows Claude is accelerating AI development—a possible path to recursive self-improvement, or AI autonomously building a more capable successor.

It’s happening faster than we thought, and the implications deserve greater attention. https://t.co/OVVPJO7VQx

[Changkun Ou @ouchangkun]

@OpenAI nice! why can't pro users use this?

[Changkun Ou @ouchangkun]

what is claude doing hhere? @bcherny

[Changkun Ou @ouchangkun]

@antonosika there will always be people who wants to delegate more about decisions made behind a platform, but also people who prefer transparency around it, not necessarily always inspecting but presence. this types of selection has to be designed well so that user trusts remain in the loop

[Changkun Ou @ouchangkun]

@diptanu try https://t.co/ApE6xExCt5

[Changkun Ou @ouchangkun]

@bcherny do you agree?

[Changkun Ou @ouchangkun]

Claude Opus 4.7 is a horrible model. Its performance is significantly worse than the Claude Code experience with 4.6. It shows terribly narrow-minded behavior, poor intent prediction, and often fails to align with the user’s actual intent unless explicitly corrected.

[Changkun Ou @ouchangkun]

Instead, it relies too much on its own phrasing, such as “verified,” and similar expressions. The model also tends to over-prioritize verification, but the verification often goes sideways, burning more tokens without properly completing the task.

[Changkun Ou @ouchangkun]

@Lovable 😄 the ad would be a lot cooler if the guy can just talk to the phone, instead of typing.

[Changkun Ou @ouchangkun]

@oran_ge https://t.co/oiwf7Tf0h7

[Changkun Ou @ouchangkun]

Am I the only person who thinks it is worse than 4.7? I committed few hundreds of commits over the weekend, it a lot more narrow minded (aka aligned) but I spend more effort to steer it towards correct. Bug fix back and forth.

[Theo - t3.gg @theo]

How are people feeling about opus 4.7 so far?

[Changkun Ou @ouchangkun]

@MaiYangAI 最近产生了一些不同的看法，我对未来的构想和文中说的完全不一样，既不是基于现在的工作流，也不是纯AI first

[Changkun Ou @ouchangkun]

I agree majority thoughts here. However, this is not surprising and we all know structure limits the growth and people cognition. Transformation means a complete rethink of a known structure as well. The fundamental assumption is: will the core stablize?

[Peter Pang @intuitiveml]

https://t.co/oWr2pTk4Hu

[Changkun Ou @ouchangkun]

@diptanu I think the choice of where to run harness is different by whether builder owns the harness implementation. The core problem is where do you draw the boundary of harnesses. This is tight to the model capability which is forever shifting based on the data and training.

[Garry Tan @garrytan]

New item in my SOUL md tonight

[Jürgen Schmidhuber @SchmidhuberAI]

Neural Computers
https://t.co/XVk9bCLGbm

[Changkun Ou @ouchangkun]

@computesdk @vercel Hey, can you benchmark https://t.co/PWXML2ldSs?

[Changkun Ou @ouchangkun]

@MaiYangAI 想好 prompt，schedule，睡觉