1. Introduction
As I kept using the same conversation with dots on Windows, the response of on-screen operations gradually became slower. This article is a record of that observation.
By "dots," I mean my personal AI agent that runs continuously in OpenAI's environment and keeps working on the tasks I give it.
What this article deals with is latency in the UI itself: typing, rendering, and switching between conversations. dots is an always-on agent that runs on the cloud side, and the conversation itself proceeds outside my local PC. The time the AI takes to generate a response, or the time dots takes to finish a task I asked for, is a separate matter from how sluggish the screen in front of me feels. When I say "heavy" in this article, I mean the latter: the UI operation latency.
2. Symptoms
The operations that became slow were the following:
- The time between typing a character and seeing it appear on screen
- The response to send actions and button presses
- Scrolling and screen refreshes
- Loading and switching between past conversations
At times, each operation felt like it took about 1–3 seconds (these are impressions, not measurements). The longer I kept using the same conversation, the heavier it tended to get. In a newly started regular chat, by contrast, I did not feel this slowness. Once it became considerably faster again, but later the problem returned.
3. Environment
- The machine is a Windows PC with 32 GB of memory
- The latency remained in the same dots conversation, both in the app and in the browser
- On the same PC, Notepad and newly started regular chats felt fast
That said, the fact that Notepad and new regular chats felt fast does not rule out a problem with the PC as a whole.
I do not have the exact version numbers of the app or OS I was using. I also did not measure the latency quantitatively, and the cause has not been identified.
4. What I tried
I tried the following, but could not confirm any lasting improvement:
- Display-related measures: I tried changing settings inside the app, but could not confirm any lasting improvement. On the Windows side, I confirmed that the transparency and animation effects were Off (their state before that check is unknown)
- Cleanup such as organizing tabs and closing the browser: memory usage went down, but the latency in the same conversation remained
- Closing, reopening, and restarting the app: it did become considerably faster once in the morning, but that was after a Windows Update and a restart, and before I changed any display settings. The effect of another restart in the afternoon has not been confirmed
- Opening the same dots in a different browser: I did not feel any major improvement
Since I tried these in combination, I cannot separate out how effective each individual measure was. The temporary improvement that day happened before the display-related measures, so it cannot be credited to them.
Separately, entering text into the conversation from my iPhone felt faster, so I am considering an arrangement where I work on the PC but split off just the conversation input to the iPhone. This is based on a comparative impression; I have not measured any improvement.
5. Current status
In the public Issues, there are multiple similar reports of operations becoming heavy within the same conversation or during long sessions. The related public information is as follows:
- Issue list: https://github.com/openai/codex/issues
- Issue #38547: https://github.com/openai/codex/issues/38547 — A contributor explained that a specific contributing cause had been addressed in a particular version, and the issue is now Closed
- Issue #20214: https://github.com/openai/codex/issues/20214 — A contributor explained a fix for one contributing factor in slowdowns during long sessions. Since then, there have been posts reporting both improvement and persistence
- PR #40672: https://github.com/openai/codex/pull/40672 — A related code change has been merged into main
However, each of these is an explanation of work on an individual cause; none of them means that all symptoms are fixed, and they are not a guarantee by the company. Nor is it certain that a merged change has reached every distribution channel.
As for my own case, I reported it to OpenAI support and received confirmation that the report was received. I have not received a reply, a root-cause determination, or a fix timeline, and the matter is currently at the stage of having asked for their handling. I also do not know whether my case shares a cause with any of the public Issues.
6. Summary
- Continuing the same conversation with dots on Windows left latency in typing, rendering, and conversation switching
- I tried settings changes, cleanup, restarts, and more, but could not confirm any lasting improvement
- The cause has not been identified, and I am at the stage of having reported it to support and asked for their handling
Top comments (1)
화면 조작 지연과 모델의 응답 생성 시간을 분리하고, 여러 조치를 함께 바꿔 원인을 특정할 수 없다고 적은 부분이 좋네요. 다음 재현에서는 같은 짧은 입력·스크롤·대화 전환을 새 대화와 긴 대화에서 각각 측정하고, 앱·OS 버전과 작업 시각도 함께 남기면 비교가 쉬울 것 같습니다. 변경이 병합됐다는 사실과 실제 배포판에서 증상이 나아졌다는 확인을 구분한 점도 중요한 기록 방식이라고 생각해요.