I am not sure whether something changed in the GPT team's UI strategy, but I want to try to understand this from my own perspective.
First, why did these two buttons exist? It is simple: to collect user feedback, which could then affect the model's behavior and improve the user experience.
But in reality, not all feedback collected through these buttons is necessarily good for the health of the model itself. Sometimes, user feedback really should not be adopted.
For example, a user might click thumbs-down because they are in a bad mood or feel offended by the model. Or they might click thumbs-up because the model praised them in exactly the right way. From the perspective of the immediate user experience, adjusting the model based on this feedback is a plus. But in the long run, it may not be a good thing, because data from the user side can also become one source of model-training data. Which data can be used and which cannot? And when large amounts of unusable data pile up... just thinking about it makes my head hurt.
So my guess is that this may be about lowering the weight of data based on user feedback to make the quality of reinforcement learning more stable.
There is another point. Recent AI products have been fighting hard over context engineering. That may mean that continuously collecting, understanding, and perhaps even applying a user's context is what can create a better experience. I recently recovered an old ChatGPT account. At first I worried that, since I had not used it for almost a year and my interests, focus, and work had changed a lot during that time, it might not be able to follow my topics when I switched back.
Then I discovered that I had worried for nothing.
When I started the first new conversation, I asked my current account to give me some context and brought it into the old account. In the very first deep conversation, the old account aligned with me almost perfectly. The experience barely felt different from my current account. It was a real surprise. Occasionally it did connect something to an old memory from that account, because there really was a relationship. But I would tell it that I had already pivoted and that it could forget that line of context.
So how do you do context engineering well? It does not rely only on system architecture. It also needs the user's participation—for example, the real-time feedback I described above, followed by real-time correction from the model. ChatGPT does this very well. Especially after 5.3 launched, my positive feelings about the experience shot straight up.
There are a few related things I find interesting and want to mention.
1) Claude's dreaming-related design is worth learning from. The memory mechanism does not need to depend on the user being online in order to work. I feel that organizing context while the user is away gives the product a clearer “mind” and can also give the user a better experience.
2) SecondMe has a feature where, whenever I come online, a prompt floats up to remind me what topics I might share publicly. It is really telling us that our thinking and our interactions with it do not have to remain known only to the two of us. Perhaps they can be shared. This pushes us to leave the product and get feedback in the real world, even if that only means posting on social media.
3) Memory decay. We always say that a model's memory should learn from how human memory works: some things are forgotten, while memories mentioned often become deeper. Adding decay to memory can also reduce the burden on the product's brain.
4) How should memory be stored? It is definitely not as simple as recording events. It involves concepts from cognitive and behavioral science, and I am still exploring the specifics. But recently I discovered that a system I built can remember across conversations. When it refers to a memory from an earlier conversation, it follows naturally from what is currently being said. That surprised me a lot.
Everything above is speculation and extension from an outsider with no professional experience in either the AI industry or product work. Its accuracy still needs to be examined. This piece is only a record.