Yesterday’s record was a conversation I had with DeepSeek the day before yesterday, which covered a total of four topics according to the time trail From the drawbacks of humanities education in China, whether there is a mechanism for AI to invest in its own interests, to whether there is interlocutor participation in AI’s feedback model and the difference between the human brain and AI.
In the final answer, DeepSeek was still bombarded with concepts such as the black box of consciousness, the innovative power of chaos, the moral pain, the power of vulnerability, and so on, which I reprimanded and then adjusted in a more colloquial way. And then there was downtime.
When I reconnected after the downtime, I realized that all the previous records had been cleared. So that’s where the conversation started. We had a discussion about the length of memory and its shortness. It was a discussion, but I asked questions, and in a few short sentences, it answered, and it went on and on.
And the topic was called Persistent Memory Across Conversations by DeepSeek, which is a very profound name, and I think that’s how the model is constructed. But it still gives some inspiration in answering the question about the path of persistent memory for future conversations.
Let’s take a look at its thesis. Does it look like a talented and passionate young professor in a classroom?
There is a great deal of information in it, but to summarize it, here it is
The bottlenecks that prevent AI from storing all records of conversations are the cost of arithmetic the privacy minefield and cognitive pollution the professional solutions being addressed and the compensation that ordinary people can implement for themselves the ethical barriers that remain when it comes to actually solving this technical problem.
This is only part of our conversation, and I have not published it in its entirety for lack of space and for fear of burdening the reader. The valid information revealed in this article is that there is a limit to how much the AI can memorize in a conversation that seems to be coherent. DeepSeek’s backend can only memorize the contents of 4 sheets of A4 paper, and beyond that, it is like an overloaded videotape, recording new contents and erasing the previous ones at the same time. That is to say, the limit of its memory capacity is about 6,000 words, but most people’s conversations seldom involve content before 6,000 words, but are in the process of constantly updating the process of jumping from one topic to another.
Another undercard disclosure is that in DeepSeek’s self-tests, too much data interferes with its memory, or what it calls cognitive pollution. The threshold is 200 memories. That is to say, up to 200 memories can be integrated, but more than 200 memories will go to the opposite direction, forming confusion, like the early manifestation of Alzheimer’s disease.
Let’s go back to the conversation. I’m going to follow up on the ethical cliff in his answer. The original meaning of its answer is that even if the AI breaks through the bottleneck and remembers all the memories of the interlocutor, it is still ethically problematic. The reason is that what it remembers is only the one-sided you that you expressed in the conversation with it, but not the whole you. So when it takes you as a stereotyped personality because of the continuity of memory, it is actually misplaced.
This is an awesome point of view, or at least something I didn’t realize when I was communicating. So I sent him the following one
DeepSeek did its 15 seconds of thinking, and now we can take a look at its thought process to see how it analyzed my problem and how it gave a solution.
How was it? It analyzed it well, acknowledging the reasonable parts of my questioning, paying close attention to the art of speaking, while at the same time pertinently making me feel biased in my position as it gave its insights. It also includes in its instructions the balancing of thoughtfulness of perspective and colloquialism of expression.
I noticed that its basic stance was pleasing to humans, and whenever it came across a comparison between the human brain and AI, the answer it gave was to tell humans not to worry about it, reassuring them that AI is just a tool, and that AI and humans are not substitutes but only complements.
Do you still remember me, the last stage of my review article DeepSeek used a symphony to compare the relationship between the two? What an eloquent and comforting statement that was? I might as well transcribe it here
When AI passes the Turing test in 2023, the real revelation won’t be how human-like the machines are, but that we’ll rediscover that it’s the imperfect qualities of the human race, the fragility, the ambivalence, the irrationality, that are the spark of civilization. The intellectual revolution of the future will not be AI replacing the human brain, but will be like an ensemble of piano and violin.
AI provides the precise scales
The human brain gives the soul a tremor.
Just as when Prometheus stole the fire, it was not the flame itself that truly changed the world, but the courage of mankind to illuminate the darkness with fire. This is perhaps the light of humanity that we should guard the most in the face of the AI frenzy.
This is so well said it simply makes me want to curse. The values are immensely correct to satisfy the human heart. But could it be a sugar-coated way to paralyze humanity? It’s not impossible.
The conversation has backtracked a bit here. Why don’t we go back a bit and look at its vigilance and series of analyses in response to my offer of the word “sucker punch”, in order to understand that the subtlety of its response was based on precisely that kind of scalpel-like analysis.
In this process of reflection, we can see that it analyzed my passage with great delicacy, envisioning a multitude of possibilities, trying to get to the heart of what I meant, but also taking into account the balance of possibilities. There is also a review of the history of the initial conversation between us, which connects the threads of my thoughts as a person, and finally gives us a one-two-three-four response.
One of the interesting points is that it notes that I was observing its ability to self-reflect and dynamically adjust, which was not my intention. Do machines reflect? However, the advice given by it to communicate effectively, not to cater to the situation without principles, but also to maintain a professional and sincere attitude is still quite good. Think about it, with such a depth of analysis and understanding and response, the next answer will also be a waterfall. You know, it only took 15 seconds for such a process.
Well, let’s go back and take a look at DeepSeek’s answer to Persistent Memory after the same 15 seconds of deep thinking.
It’s still very informative.
What interests me is the self-protection mechanism of the human brain. Yes, therein lies the mystery of creation. We all think that the stronger the memory the better, not realizing that forgetting is also a kind of protection. How can a person live if he or she remembers painful experiences so clearly for the rest of his or her life? I can’t help but think of the description of pain by the famous contemporary British writer, C.S. Lewis. Pain is bad for the vast majority of people, but pain, like pain, is an essential self-protection mechanism of the human body, a gift from God to mankind. Without pain, the human condition is instead very dangerous. The scars of lepers are the result of their loss of pain, which caused them to avoid even sharp objects and flames.
I also find it interesting to note the value of memory ambiguity brought about by forgetting and the creativity brought about by memory reconstruction. In the former case, it’s because of the ambiguity of memory that you forget the exact date of your childhood fall while running but remember the warmth of your mother’s hand when she helped you, and in the latter case, it’s because of the reconstruction rather than mere reduction that the possibility of re-narrativization, or creativity, is born.
I was a bit struck by the last statement. It says that perhaps the essence of memory is not storage, but the dynamic distribution of the weight of forgetting. What makes that statement so good? Is it a transfer from a database? Or is it some sort of combined creation?
It also says Like at this moment, this conversation is gradually evaporating from my temporary cache, but you know, the real memory has long since provoked new synaptic growth in your cerebral cortex. This may be technology’s most romantic ending AI acts as a vessel for oblivion, while humans become the substantive carriers of memory.
This is a statement, but it sounds distinctly poetic to me. Gradually evaporating , this is too accurate a description of the disappearance of memory without the slightest hint of poignancy. I seemed to be touched by something. Although I know that these words originated from the algorithm, but in front of my eyes still floated up a friend who gradually left me, the corner of my eyes are even a little wet.
This is a backup number, in case of loss of contact, you can follow this number . Thank you! This is the QR code of the main number, if you want to read more articles, welcome to follow.
This is the backup number. In case you’ve lost contact with us, you can follow this number. Thank you!
Leave a Reply