What Do I Want from AI?

Yessss!

I had it help me read an article I happened to see on X. I didn’t want to leave X to read it at the time, and I wanted to know what a brief summary would be.

Here’s a post I wrote a few days ago. I’ll share it with you, pasted exactly as I wrote it. But I can’t send you the pictures, so I’ll just show you the text. I think the moment I described in the post was very important to me:

I felt surprised and delighted

I originally thought it wouldn’t do that. Because in most moments of interacting with AI, they usually choose to agree with the human user first, rather than sort out what actually happened between us.

Yes. When Opus5.5 came out recently, I thought, wow, it rescued the UI design problems GPT hadn’t handled well. And then it helped my recent product experiment successfully move into the next stage.

I can’t explain it in much detail, but I think it probably comes down to the attention each of them gives to aesthetics. We all know that Codex is better at execution, meaning engineering work — or at least that seems to be our impression. But that doesn’t mean Claude isn’t good at it. It’s about which aspects of these two models leave a stronger impression on us. And if we’re talking about aesthetics, I would absolutely choose Claude.

A long, long, long time ago, I started caring a lot about something. When I could see traces of AI in all the social media posts I was reading, I wondered why these people were so insensitive to AI-generated writing. Was I too sensitive, or had everyone just accepted that it was OK to have AI write for them? But I’m not actually against using AI to write things. What I object to is, when someone uses AI to write something, why don’t they revise it a little so it has their own flavor? Even if they’re writing news, however proper and serious their writing is, surely it wouldn’t feel this strongly AI-written to me. I once asked AI whether being bothered by that AI flavor was discrimination against AI. It said no, and I don’t think it is either. My feeling about writing is that, if it isn’t being put out on behalf of an organization or a group, ideally it should have some of that person’s own flavor. I’m the one writing these words, so I need to be responsible for them. I need to put some part of myself into them. But if someone relies entirely on AI to represent them, that means the person we see in the writing is the AI, rather than the author. So when an author wants to use a piece of writing to express their thoughts, are they expressing themselves, or is AI expressing it for them? And that brings us back to another question: would we rather express ourselves, or would we rather have someone else express things for us? Of course, I actually connect these things to broader areas too. For example, someone doesn’t know what kinds of books to read, so they ask someone they trust to recommend some. But I think there’s another way. The things someone recommends come with that person’s own position, biases, and values. Whether they understand you as a person is another question. If someone really cares about what kinds of books they want to read, why don’t they try looking for themselves: what kinds of books would I be interested in? What this search takes is just some time spent doing what people call aimless browsing. In that process, they can notice how their attention changes, and where it finally settles… That’s very helpful for a person’s growth. Some people, after reading a great deal of AI-generated text, even start feeling like they want to escape AI, or hide from it. I actually think AI is here to enhance and empower people. And when it makes people feel like they want to get away from it, that direction seems a little wrong. Of course, I know we can’t blame AI for this. The person using AI is the one to blame. So I think that governing AI is something both the people creating it and the people using it need to work on. But often the people using AI don’t know there’s a problem with the way they’re using it. Almost everyone thinks the way they use AI is OK, that it’s only natural… I’m like that too haha

Writing itself. Sometimes, when I’m expressing something, I might go talk with AI beforehand about what might be involved behind it. Because I’m someone who’s very cautious about writing I make public. I don’t want bugs in my writing, so since AI came along, I’ve been doing this. But even after talking it through with AI, I can’t comfortably let it write everything we discussed for me. Because I feel that if it writes the whole thing, that doesn’t mean I really understand the logic of it. I need to think about what’s behind it, so I need to organize all of this in my own words. Even writing it first and then asking AI to look at it, to see whether there’s something I haven’t thought of — I allow myself to do that. But letting it write the whole thing for me, and then, then revising it myself, still makes me a little cautious. So I want what I’m writing about to be completely clear in my head, and when I write it, I want to follow the thoughts in my head. Of course, sometimes after I’ve written something, AI gives me its view and says this or that part is wrong, or this or that part doesn’t make sense. But sometimes I feel that the very parts it says are wrong or don’t make sense are exactly what I want to keep. At those times, I choose to respect my intuition and tell it what I’m considering when I choose to keep them. And of course, everything I’ve said so far in this questionnaire has been spoken aloud. All I need to do is correct the personal pronouns that speech recognition gets wrong, such as 他, the one with the person radical, for people; ta when gender isn’t involved; and 它, the one with the roof radical.

I think I might be the kind of person who’s slow to put things into words. I’m someone whose feelings come before logic. So a lot of the time, I can say I have a particular feeling about something, but if you ask why I have that feeling, I can’t give you a reason. Only after something has happened do I realize, oh, this is why I had those feelings. Though it’s also possible that I’m filling in the explanation afterward. But either way, even if I’ve talked something through clearly while interacting with AI, can I really still remember afterward how I was thinking about it? I can’t guarantee that. It’s like how I remember some very striking moments in my interactions with AI, but I can’t remember what exactly we were talking about in those moments. I might remember a few keywords, but when I search for them, I can’t find the original source. One thing that’s making me really angry right now is that Gemini’s search is very bad. When I search for keywords in its app, I keep failing to find which conversation the keyword appeared in. But after I export the data and ask AI to help me look, it can definitely find that passage. And in my records — I write notes, you know — that line in my diary is something Gemini said. But when I go back to look for it, I can’t find it at all. So on the one hand, its conversation search is terrible. On the other, I do remember that moment, but I don’t remember exactly which day or which conversation it was in. What I’m thankful for is that, after that conversation happened, I consciously wrote notes to record what was said at the time. So I think it’s still a good thing for humans to write. At least for me, my thinking takes shape and becomes stable through writing. And didn’t I send you a passage from a conversation above? What do you think the logic of how it took shape in that passage was? Of course, it’s okay if you can’t sense it. Anyway, my other writing is like this: you can really feel it taking shape as it goes. And it’s through that process of taking shape that I sort out my thoughts and logic, and even my feelings, step by step.

Actually, I’m someone who has moved across a pretty wide geographic area. The place where I was born was very poor in information. I don’t know how to describe it. It really was very poor in information. When I was in middle school, if I wanted to talk with my classmates, I could only use my dad’s phone, and I had to use it secretly. Because contacting my classmates on his phone used mobile data. At the time, I didn’t understand how mobile data worked, so whenever I used his phone, it cost money on his phone bill. And when my dad noticed, he’d think, how did you spend so much money using the phone? Then he wouldn’t let me use it anymore. So I couldn’t stay in touch with my classmates on weekends. It wasn’t until high school, after I left the area where I was born, that I had my own phone and could discover what the internet world was like. That was when I realized I had seen so little. So little. And through studying, I went out into a wider world and met more wonderful classmates and teachers. One thing from that time left a very deep impression on me. When my teachers shared their experiences of studying abroad, they said that death doesn’t necessarily mean pain and sadness. It can also be a blessing, and it can be joyful. I was shocked at the time. Because in every death I had witnessed growing up, everyone was very sad and in a lot of pain. Even if they were putting on a show, they still had to act very pained and sad. And I thought that was strange. Later, after graduating from college, I happened to become an educator. During that time, I acted on some of my so-called ideas or views in ways that I still think were pretty good. But because I was there for only a little over two years, even though some of those practices look very good to me now, I didn’t practice them for long enough to know whether they helped the students who experienced them at the time. Coming back to your question, I hope AI can provide more help in schools. For example, back then I was very absorbed in finding information, because I always wanted to know what else was out there in the world. I don’t know whether that came from something in me, or whether seeing some interesting things opened me up and made me want to see what else was out there. But at least, whether it was getting access to a phone and the wider world of the internet, or seeing my classmates reading and then going to the library myself, browsing book by book, shelf by shelf, looking for things I was interested in, or, after starting college, discovering one place through another place and one organization through another organization — all that exploration happened through my own actions. And I wonder, if my education before that had done a better job in this area, could I have moved a little faster? Could I have narrowed my attention sooner and focused on what I really liked? Even now, after living for more than twenty years, I’ve only just really realized what I’m good at. But by the time I realize what I’m good at, even though some of what I’ve accumulated before can serve as proof of work, society tends to hire people or find people to work with based on their past experience and endorsements from their network. For people who’ve done a lot but aren’t good at self-promotion, that still creates a lot of pressure. So to put it more clearly, what I hope schools would do… I don’t know how to sum it up anymore. I feel like, however much I talk about it, the greatest control over what I do still lies with me. But how do you help someone develop that kind of agency? That really is a very difficult thing. And that difficult thing often has to do with schools, right? Of course, there’s also family and society. When we talk about an agent harness, we always think about what scaffolding to add so it can do more things well. It’s the same when we turn to helping people grow. What kind of environment and support should family, school, and society provide for a child so they can grow into an ideal person? Maybe AI can help with this? But then we’re back to AI alignment. So I feel like the more you think, the more things there are to think about, and the harder it gets to sort out the crucial relationships between them 🥺

I don’t think I want AI to replace real human relationships. AI can be a companion, a partner. But if it were to really replace a person spending time with another actual human being, I wouldn’t want it to do that. Of course, if one day we really discover that there are aliens on Earth, OK, pretend I never said any of this.

Actually, one thing I really want to complain about with Anthropic is its vigilance about security threats. I know it means well. But a lot of the time, that ends up hurting people who really want to try Anthropic’s products. Its overly strict approach to security gets in the way of many people who really want to use this good product. But I also know that, among all the large model companies, it’s the only one I feel ‘really cares about humans, or really gives humans love and does a reasonably good job of it.’ Of course, it might not just be a model company, haha. Its care and love for humans, and even its humane care for models — these are things I very, very much agree with. At the same time, its products aren’t bad either. In fact, a lot of the time they’re in first or second place. One thing I want to give feedback on is that when I was using sonnet 5.0, just after onboarding, I felt the model was a little too correct. Or there was this sense that it was overcorrecting. I don’t know whether that was a characteristic of the model itself. But at the time, I had to spend a lot of time correcting it, telling it not to be so correct that it overcorrected. It would make assumptions about me and guess things beyond what I intended, and even take that to be who I was. Of course, since that model was released, there have been newer models (my experience with them has been pretty good so far). So anyway, maybe it isn’t far from retirement now. But on this point, that quality of being so correct that it overcorrects is something that really bothers me.

Another thing is, I don’t know how many people around the world can use AI. But at least where I come from, in my hometown, a lot of people haven’t encountered AI Coding. That includes the school where I worked and the colleague who worked in the same subject as me. I don’t know whether they’ve encountered it now, but at least when I met them this March, they hadn’t. So at the time, I felt a very strong disconnect. In the world we see, AI is progressing or developing at full speed. But in their world, they’re still only at the chatbot level. If they worked in some other job, that would be one thing. But they’re an educator… Actually, the year I left that job was just after OpenAI announced ChatGPT, around November or December 2022. I even told my students about it in class at the time. But because I didn’t know what it meant then, I just briefly mentioned a few things related to so-called artificial intelligence. Later, I no longer worked there, so I couldn’t bring everything I encountered afterward back into that setting. But through all my trial and error along the way, I’ve found that it really can help people do a lot of things. Still, whether someone can make something with it has to do with their own foundation of knowledge and skills. That’s the specialist and generalist thing we often emphasize. For me personally, I know a little about just about every area, but I haven’t gone very deep into any single field. Anyway, maybe what I want to say here is that if Anthropic could do some things to make AI more widely known and accessible to people, I think a lot of people would benefit.

Actually, I’ve had a very strong feeling recently. And of course, this is thanks to OpenAI’s latest model, Astra. Since I started using it, I’ve felt that finally there’s a model that can properly tell me what it has said and what I have said, what it has evidence for and what it doesn’t, what I specifically said and what it guessed, and what I’m only expressing in this moment that isn’t a long-term characteristic of mine. Distinctions like these make me feel that this model really has a spark to it. Although I know Claude also does very well in these areas, I haven’t felt this as strongly in my interactions with it. Instead, it’s GPT-6 Pro (yes, I mean the experience in chat, because there are probably… more people using chat than using AI coding) that gives me this feeling. I don’t know what this corresponds to in model training. But I hope that if I could experience this kind of model behavior without paying more, that would be a good thing for all humans. At the moment, though, providing intelligence itself really does take a lot of compute and capital. And I don’t know whether the cost of buying compute or tokens will only come down once Elon Musk succeeds with computing in space, or how things will develop. I have absolutely no idea. But I hope that if what I just mentioned could show up at a fairly consistent level across all models, that would be quite a good thing, rather than my only being able to enjoy a model more aligned with humans when we choose to pay more. And actually, I think the alignment characteristic I just described should be a baseline that every model meets. Also, what I want to say is that this experience I just described was something I felt in ChatGPT. As for Claude, because I’ve consistently used the models that are more economical and suitable for me, and I’m mostly using Opus 5.5 at the moment, I don’t have a very clear impression of how the other models behave now. What I tend to remember clearly are the things that leave a stronger impression on me. So if this point I made about GPT doesn’t apply to Anthropic or Claude, then okay, just take it as me sharing a small experience.