avatar
Ye Zhang 📜
15,114 Twitter followers
Follow
Cofounder @Scroll_ZKP. Doing research about zero-knowledge proof, system, and hardware.
Posts
avatar
Ye Zhang 📜
03-15
Thread
For those who don’t know, this is my favorite feature in ChatGPT: You can branch the current conversation into a new chat while keeping the same context. My usual pattern is this: I’m learning something new, the answer introduces unfamiliar concepts, and I want to dig into those concepts more deeply. But doing that in the main chat can taint the original context. Branching solves that. I can explore those questions in a separate chat, then come back to the main conversation. However, I usually have too many follow-up questions, so it turns into branches of branches of branches, and it quickly becomes hard to manage. I think the best chat app for learning should have a tree structure. You could keep exploring new concepts on different branches, while still being able to manage everything and all context easily in a visual way. I’ve wanted to vibe-code something like this, but I’ve been too lazy + all my past context is already in ChatGPT + I’d rather use my subscription than build on the API. Still, I think this is a real need for power learners like me. Another need is having some separate thread monitor the conversations so it doesn’t go down a rabbit hole, because the user (me) is the only one steering the whole conversation...(similar to the problems mentioned in last tweet, I keep asking random distracted questions and need some "advisor" to stop me stuck in the same place with a global view). @sama @ChatGPTapp @JustinBleuel
STUCK
0%
avatar
Ye Zhang 📜
02-28
Today, Anthropic is big enough to say no. However, most people don’t really have “choice” unless they’re big enough to matter. Terms of service is basically a binary button: accept or leave. And if you need the service, “leave” isn’t realistic. If every major LLM collects your data by default, how many people can truly opt out? (Yes, privacy-preserving models exist. But if those modes are much weaker, they aren’t real alternatives. Many of those companies also die because they lack access to data as a flywheel. You have to survive long enough and be at the table to even talk about your values.) When intelligence and distribution concentrate in a few hands, companies don’t just build products. They set the rules. Everyone else adapts. This reminds me of that tweet a while ago: Gemini basically required activity collection to use it, or your AI would forget every conversation 👇
Ye Zhang
@yezhang1998
12-21
Gemini has gotten way better technically, but its privacy model is still a hard no for me. You're forced into this tradeoff: • Activity ON → Keep chat history, but chats can train Google's models • Activity OFF → No training use, but history deleted after 72 hours That's
ANTHROPIC
14.01%
avatar
Ye Zhang 📜
12-21
Recently, many people on my timeline have started switching to Gemini. I want to offer a heads-up: I don't feel Gemini is significantly better than Claude + GPT, and I find its data collection and privacy settings completely unacceptable. With Gemini, you're forced to choose: 1. Turn off data collection entirely (meaning all your chat history is only retained for 72 hours) 2. Turn on data collection completely (meaning your chat history will be used for Google's future training). Any other large model, like Grok, GPT, or Claude, allows you to retain conversation data but choose not to use it for training. Only Google deliberately doesn't do this… I think this is incredibly restrictive, since a 72-hour chat memory box isn't very useful… Google will do a great job; they have data from Gmail, Drive, YouTube, and so many chats. They'll improve, but imagine a day when much of your information directly becomes part of the model's pre-training knowledge—how terrifying would that be? Furthermore, if everyone starts this cutthroat competition, forcing users to submit data, it will only worsen the entire ecosystem.
Ye Zhang
@yezhang1998
12-21
Gemini has gotten way better technically, but its privacy model is still a hard no for me. You're forced into this tradeoff: • Activity ON → Keep chat history, but chats can train Google's models • Activity OFF → No training use, but history deleted after 72 hours That's
loading indicator
Loading..