A bit of PR puffery, but it is fair to say that between Gemini and others it’s now been clearly demonstrated that OpenAI doesn’t have any clear moat.
That means responses can be far more tailored - it knows what your job is, knows where you go with friends, knows that when you ask about 'dates' you mean romantic relationships and which ones are going well or badly not the fruit, etc.
Eventually when they make it work better, open ai can be your friend and confident, and you wouldn't dump your friend of many years to make another new friend without good reason.
Perhaps it'll be easy to migrate memories indeed (I mean there are already plugins that sort of claim to do it, and it doesn't seem very hard), but it certainly is a very differentiating feature at the moment.
I also use ChatGPT as my daily "chat LLM" because of memory, and, especially, because of the voice chat, which I still feel is miles better than any competition. People say Gemini voice chat is great, but I find it terrible. Maybe I'm on the wrong side of an A/B test.
The real challenge for Google is going to be using that information in a privacy-conscious way. If this was 2006 and Google was still a darling child that could do no wrong, they'd have already integrated all of that information and tried to sell it as a "magical experience". Now all it'll take is one public slip-up and the media will pounce. I bet this is why they haven't done that integration yet.
Many people slowly open up to an LLM as if they were meeting someone. Sure, they might open up faster or share some morally questionable things earlier on, but there are some things that they hide even from the LLM (like one hides thoughts from oneself, only to then open up to a friend). To know that an LLM knows everything about you will certainly alienate many people, especially because who I am today is very different from who I was five years ago, or two weeks ago when I was mad and acted irrationally.
Google has loads of information, but it knows very little of how I actually think. Of what I feel. Of the memories I cherish. It may know what I should buy, or my interests in general. It may know where I live, my age, my friends, the kind of writing I had ten years ago and have now, and many many other things which are definitely interesting and useful, but don't really amount to knowing me. When people around me say "ChatGPT knows them", this is not what they are talking about at all. (And, in part, it's also because they are making some of it up, sure)
We know a lot about famous people, historical figures. We know their biographies, their struggles, their life story. But they would surely not get the feeling that we "know them" or that we "get them", because that's something they would have to forge together with us, by priming us the right way, or by providing us with their raw, unfiltered thoughts in a dialogue. To truly know someone is to forge a bond with them — to me, no one is known alone, we are all known to each other. I don't think google (or apple, or whomever) can do that without it being born out of a two-way street (user and LLM)[1]. Especially if we then take into account the aforementioned issue that we evolve, our beliefs change, how we feel about the past changes, and others.
[1] But — and I guess sort of contradicting myself — Google could certainly try to grab all my data and forge that conversation and connection. Prompt me with questions about things, and so on. Like a therapist who has suddenly come into possession of all our diaries and whom we slowly, but surely, open up to. Google could definitely intelligently go from the information to the feeling of connection.
Do people want an LLM to "know them"? I literally shuddered at the thought. That sounds like a dystopian hell to me.
But I think Google has, or can infer, a lot more of that data than people realize. If you're on Android you're probably opted into Google Photos, and they can mine a ton of context about you out of there. Certainly infer information about who is important to you, even if you don't realize it yourself. And let's face it, people aren't that unique. It doesn't take much pattern matching to come up with text that looks insightful and deep, but is actually superficial. Look at cold-reading psychics for examples of how trivial it is.
I get the impression that most non-techies have either never tried "AI", or regard it as Google (search) on steroids for answering questions.
Maybe more related to his (sad but true) senility rather than lack of interest, but I was a bit shocked to see the physicist Roger Penrose interviewed recently by Curt Jaimungal, and when asked if he had tried LLMs/ChatGPT assumed the conversation was about the "stupid lady" (his words) ELIZA (fake chatbot from the 60's), evidentially never having even heard of LLMs!
I didn't tell her to download the app, nor she is a tech-y person, she just did on her own.
It creeps me out when a past session poisons a current one.
Their 'memory' is mostly unhelpful and gets in the way. At best it saves you from prompting some context, but more often than not it adds so much irrelevant context that it over fits responses so hard that it makes them completely useless, specially in exploratory sessions.
https://cdn.openai.com/pdf/a253471f-8260-40c6-a2cc-aa93fe9f1...
Note that since 2024, Google no longer has your location history on their servers, it's only stored locally: https://www.theguardian.com/technology/article/2024/jun/06/g...
This sounds like first-mover advantage more than a moat.
That may seem minor, but it compounds over time and it's surprising how much ChatGPT knows about me now. I asked ChatGPT to roast me again at the end of last year, and I was a bit taken aback that it had even figured out the broader problem I'm working on and the high level approach I'm taking, something I had never explicitly mentioned. In fact, it even nailed some aspects of my personality that were not obvious at all from the chats.
I'm not saying it's a deep moat, especially for the less frequent users, but it's there.
I’m not saying it’s minor. And one could argue first-mover advantages are a form of moat.
But the advantage is limited to those who have used ChatGPT. For anyone else, it doesn’t apply. That’s different from a moat, which tends to be more fundamental.
For instance, I've been struggling against a specific problem for a very long time, using ChatGPT heavily for exploration. In the roast, it chided me for being eternally in search of elegant perfect solutions instead of shipping something that works at all. But that's because it only sees the targeted chats I've had with it, and not the brute force methods and hacks I've been piling on elsewhere to make progress!
I'd bet with better context it would have been more right. But the surprising thing is what it got right was also not very obvious from the chats. Also for something that has only intermittent existence when prompted, it did display some sense of time passing. I wonder if it noticed the timestamps on our chats?
Notably, that roast evolved into an ad-hoc therapy session and eventually into a technical debugging and product roadmap discussion.
A programmer, researcher, computer vision expert, product manager, therapist, accountability partner, and more all in a package that I'd pay a lot of money if it wasn't available for free. If anything I think the AI revolution is rather underplayed.
Branding isn't a moat when, as far as the mass market is concerned, you are 2 years old.
Branding is a moat when you're IBM, Microsoft (and more recently) Google, Meta, etc.
In the onboarding flow, I can ask you, "Do you use another LLM?" If so, give it this prompt and then give me the memory file that outputs.
The harnessing in Google's agentic IDE (Antigravity) is pretty great - the output quality is indistinguishable between Opus 4.5 and Gemini 3 for my use cases[1]
1. I tend to give detailed requirements for small-to-medium sized tasks (T-shirt sizing). YMMV on larger, less detailed tasks.
I don't know how much secret sauce is in CC vs the underlying model, but I would need a lot of convincing to even bother with Gemini CLI again.
Claude Code is much better than Gemini CLI though.
If it is so clear, then investors will want to pull their money out.
It's not about who's the best, it's about where the market is. Dogpiling on growing companies is a proven way to make a lot of money, so people do it, and it's accelerated by index funds. The REAL people supporting Google and Nvidia isn't wallstreet, it's your 401K.