Post

im very curious how this is gonna play out, i see the risk as well, but two forces go against it i think: more user data makes the problem of creating good llm results much harder. its basically why all all these harnass memory providers like mem0, honcho and Hindsight suck, they pollute context

using memory/user data is much more a problem of what data not to use than what data to use, and getting this wrong results in worse output than no memory/personalisation so far my experience with llm memory providers is that it is very easy to overshoot and get worse results than no memory

the 'organise-my-life' assistant's value comes from the ability to integrate with every possible service. so even google/msft will have to balance providing open data in some form as well, otherwise theyre gonna get shut out in retaliation

Personally, I hope Ethan is right and they are being timid, because their potential ability to lock customers into an agentic environment where all their email and other data already lives seems overpowered to me.

in order to really make “we have your data” a moat, Google would have to close off access to gmail and drive and calendar to external tools, and they understand their role as a platform differently. which is probably wise, closed is a short-sighted strategy

I would describe their problem as “they have a hard time doing product development” rather than timidity

Look by contrast at something like OAI’s Sora. Technically it was impressive at the time. But it was a niche product; most people don’t need to generate videos that often. Personal assistant who tracks your todos by observing your email and what you’ve written lately? Widely needed.

(Also potentially intrusive and troubling, so you need to get the roll-out exactly right.)

I suspect it will look like it does now. To satisfy anti-trust, there will be ways to evacuate your data. Most people will not and be stuck in a hall of mirros where the hyperscalers retain massive margin on housing data with AI.

Coding already ran this experiment. Copilot had the deepest integration in the market - GitHub, VS Code, the enterprise deals - and it still lost the practitioners to a bare model in a terminal once the pricing changed.

Anecdotal, but in my circles Copilot was losing well before the pricing change — its performance and result quality kept degrading. I had paid yearly and by the end I was only using it for last pass code review. Not performing was what killed it, pricing change was MS pissing on the open grave.

I have never liked it either, but I remember some of my vibe-coder friends were paying $10/month on Copilot, while everything else cost $20.