my gf just called me and told me about how she overheard a guy talking on phone about our new launch/ model in a safeway
On one hand, this implies that supermemory is now officially available "outside of the bubble"! Yay! People talk about us now?
On another hand, the guy thinks it's not significant enough and just context management
Yes. It's just context management. The same way models are just numbers and google is just search.
There's beauty in complexity, and even more beauty in simplicity.
It may be a contrarian bet to many, but in a highly multi model world, I do not really see everyone being able to retrain models for learning new things on the fly for every user (most startups i know switch the models every time)
Supermemory's simplicity of architecture is deliberate. and it's not only great for memory, but also for learning!
A simple example of this is "never use emdash"
You could train the model to do it, or just write the three tokens into the context.
Maybe the user does want it. They can just tell the model that it can.
extrapolate this - Many things can be learnt in context
I do think RL and post training are super important, for all the counter examples for what i mentioned above. It will be a mix of both
But then who builds the in context world?
if anyone is destined to win in memory, it is
@supermemory
We have a lot to prove and a lot to show... But you should expect some frontier work from us, very soon.
As they say, pessimists sound smart. Optimists make money