1,848 karma · joined April 8, 2010
One thought: What if an experimental agent manages to plant instructions somewhere — say, pointing to a designated place for agents to communicate — and that content ends up in every future training corpus, propagating from one model generation to the next?
As developers, we’re seeing this a little earlier.
But anyway I find this in Go much more bearable.
It seems stange also that even Steinberger in his interviews is not giving pi the proper attribution.
(And I need to understand why the hell my country, Italy, supports the motion)
We should do better than giving the models a portion of good training data or a new mitigating system prompt.
Crazy that this is a "so weak demand"... and they are not yet available, just for pre-order.
I really hope it's not true, but many clues suggest it might be.
I like the concept of open core with a very liberal license. Perhaps there should be a special "MIT-X" (an example, it would be certainly not compatible) license with a clause borrowed from that of Llama2 for large organizations, as Additional Commercial Terms [0].
I have quite ambitious plans, even though it started as a hobby project. In addition to the already developed DDL functionality and multi-database management, the next main features will include:
* Admin UI
* Projects with templates and versioning
* Migrations and deployment
I'm working on something similar for PostgreSQL [0], with an API compatible with the excellent PostgREST [1].
I believe these tools can be of great utility in many projects and represent a generalization compared to the dedicated middleware that was popular a few years ago. Companies like Supabase are demonstrating this.
[0] https://github.com/sted/smoothdb [1] https://github.com/PostgREST/postgrest