@SebAaltonen may i present you https://t.co/SgfSMmbast ?
The strategy of not hand holding AI creates bloat... you need to know what tools exist and properly guide the implementation with concrete info like: "use this library, don't do that e.g. write python to read PNGs in a C++ project lol
@Diodotus__l@PicturesFoIder I don't think it's fair to reduce a bad person to "jewish". It's a bad person regardless of their origin and should be held accountable by the same laws as everybody.
@ThePrimeagen So the OS has an "MCP"-like API, allowing an agent to do I/O like a user would do with a mouse and a keyboard?
And the agents are using that to test various user journeys?
@rfleury@toncijukic@SebAaltonen Tests, code review or QAing the application withoutbare for me the main things I use AI for.
Agents are the testers and users I don't have giving some sort of external feedback without my own bias.
@AgileJebrim@SecPerkinsStan I mean. If/else is elementwise multiplication with a mask. Fixed for loops can be chained multiplications.
Lots of logic can be converted in matmuls/data oriented design/ecs, no?
The true branching logic can stay on cpu, but most of the work can be a tensor.
Where am i wrong?
@TheAITradeoff@FireworksAI_HQ If you still write most of the code and use the model just to "fill the boilerplate" (e.g. tests/code review at each step/docstrings), then having this kind of speed is amazing.