Post

Ed
@ed3d.net

one thing I do wonder about when it comes to this attitude is because I settled, early on, on a couple of things: - don’t spend less to achieve a goal at the cost of quality - it is all about latency and not about bandwidth

Ed@ed3d.net

legit, I do not think we need better large models than we have now and I’m still not seeing substantial improvement in the places where models are struggling. the trend is to topically RL harder and to try to just make them guess your under-specified prompts better and I think that’s a mug’s game

Ed
@ed3d.net

a polytoken ticket, run through the normal workflow, takes between four and six hours. I have accelerated ones for faster turnarounds at the cost of more human interaction but my stuff is mostly AI container ships putt-putting across the ocean

Ed
@ed3d.net

Polytoken is closing in on a million lines of code, counting the test suite (hovering around 50%). It has basically all been built this way.

Ed
@ed3d.net

And not to toot my own horn too much, but one thing I’m proud of is that there are very few regressions in that time. And all but two I can think of were my own meatbag fault—bad design and insufficient guardrails on flagging changes for unexpected disruption

0 reposts9 likes

No replies yet.