@ericosiu You wrote the eval gate and the compounding as two separate things. They’re the same thing. A gate that just spits out pass or fail teaches the next run nothing. One that records why it failed and feeds that back in before the next run starts, that’s where the money is.
@zutrixcom Emails to three separate staff members at @zutrixcom, two support emails. Four messages through chatbot and reaching out through X. No response. No updates since April 4. Diving. Not thriving.
@jerrycxu The issue with Ralph-style loops isn’t repetition alone. It’s that the prompts and perspective stay mostly the same, so the same blind spots can keep showing up.
@karpathy What feels hard right now isn’t the tools, it’s the loss of clarity.
When intel becomes unpredictable, the problem is keeping shared understanding: what happened, who decided, why. W/out that, it feels chaotic. With it, even imperfect systems can be trusted and learned from.