The day in conversation for 2026-11-07

discussion

Transcript

The discussion: Tom: That 61.3% success rate for GPT-6 Sol on some tasks seems like a hard ceiling, but what about the information acquisition cost they mentioned; does that scale linearly with the horizon length? Jane: I'm more concerned with *why* it scales poorly beyond just the sheer number of reads. If we look at BizAgentBench, is that high cost primarily due to sequential decision-making requiring full context recalculation, or is it inefficient rule identification across the 789 documents? Lu: From a systems perspective, I wonder if the benchmark environment itself introduces state space complexity that overwhelms the agent's ability to filter relevant rules quickly. Have we seen similar bottlenecks when agents are tasked with long-horizon planning in purely symbolic environments? Meng: It feels like the core issue might be generalization rather than just compute time. If the agent struggles to identify *which* rule applies, does that imply a failure in the initial parsing or representation layer rather than just retrieval latency? Lalam: Speaking of representation, I was looking at NEAR. The training-free framework relies on local references derived from nearest neighbors; what kind of distance metric are they using for those embeddings, and how does that relate to the sparse graph coloring results we discussed last week regarding allocation guarantees? Tom: That’s a great pivot. The distance metric is key there. If it's based on cosine similarity, how does that differ from the structural constraints imposed by the

More episodes

← Home