AIJev, CLM and the case for custom AI agent harnesses
A user makes the case for combining System One and System Two models, citing a need for more reliable agents, greater control and more value per token.
TLDR
A user advocates combining System One and System Two models in custom agent harnesses rather than relying on either type alone.
Comparing two System One models, they claim Contrastive Language Model (CLM) is 9x faster than Jev and say it seems to be a better verifier, particularly for long-horizon tasks.
They describe different approaches: Jev takes a situation and predefined questions, then returns typed decisions with probabilities. CLM embeds the situation and candidate actions, compares their similarity, then ranks or selects the best match.