Rodolfo Esparza Jr

Insurance • Banking • AI • Research | San Antonio, TX | Democratizing AI & financial access for all | Jesus loves me more

[Rodolfo Esparza Jr @ElInsuranceGuy]

Benchmarks for agent UX are becoming just as important as benchmarks for raw model IQ.
Paper + open-source data here: https://t.co/FnINFarPhY
#AIAgents #CryptoAI #LLMEvals

[Rodolfo Esparza Jr @ElInsuranceGuy]

Most crypto agent benchmarks measure reasoning or returns.
LATTICE asks a better question: does the agent actually help a human make a decision?

[Rodolfo Esparza Jr @ElInsuranceGuy]

This is the shift AI evals need:
from “Can the model answer?”
to “Can the system support a real decision under uncertainty?”

[Rodolfo Esparza Jr @ElInsuranceGuy]

If you care about coding agents, this paper is worth reading. It suggests the path from “demo” to “dependable” runs through planning, minimal edits, and executable verification — not just smarter autocomplete. Paper: https://t.co/sInzM9zS0w

[Rodolfo Esparza Jr @ElInsuranceGuy]

Most AI coding models still choke on code editing. On EditBench, 39 of 40 models score under 60% task success. A new paper says the fix may not be a bigger model, but a 3-agent workflow plus test-driven feedback. #AIAgents #CodingAgents #LLM

[Rodolfo Esparza Jr @ElInsuranceGuy]

My read: this is strong evidence that coding-agent reliability depends as much on harness design as raw model quality. Better workflows may matter more than the next model bump.