AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

Eric Zelikman on STaR, xAI, and Starting Humans&

· No Priors

Eric Zelikman, formerly of Stanford and xAI, discusses his research trajectory from the STaR (Self-Taught Reasoner) paper through Quiet-STaR to his work on Grok 2/3/4, and announces his new company Humans& focused on EQ rather than IQ. He explains how STaR iteratively trains models on their own successful reasoning traces and how scaling it revealed no obvious plateau on multi-digit arithmetic. At xAI he contributed to pre-training data for Grok 2, the reasoning recipe for Grok 3, and agentic infrastructure for Grok 4, noting that current frontier models are 'jaggedly' smart—capable of PhD-level HLE questions but tripped up by trick questions, and very sensitive to context quality and verifiability of the task. Zelikman argues the industry is locked in a task-centric, single-turn optimization regime driven by easy credit assignment, which is why benchmarks and progress metrics still revolve around isolated IQ-style tasks rather than long-term human collaboration. He frames his new venture Humans& as a bet that 'EQ'—models that understand people's long-term goals, express uncertainty, ask clarifying questions, retain memory, and reason about consequences across timescales—matters more than further IQ scaling. He contrasts this with the dominant paradigm of removing humans from the loop to extend autonomous task horizons, arguing that empowering humans to grow the pie is more likely to solve fundamental problems than fully autonomous AI agents.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.