Interview
Founder Eric Steinberger on Magic’s Counterintuitive Approach to Pursuing AGI
- General domain reasoning, long-horizon tasks, and reliability issues are expected to persist until solved through increased inference and test-time compute, with optimizing resource allocation per token identified as a remaining major developmental hurdle.
- AI software engineering is predicted to achieve higher automation levels and a "colleague-tier" status within a timeframe significantly under ten years, driven by models that manage tasks without specific bug instructions once they surpass human competence.
- Economic automation is projected to proceed rapidly without resistance as long as inputs are reduced for equivalent or greater output, with significant societal impact anticipated within the next one to five years.
- The company plans to deploy imperfect products to reach a state where models function as superior colleagues, though the ultimate goal remains a fully autonomous agent capable of managing human workflows.
- Future value accrual is predicted to concentrate at the AGI and hardware levels rather than application layers, with the company expecting to operate with a team size not exceeding the tens of people even during large-scale deployment.
- Long-context capabilities require models to fully retain context windows without implicit priors, critiquing current "Needle in a Haystack" benchmarks as insufficient for measuring true long-context performance.
- Research progress is tracked closely to an internal AGI roadmap, with plans to open-source specific evaluation methods to allow architectural comparison while maintaining that the specific algorithmic details remain to be fully resolved.
- The field is expected to continue rapid advancement with models surpassing existing benchmarks, contrasting with public perceptions of stagnation, and alignment strategies may be solved by cloning capable agents at scale if return on investment is demonstrated.