OpenAI chief scientist Jakub Pachocki stated the company is approaching models capable of working indefinitely in a coherent way, similar to human researchers.1 "I think we will get to a point where you kind of have a whole research lab in a data center," Pachocki said.1
The shift from code assistance to autonomous research capabilities is creating a split AI infrastructure market. NVIDIA RTX PRO 6000 Blackwell workstations target high-end training workloads, while companies optimize for inference efficiency.2 Palantir shares jumped 6% as investors bet on AI infrastructure demand.1
Pachocki explained that simple boosts in capability now enable models to work longer without human help.1 "I think we are getting close to a point where we'll have models capable of working indefinitely in a coherent way just like people do," he noted.1
New technologies are emerging to support this transition. DiLoCoX-107B enables distributed training across decentralized infrastructure, while RL-KPI frameworks optimize reinforcement learning for extended autonomous operations.1
Pachocki acknowledged regulatory challenges accompanying concentrated AI capabilities. "I think this is a big challenge for governments to figure out," he said.1 He advocates deploying powerful models in sandboxes isolated from systems they could damage or exploit.1
The infrastructure investment thesis centers on companies providing hardware and platforms for both training and inference. As AI transitions from assisted coding to autonomous research, demand is shifting toward systems supporting extended operation without human oversight.
Startup funding dropped sharply in March 2026, suggesting investors are consolidating capital into proven AI infrastructure plays rather than speculative applications.3 The bifurcation between training and inference infrastructure creates opportunities in specialized hardware, cloud platforms, and orchestration software.
Safety concerns remain unresolved as models gain autonomy. Pachocki's sandbox proposal indicates OpenAI recognizes deployment risks, though specific implementation details and regulatory frameworks are still being developed.1
