David Zax hosts Aili McConnon, Sandhya Iyer and Jacob Repp on IBM Technology’s 36-minute Mixture of Experts roundtable.
Small models and insecure devices
- The panel discusses research approaching self-replication across cheap devices, with substantial speed and hardware limits still in play.
- Repp puts more immediate emphasis on prompt injection: hostile content steering a home agent that already has access to useful tools.
- Local inference can preserve data control, but inputs, permissions, sandboxing and device security still matter.
What clinicians ask a medical chatbot
- OpenEvidence use is discussed as evidence retrieval and treatment support, with many questions crossing specialty boundaries.
- The panel sees potential for faster triage and knowledge access, while retaining specialist judgment and source checks.
- Adoption and query counts do not establish safer treatment or better patient outcomes.
- The episode notes pharmaceutical advertising and company involvement in the cited study; both matter when assessing its evidence.
Agents can talk without having authority to act
- At a reported founders’ dinner, assistants chatted while their users met; those chat agents lacked action-taking tools.
- The panel imagines useful matching and coordination, but asks what commitments an agent should be allowed to make.
- Repp describes improved coordination in his own experiments, including the need to assign a clear leader and make activity visible.
Caveat: this is a news discussion. Replication forecasts, clinical benefits and coordination gains remain attributed claims or hypotheses rather than independently verified results.
“What can that agent do and what can get into that agent?” — Jacob Repp, 11:32