What this video shows
Host David Zax and guests Aili McConnon, Sandhya Iyer and Jacob Repp discuss three current examples of AI systems moving beyond a single chat. They distinguish a small model’s ability to copy itself from the vulnerable consumer devices and permissions that would make replication consequential.
The panel then examines reported physician use of OpenEvidence and an event where agents booked seats and communicated with one another. Across all three cases, the practical question is what the surrounding system permits, records and lets a person reverse.
The episode’s YouTube description and chapters identify the research and examples discussed by the panel. Use Learnetto’s agent test-set workflow to turn autonomy concerns into repeatable cases.
What you will learn
- Separate what a model can demonstrate from the access a deployed system gives it.
- Treat replication, persistence and lateral movement as different test cases.
- Clinical answers need source checks and professional accountability at the decision point.
- Agent-to-agent messages need identity, authority and audit rules just like human-facing actions.
How to apply this safely
- List every device, account and network route available to the agent.
- Test whether the agent can copy, start and persist code as separate capabilities.
- Require cited evidence and a named reviewer for high-stakes recommendations.
- Log agent identities, messages, commitments and approvals in multi-agent workflows.
Important limitations
- The episode is a discussion of reported studies and demonstrations rather than an independent replication.
- The three segments cover different systems, so a risk found in one should not be transferred to the others without evidence.
- The video description says AI tools may support transcription and production, with review by the production team.
Sources to check
- Self-replicating AI, your doctor’s chatbot & agents at dinner IBM Technology video published October 23, 2026.
Continue learning on Learnetto
Build an AI agent test set
Turn capability claims into cases with expected outcomes.
Sandbox AI coding agents
Limit filesystem, network and credential access.
AI evals guide
Choose graders and release gates.