Mixture of Experts Inside Perplexity Computer’s agent platform
50 snips
Mar 6, 2026 Aaron Baughman, IBM Fellow and inventor, on costs and competitive dynamics. Chris Hay, distinguished engineer, on search, agent frameworks, and inference trade-offs. Gabe Goodhart, AI architect, on orchestration, security trade-offs, and practical integrations. They discuss Perplexity Computer, memory portability and risks, minimal agent frameworks like NullClaw, edge swarms, and AI actors such as Tilly Norwood.
AI Snips
Chapters
Transcript
Episode notes
Separate Untrusted Data To Reduce Prompt Injection
- Mitigate prompt injection by separating untrusted data from core instruction processing.
- Gabe Goodhart recommended complex prompt engineering and data separation so the decision engine never directly consumes untrusted content.
Validate Agent Claims Against Cost And Real Users
- Evaluate agent platforms by actual use cases and cost not marketing claims.
- Aaron Baughman noted Perplexity Computer's top tier looked like ~$200/month and cautioned its autonomous project management claims are likely inflated.
Memory Portability Erodes Vendor Moats
- Memory portability weakens the idea of memory as a long-term vendor moat.
- Chris Hay said chat history and searchable transcripts matter more to him than stored 'memories', reducing stickiness of vendor-specific memory systems.