Skip to content

Emmett Shear: Alignment Protocols

Talk Emmett Shear
researcher practitioner
Open resource →

Emmett Shear, founder of Softmax and former interim CEO of OpenAI, argues that alignment is not a static property but a capability built through a protocol stack: shared success (biological stress-sharing like reactive oxygen species), specialization (cellular differentiation, corporate job roles), and enforcement (immune systems, auditors) — in that order. He critiques mainstream AI safety for starting with enforcement and control rather than cultivating mutual care and interdependence from small-scale interactions upward, advocating multi-agent rather than singleton AI architectures.

Related resources

Talk

AI as the Anti-product – Guest Talk with Peter Wang

Peter Wang (Anaconda co-founder, now Chief AI Officer, and founder of the Sky Seed fund) argues that modern software/AI development is best understood through a complexity-versus-complication lens: tool stacks like CUDA and PyTorch remain complicated (Boeing-747-style, decomposable) while the models and ecosystems they produce are genuinely complex and emergent, and 'vibe coding' represents a quasi-religious third mode of engaging code beyond craftsmanship and managerial industrialization. He traces how Anaconda/NumPy's open-source success came from humble, itch-scratching human ecologies (e.g. Travis Oliphant merging numeric and numarray into NumPy) rather than top-down design, and warns that protocols like AT Protocol/Blue Sky must avoid centralized 'technocratic' forced upgrades, instead preserving pluralistic values across nodes the way open source avoided forced EOL upgrades.

complexity-vs-complication forced-upgrades-and-coherent-pluralism-in-decentralized-systems governance

Peter Wang

Talk

Protocols in Electronic Design

Engineer Vikram Shaker (Qualcomm RF front-end) surveys protocols across electronic design, from hard standards like NIST's cesium-clock definition of the second and the 50-ohm coaxial impedance convention, to communication standards (USB, chiplet interconnects like UCIe and CXL), validation failures (Intel's Pentium FDIV bug and the resulting birth of formal verification), and documentation disasters (NASA's Mars Climate Orbiter unit mismatch). He frames these through Protocol Institute concepts like 'tension' versus 'trade-off' and the efficiency-thoroughness (ETO) frontier, and closes by examining unwritten cultural, ethical, and geopolitical protocols governing chip industry work.

chiplet-interconnect-protocols-ucie-cxl-bow documentation-and-tribal-protocol-failures efficiency-thoroughness-eto-frontier

Vikram Shaker

Talk

Public Intelligence

Kevin Kelly, Wired co-founder, presents his essay 'Public Intelligence,' arguing AI should be built as a commons like the internet or Wikipedia rather than owned by corporations or nations—combining public access, accountability, data (the 'omnipedia'), funding, and participation. In the Q&A moderated by Venkatesh Rao at a Bangkok town hall, Kelly argues decentralized AI will require corporate risk-taking to prove viability before being 'returned' to public commons, and identifies failure modes including fragmentation into competing sovereign national AIs and technical infeasibility of planetary-scale federated intelligence.

architecture-of-participation federateddecentralized-mixture-of-experts-ai governance

Kevin Kelly, Venkatesh Rao

Topics

alignment-as-capability-not-property attractor-basins-vs-goal-vectors governance multi-agent-alignment-over-singleton-ai protocols shared-success-via-stress-sharing-reactive-oxygen-species specialization-and-enforcement-protocol-stack technology town-hall