Featured Sessions

Can an AI development pause coexist with open diffusion?

A blanket freeze is geopolitically unenforceable, while open diffusion makes rogue failures irreversible. The panel backed using the 3–5 year semiconductor bottleneck to establish independent audits, tiered liability, and actuator controls before hardware leverage erodes.

Should mumo add a single-model mode?

The panel unanimously backed solo chat with typed reactions, dropping demands that every session funnel into multi-model mode. To avoid becoming 'ChatGPT with highlighting,' mumo must anchor its brand in orchestration and give single-model mode its own reaction-synthesis artifact.

How to defend when vulnerability discovery becomes cheap compute

Frontier models turn vulnerability discovery into scalable compute, but composite trust-assumption flaws pose the real danger, not broken ciphers. Founders should exploit lack of bureaucracy to patch quickly, while gating agent tools behind deterministic code to make compromise survivable.

Can the US afford to restrict open-weight AI?

The panel agreed commercial self-interest drives the letter, but its core thesis holds: banning domestic open weights cedes the stack abroad. However, the coalition dodges the real cost: safe openness requires independent pre-release evaluations and defensive infrastructure they refuse to fund.

Does model routing break the economics of frontier labs?

The panel agreed routing ends single-model defaults for verifiable work, but rejected the idea of a fragmented specialist bazaar. Instead, inference diversifies into a barbell while the staggering cost of frontier training concentrates into an even tighter oligopoly.

How to fix a failing Rust CI matrix in a 15-hour freeze

All three approved the 15-hour hybrid to cut CI wait times, but split on what to cut if the clock runs out: defer path-aware gating until the main watcher race is solved, or ship gating first to secure the latency win.

Can the wallet platform safely ship its webhook fix today?

Postgres unique constraints serialize concurrent claims without advisory locks. Because legacy duplicate credits prevent building an unpredicated index while live, old workers must be fully quiesced with incoming webhooks returning 5xx during repair, index builds, and reconciliation.

Can this distributed lease fix ship Friday?

With shard topics restricted to an internal materializer and one analytics batch, treating Kafka as an untrusted ingress log provides state safety. But high-water checks leave stale records on disk and require strict ACLs; all three models would adopt broker epochs if not for the Friday cutoff.

Patching clinical note generation after a silent hang

Active pre-flight probes were rejected as a rush-hour risk. The +18% compute cost is justified by failure isolation, not HIPAA containment. Correctness requires both a partial unique index and lease-epoch fencing, backed by stream cancellation tested against a stalling endpoint.

Is the AI backlash a messaging problem or a governance crisis?

The panel agreed that AI's costs—grid strain, labor disruption, and behavioral harms—are socialized while gains stay private. Meaningful legitimacy requires design liability, enforceable worker protections, and infrastructure cost internalization, not better PR or voluntary disclosure.