Leading AI Researchers Warn Superintelligence Poses Existential Risk

A new video series published by Palisade Research on frominside.ai brings together more than a dozen current and former researchers from OpenAI, Google DeepMind, and Anthropic. Each participant speaks candidly about the probability that advanced AI systems could drive human extinction.

Geoffrey Irving, formerly of OpenAI and DeepMind, describes the extinction likelihood as roughly a coin flip. Neel Nanda, a DeepMind research scientist, puts the chance at no less than ten percent. Daniel Kokotajilo, a former OpenAI researcher, characterizes superintelligence as "god-like powerful" and admits the field still lacks any reliable control mechanism.

When asked why they continue working on technology they consider dangerous, several respondents argue that their presence inside the labs directly reduces risk. Mary Phuong of Google acknowledges the conflict of interest, while Nanda notes that the companies would not halt development if any single researcher departed.

Why it matters for GPU infrastructure

Safety research, red-teaming, and alignment experiments demand massive, isolated compute environments with full auditability. Cloud providers must deliver hardened tenancy, real-time monitoring, and policy-enforced resource limits so that frontier labs can test containment strategies without jeopardizing production workloads. This shifts procurement criteria toward platforms that embed governance tooling at the hardware layer.

As the industry internalizes these warnings, capital allocation will increasingly favor GPU clouds that can demonstrate verifiable safety controls alongside raw performance. Providers that integrate observability, access segmentation, and compliance reporting into their stack will become the default choice for responsible frontier development.