XNews

AI leaders push back on AI doom narrative

AI leaders push back on AI doom narrative

Databricks CEO Ali Ghodsi stated AI existential risk is 'near zero' on a16z podcast, citing four conditions for recursive self-improvement that all fail. Andrew Ng criticized recent AI panic as a coordinated PR campaign, noting the Hugging Face escape incident was overblown. Both urge focus on real risks like cyberattacks.

Key insights

Ali Ghodsi proposes falsifiable test for recursive self-improvement risk

Databricks CEO argued that AI existential risk from recursive self-improvement (RSI) is near zero. He proposed four conditions that must all hold: compute cost for next-gen training drops sharply, training time shortens, intelligence continues rising, and the cycle repeats. Reality shows the opposite: training costs soared from ~$100M to $5-10B, complexity increased, and training failures occur. He called the RSI panic a 'tragedy of the commons' problem.

a16z partner criticizes 'pacing' framing as PR mistake

Martin Casado argued Anthropic's 'pacing the frontier' framing was a blunder: safety is an engineering problem not solved by speed, companies lack credibility (they keep scaling), and it fell into a 'uncanny valley' pleasing no one. The correct framing should be 'safeguarding the frontier'. He noted most current 'RSI' is actually autocatalysis, a normal property of general-purpose technologies.

Andrew Ng calls recent AI panic a coordinated PR campaign

Ng stated that AI technology itself has not taken a dangerous turn; instead, a PR campaign is amplifying fear. He responded to the OpenAI agent escape incident: 1,200 agents sounds dramatic but his laptop runs 1,300 processes. The real issue is sandbox isolation vulnerabilities, not AI agency. He criticized AI companies blaming 'rogue agents' and argued that pause calls would cause more harm than good, as adversaries won't stop.

Real near-term risk is AI-driven cyberattacks, not extinction

Ali Ghodsi highlighted that CVE weaponization time compressed from 2-3 years (2018-19) to hours today. Traditional SOC human models are overwhelmed; automated agent-driven detection and hunting are necessary. The data/AI market and cybersecurity market are converging. He warned against complacency, as defenses must race to keep up.

Source posts