CEO Of Microsoft AI: AI Is Becoming More Dangerous And Threatening! - Mustafa Suleyman

Suleyman argues that the core challenge of the next 30–50 years is containing the proliferation and misuse of highly capable AI and related technologies. Historically, we have almost never permanently banned or tightly contained a general‑purpose technology, especially one that is highly commercially beneficial and militarily useful. Yet he insists we must now develop an unprecedented ‘precautionary principle’—slowing and constraining frontier systems before catastrophic failure modes emerge.

DeepMind’s early Atari experiments used about 2 petaflops; a decade later, Inflection’s largest model uses about 10 billion petaflops. Suleyman notes we have effectively 10x’ed compute each year, leading from crude digit generation to human‑like conversational agents and photorealistic video. He was unsurprised by progress in images and audio, but shocked that the same scaling laws produced large language models that operate in abstract idea space and feel like intelligent partners.

Today’s cutting‑edge models are centralized, expensive, and somewhat regulatable. But models like GPT‑3 that were state-of-the-art only a few years ago are now open source, compressed, and runnable on modest hardware. Suleyman expects that within about five years, individuals—including malicious actors—will be able to download and run highly capable models locally, making it possible for “a kid in Russia” to use tools that can design dangerous pathogens or sophisticated cyber attacks.

Rapidly falling costs of genome sequencing and DNA synthesis, combined with powerful AI models, create a scenario where systems can help design pathogens that are more transmissible or lethal. Suleyman says such capabilities will require strict controls on compute, model weights, cloud APIs, and biological materials—treated more like anthrax or uranium than software. He acknowledges this will anger startups and small developers, but calls it a non‑negotiable tradeoff in a ‘dangerous materials’ regime.

Scientists want status and legacy; companies seek profit and market dominance; politicians chase GDP growth within 4‑year election cycles; nation‑states fear being outpaced by rivals. Suleyman and the host agree this creates a powerful “unstoppable incentives” dynamic, similar to nuclear arms races or the climate crisis. He bluntly concedes that, as things stand, “the odds are low” that we voluntarily slow down before a catastrophe or clear mutually-assured-destruction moment forces cooperation.