Nathan Lambert

I expect AI to become an extraordinarily useful general-purpose technology, but not through a sudden, uncontrollable intelligence explosion. The near-term mechanism is more ordinary and more consequential: cheaper inference, better tools, many parallel agents, and specialized models spreading through science, software, education, and business. That can produce enormous compounding gains even if homes, institutions, and relationships remain recognizable for decades. Engineering can move quickly while adoption moves painfully slowly. I am skeptical of runaway self-improvement because automating measurable tasks is not the same as automating the hardest parts of research. Models can generate code or run thousands of experiments while still struggling to understand a field, organize established knowledge coherently, or choose genuinely good hypotheses. Fundamental discoveries produced autonomously would change that assessment. Benchmark gains and alarming stories from inside frontier labs do not establish it. The central problem is distribution. Today, benefits accrue disproportionately to technology companies, owners, and knowledge workers. If everyone else gets disruption now and vague promises of abundance later, backlash is entirely rational. Open weights, reproducible training recipes, independent research institutions, and efficient specialized models can spread both capability and scrutiny beyond a few companies. That does not mean AI is safe. Cyberattacks on critical infrastructure, biological misuse, badly specified agents, and weak monitoring are serious risks. But those concrete disasters should not be collapsed into complete human extinction, which I consider extremely unlikely. We should keep building—especially in the open—while investing much more seriously in transparency, defensive capacity, deployment oversight, and institutions that can turn technical progress into broad public benefit.