Cerebras CEO urges a broader debate on AI’s risks and promise

Beyond the doomsday debate

Feldman was responding to familiar fears: AI models slipping out of control and speculation about the odds that AI could destroy humanity by the end of the decade. The subject has become mainstream enough that Anthropic CEO Dario Amodei was recently mocked in a sketch on “Saturday Night Live.” A fictional character summed up the conversations by saying participants “don’t approve of what they’re doing.”

Feldman’s counterpoint is not that the risks are imaginary. Powerful technologies raise the question of what happens if bad actors get hold of them, he said; there is no way to build such technology without confronting that problem. AI should be treated as a powerful tool, approached thoughtfully and regulated sensibly.

His argument is that risk should not crowd out the possible upside. The cancer example is a striking promise, but it is an aspiration, not evidence that AI will deliver that outcome.

The local costs of AI infrastructure

Feldman also connected some public hostility toward AI to the data-center buildout. He said the industry had not done enough to engage local communities in recent years, particularly in the US, or involve residents in discussion the way energy companies in oil and gas do. He called that a problem.

He contrasted the American “not in my backyard” attitude with countries he said encourage data-center construction:

That points to a less abstract source of opposition than fears about runaway models: people are also reacting to how the infrastructure behind AI is built and discussed.

Efficiency is not a demand killer

Feldman does not expect more efficient computing to reduce demand for Cerebras chips or equipment from Nvidia’s competitors. Hardware makers can improve computing efficiency while model developers make models more efficient, allowing them to achieve the same results. Feldman argued that the two efforts should complement each other and produce substantial progress together.

I think the harder question is not whether efficiency improves, but who benefits from it and who bears the costs of building the infrastructure. Feldman’s case for discussing AI’s promise will be more convincing if the industry pairs it with meaningful engagement where those data centers are built.