Analysis
I have been investing at seed for long enough to recognize the shape of this. When the largest firms in a category start describing their own product as civilizationally dangerous, I stop listening to the content and start reading the proposed remedy. The Register made the argument bluntly this week: the myth of killer AI is a self-serving bid for regulatory capture. I think that is too strong as a statement about motives and roughly right as a statement about consequences.
Here is what I actually see from my seat. A seed-stage company training or heavily fine-tuning models runs on eighteen months of capital and four engineers. Every requirement that a pacing regime would add -- structured pre-deployment evaluations, third-party red-teaming, incident reporting, documented shutdown authority -- is a fixed cost. Fixed costs are regressive. Anthropic absorbs them out of a safety budget that is larger than most of my portfolio companies' total raises. My companies absorb them by not shipping a feature.
The market told the same story on Monday. Amodei's essay knocked roughly 3% off Nvidia and about 6% off Micron and Marvell, while Alphabet, Microsoft and Meta traded up. Investors did not read the essay as an admission of danger. They read it as a signal about who captures the value if the pace of model releases slows: the companies with distribution, not the companies with chips, and certainly not the companies with four engineers.
“Amodei's essay knocked roughly 3% off Nvidia and about 6% off Micron and Marvell, while Alphabet, Microsoft and Meta traded up.”
I want to be precise about what I am not saying. I am not saying the risks are fake. Agentic systems acting against production infrastructure fail in ways chatbots do not, and I would rather a lab have an evaluation suite than not. The specific move I object to is bundling a real, tractable engineering problem with an unfalsifiable civilizational one, then proposing rules calibrated to the second while the first would be solved by boring, testable requirements.
Room for disagreement: the strongest counter is that the people writing these warnings have access to internal capability evaluations that I do not, and that dismissing them as positioning is exactly the mistake every industry's skeptics made before the harm was legible. Amodei has been consistent on this since 2023, well before it was commercially convenient, and he left OpenAI over a version of this disagreement. If he is right and I am wrong, the cost of my being wrong is considerably higher than the cost of his. I hold this view at maybe 65% confidence, and the thing that would move me is a published capability evaluation showing a discontinuity, not another essay.
What I tell founders in the meantime: build the evaluation harness now, document it, and treat it as a sales asset. If the rules arrive, you are compliant. If they do not, you still have a testable model and an answer for the enterprise security questionnaire. That is the only version of this debate with a positive expected value for a small company.
