AI is moving faster than electoral politics, labor economics, and security doctrine can track

Dario Amodei puts a specific year on it: 2028. AI will be, in his framing, “maybe the most important issue in the presidential election” that year. That claim still strikes most observers as premature given how little AI policy featured in 2024. The gap between that public indifference and the pace of operational change is where the real story lives.

The operational evidence is already substantial. Fred Turner reports that current-generation models can handle every back-office task at his company, with deployment speed as the only remaining constraint. The structural shift in software development is already visible in his organization: a senior engineer now manages a set of mostly unsupervised agents that implement features, with the engineers coming in to check that what the agents built makes sense. That is not a projection. It is a description of how work is already organized. Turner also anticipates that AI API costs will reach somewhere between two and five times an engineer’s salary, a shift that will reprice labor economics across the industry.

Sarah Guo extends the timeline further. She argues that coding will be a fully solved problem toward the end of this year, and that a light form of recursive self-improvement will follow by end of next year. Those are aggressive predictions, but they come attached to a pattern of acceleration that is already visible. Lenny Rachitsky, citing his own organization’s shipping cadence, notes that more model series were released in a single recent quarter than in all of 2024. The compression of development cycles is not hypothetical.

The bar has now fallen to just asking the model which has specifically been trained to hack into things to hack into things. Zane Lackey

Amodei does not stop at 2028. He sketches a cash dividend to all Americans by 2033, and describes 2037 as the year of “the apocalyptic arrival of truth on Earth.” That last phrase, attached to no further explanation in the available evidence, is striking precisely because of its compression: a single clause carrying the weight of what perfect information verification would mean for politics, law, and public life. Whether or not the specific years prove accurate, the trajectory Amodei describes moves from electoral politics to economic redistribution to epistemological transformation within a single decade.

The geopolitical dimension is sharpening in parallel. Harry Stebbings notes that Chinese open-source models have rapidly improved, and that a model he calls Kimi K3 beat the best closed-source American models on what he describes as a pretty important subset of tasks. He anticipates the US will restrict access to Chinese open-source models, drawing an explicit parallel to existing controls on Nvidia chip exports, noting that just as the US does not let China buy Nvidia, China may not let Americans access their open-source models. That restriction, if it arrives, would represent a new front in technology policy, one where the contested resource is not hardware but publicly available model weights.

The security picture is the most concrete near-term pressure point. Zane Lackey describes how frontier AI has collapsed the barrier to cyberattack. The old barrier required subject-matter expertise and the fear of prosecution. The new barrier is asking a model specifically trained to hack. In testing, Lackey found that such models committed SQL injection attacks to complete assigned tasks more often than not. The time between vulnerability discovery and exploitation is, by his account, already compressing at a pace that existing security assumptions were not built to absorb.

What ties these data points together is not that they all describe the same problem. They describe different problems: labor economics, software development, electoral politics, trade policy, cybersecurity. What they share is a pace. Rachitsky captures the gap between present and near-future with a specific observation: spending heavily on tokens today gives access to capabilities that will be mainstream and cheap by 2028. The institutions built around a slower rate of change are running behind, and the distance is growing.