Korva is a scroll of the internet’s best ideas.
Task horizons are outgrowing release cycles
Models work longer
New frontier models now arrive at most every two months, sometimes faster, and the models keep stretching how long they can work. A model can already do a week-long task. Month-long tasks are probably next, then three-month ones.
Evaluations fall behind
Safety checks before a release have assumed since the GPT-4 era that evaluations fit into a short window. If a model stays on task for three months, the evaluations cannot run at that length before the next release arrives. Capabilities could degrade over that span, or alignment, or safety behavior, in ways nobody has had time to test.
The cost of slowing
For many companies, safety policies written in the GPT-4 era, when long horizons were on nobody's radar, have not been updated for agents that run this long. Slowing releases brings its own problem, a wider gap between what the labs use internally and what the outside world can get. Math already shows it: a powerful internal model is solving unsolved problems that the public cannot use.
Why did GPT-4-era safety policies assume short tasks?
Anthropic published its Responsible Scaling Policy in September 2023, and OpenAI released its Preparedness Framework that December. Both tie a model's release to capability evaluations run beforehand. These documents were written when GPT-4, released in March 2023, was the frontier, and the tasks models handled lasted minutes. Evaluations of that kind can be run and reviewed in days or weeks.
How does a model stay on task for months?
A model can only read a limited amount of text at once, so long-running agents keep state outside their working memory. They write notes and progress files, summarize earlier work, and reload only what the next step needs. This lets a task continue long after the original context would have overflowed. Context windows grew from 8,000 tokens in the original GPT-4 to over a million in some later models.
More from Dwarkesh Podcast
