Korva is a scroll of the internet’s best ideas.
Ask why AI can't work for months
Why pattern matching fails
Pattern matching works about 99 percent of the time: it has always been like this, so it probably still is. The other 1 percent is a period when something real has changed, and those are the moments when thinking from first principles beats the pattern.
The METR numbers
A report from METR put numbers on the change. A couple of years ago, AI could do about 10 to 20 seconds of human work before it needed guidance or made a mistake. That length has doubled every couple of months, and it is now hours. Describe a task that would take a person hours well enough, and the AI does the whole thing and comes back with the result.
The next question
The first-principles question is why the length can't keep doubling into days, weeks, and months. Then comes the next one: what does the world look like when everyone has an agent that can do months of work at a time? That world looks nothing like the last several years, and most of the road there will probably be covered in about five years, which in AI terms is a century.
How does METR measure how long AI can work?
METR is a nonprofit that evaluates the capabilities of frontier AI models. It gives models tasks that range from seconds to many hours long, and it measures task length by how long skilled humans take to finish them. The length of task a model can complete at a given success rate becomes its time horizon. Plotting that figure for successive models shows the trend over time.
Why does pattern matching work so often?
Most of the time the near future resembles the recent past, so using past outcomes as a guide is a sound strategy. Forecasters call this the base rate, and starting from it is a standard habit in good prediction. Its weakness is that base rates carry no information about a change in the underlying conditions. Forecasting researcher Philip Tetlock found that the best forecasters start from base rates and then adjust when conditions shift.
What would a month-long AI task require?
A task that long requires an agent to keep track of its goal and its progress across far more steps than fit in a single working memory. Systems handle this by writing notes to files, breaking work into subtasks, and checking results along the way. Each step carries some chance of error, and over thousands of steps those chances multiply. Reliability per step is therefore the main limit on how long agents can run.
More from David Senra
