Korva is a scroll of the internet’s best ideas.

Beyond one-off tasks

Most AI products take a prompt, do the task, and report what happened. Instinct treats a task as part of something longer, so a standing objective is natural to give it.

Goals people set

Users already work this way. They give it a number of pounds to gain or lose, or a mile time to hit, and ask it to help them reach it over the next three or four months.

Running a business

Some small businesses run entirely on it, with the back office built on Instinct and parts of it fully autonomous. One standing instruction tells it to keep an inventory level from dropping below one number or rising above another. The assistant uses its tools and devices to pursue that, with no one telling it to order a particular item.

Why do specific goals work better?

The psychologists Edwin Locke and Gary Latham spent decades studying goal setting and summarized the results in their 1990 book on the subject. They found that specific, difficult goals led to higher performance than vague instructions such as do your best. A specific goal tells a person what counts as progress and when to adjust. Their work became the basis of goal-setting theory in workplace psychology.

How does an inventory band work?

Operations researchers call a rule like this an (s, S) policy: when stock falls to a lower level s, order enough to climb back to an upper level S. The upper limit prevents money from being tied up in unsold goods, and the lower limit prevents running out. Herbert Scarf proved in 1960 that policies of this form are optimal under certain conditions. Retailers and manufacturers have used versions of it for decades.

How can an agent work toward a goal for months?

A language model can only hold a limited amount of text in a single run, so long-running agents keep their state outside the model. They write notes, plans, and records to storage and read them back on each wake-up. Scheduled triggers start each run, much as cron jobs have started programs on a timer in Unix systems since the 1970s.

More from Invest Like the Best with Patrick O'Shaughnessy

Tech & AIProactive agents need far more computeAn assistant that wakes and sleeps on its own spends far more tokens than one waiting for prompts, so its compute needs run orders of magnitude higher.
Enjoyed it?