AI Loops Will Climb Your Hill, Then Get Stuck There
An investor argues AI is best at squeezing the hill it is on, and people are needed for the leap to the next one.
Lenny's Podcast: Product | Growth | Career · Why companies are becoming a series of loops | Anish Acharya (a16z)Anish Acharya, an investor at the venture firm Andreessen Horowitz, was asked how company building will change as artificial intelligence spreads through every department. His answer was that AI loops will do the optimizing, while people will make the leap.
“The loop will help you climb to the local maxima, but then it plateaus,” he argues. By local maxima he means the best you can get by making small improvements to what you already do. “You need somebody to actually help you land at the base of the next hill.”
By loops he means something specific. Prompts led to agents, which Acharya describes as “models in a loop with tools and memory and skill files.” Sets of agents doing tasks become loops, and he expects them to cascade, from a loop per person up to “loops that can run large parts of the company.”
Coding shows the optimizing power. A bug report comes in, a reproduction gets generated, a fix gets created, and the fix gets reviewed. If the risk is high, a human confirms it is okay to ship, and if the risk is low, it just gets shipped. It may even email the customer to say the reported bug was fixed. That happens in five minutes.
Growth shows the limit. He describes the old brainstorm of listing possible experiments, prioritizing them, building them, shipping them, and measuring them. The loop version generates every variant and measures every variant, converges and ships the winner once the results are statistically reliable, keeps a long term holdout, and starts on the next experiment. It climbs the hill fast, then stalls.
In his strongest form of the idea, loops in coding, marketing, sales, support, and legal feed a larger business loop that could tell the chief executive that strategy or the business model needs to change. Humans stay critical in that company for sales, support, strategy, and exceptions, because thinking that goes beyond patterns the models have already seen is still limited. As he puts it, you still need a person to say, “Here’s the thing I think we should make,” and be right about it.
Questions people ask
What makes an AI agent different from just asking a chatbot a question?
An agent is AI software that works toward a goal by using tools, remembering context, and following instructions over several steps. A loop turns that into a repeatable workflow that runs every time it is triggered, across coding, marketing, sales, support, and legal.
Why is AI good at optimization but weak at the leap?
Current systems are strong when they have clear feedback, like bug reports or test results, and can iterate quickly. They are weaker at reframing the problem, noticing that customer desires changed, or betting on an unproven idea with no data behind it yet.
What does it mean to converge once results are statistically reliable?
Statistical significance means the result is unlikely to be due to chance. Teams keep testing variants until the math says a winner is real, then ship that winner while holding back one group to check long term effects.