What Happens When AI Can Improve Itself — And What It Means for Your Business

Earlier this year, AI researcher Andrej Karpathy sat down with Sarah Guo on the No Priors podcast to talk about the current state of AI agents, autonomous research, and where work is heading. It's one of the most grounded conversations on the topic we've come across — and several of the ideas have direct implications for how B2B companies should be thinking about their operations right now.

If you run a B2B company in Europe, you've probably noticed that AI conversations tend to fall into two camps: breathless hype about AI doing everything, or dismissive skepticism that nothing has really changed. Neither is useful.

What's actually happening is more specific — and more actionable — than either camp suggests. Let's walk through the ideas from this conversation that we think matter most for founders and operators.

The shift that happened quietly in December

Andrej Karpathy is one of the most respected AI researchers in the world. He was a founding member of OpenAI and later served as Director of AI at Tesla. He's not someone given to exaggeration.

So when he describes December 2024 as the moment something fundamentally changed — when he went from writing most of his own code to barely typing a line — it's worth paying attention.

What changed wasn't the AI getting smarter overnight. What changed was that AI agents — software that can take a task, work through it autonomously, and come back with a result — became reliable enough to delegate real work to.

He describes spending his days not coding, but directing: assigning tasks to multiple agents running in parallel, reviewing their outputs, and feeding the next set of instructions. The analogy he uses is a manager dispatching work to a team — except the "team" runs around the clock and costs a fraction of what human execution would.

For B2B operators, the relevant question this raises is: which parts of your business could run like this?

The bottleneck has moved — and it's probably you

One of the most useful reframes in the interview is about where the constraint in knowledge work now sits.

For years, the limiting factor was access: access to technical talent, to data, to tools. Most companies didn't have the resources to build sophisticated systems, so they worked around it.

That constraint is collapsing. The new bottleneck, Karpathy argues, is the person at the top of the workflow — the one who needs to review, decide, and feed the next instruction. If you have to be present for every step, you cap the throughput of the entire system.

The implication for founders: your job is increasingly to design systems that don't require you at every step. Define the goal clearly. Set the boundaries. Establish how success gets measured. Then let the system run.

This isn't new advice in operations. What's new is that the systems you're designing can now include AI agents doing work that previously required people.

AutoResearch: the idea that illustrates where this goes

The most striking segment of the interview covers a project Karpathy calls AutoResearch. The concept is straightforward: instead of a researcher manually running experiments, reviewing results, and deciding what to try next — an AI agent does the whole cycle, autonomously, on repeat.

He tested this on an AI training project he'd spent years optimising by hand. He let the agent run overnight. It found improvements he had missed.

His conclusion: "I shouldn't be the bottleneck here. There are objective criteria. Just let it run."

Now, you don't need to be training AI models for this principle to apply to your business. Think about any process in your company that follows this pattern:

1. Run something (a campaign, an experiment, a test)
2. Measure the result against a clear metric
3. Make a change based on what worked
4. Repeat

Outbound sales sequences work like this. Google Ads campaigns work like this. Email marketing works like this. Lead scoring works like this.

Anywhere you have a clear goal, a measurable outcome, and a repeatable process — that's a candidate for closing the loop autonomously. Not just AI-assisted, but AI-driven, with you reviewing the direction rather than executing each step.

The one condition that makes or breaks autonomous systems

Karpathy is careful to flag the key constraint: this only works where the outcome is objectively verifiable.

If you can measure whether something worked — a reply rate went up, a conversion improved, a cost per lead came down — an agent can optimise for it without needing your judgment on every iteration. If the output is subjective — does this feel right, is this the correct tone, does this match our brand — a human needs to stay closer to the loop.

This is a useful lens for deciding where to start. Look for the parts of your operation where success has a number attached to it. That's where autonomous systems create the most leverage with the least risk.

What this means for B2B teams specifically

Karpathy's analysis of the jobs market makes a useful distinction: work that manipulates digital information is changing fast. Work that involves the physical world — field sales, events, logistics — changes more slowly.

For most B2B founders reading this, the core of your go-to-market is digital: prospecting, outreach, content, advertising, reporting. That's the zone where the change is already underway.

A few areas where the implications are most direct:

Outbound and lead generation. Building prospect lists, enriching contact data, writing personalised sequences, testing messaging — each of these has a measurable output. The companies building structured, repeatable systems around these activities are compounding their advantage over those still running it manually.

Paid advertising. Campaign performance is highly measurable, which makes it well-suited to tighter feedback loops. The question isn't just "are we using AI features in Google Ads" but "have we structured our campaigns to learn and adjust faster than our competitors."

Reporting and decision-making. The value of a good operations setup isn't just the data — it's the speed at which you can turn data into decisions. Agents can help collapse that gap significantly.

What still requires human judgment

The interview is honest about where current AI falls short. The capability profile of today's models is uneven: exceptional at structured, verifiable tasks; inconsistent on anything that requires nuance, context, or reading between the lines.

Karpathy describes it as feeling like you're dealing with "an extremely brilliant systems engineer and a 10-year-old at the same time." Highly capable in specific domains, surprisingly unreliable in others.

For founders, this means the soft parts of the job remain yours: reading a client relationship, sensing when something is off in a deal, making judgment calls in ambiguous situations. These are not things to hand to an agent and walk away from.

The practical approach is to be deliberate about the division: automate the structured, measurable work aggressively, and protect your judgment for the parts that actually require it.

Where to start

If you're a B2B founder thinking about how to act on any of this, the most useful question is probably not "how do I implement AI across my business" but a narrower one: what is one process in my company where I could define a clear goal, measure the outcome, and remove myself from the execution loop?

Start there. Build the feedback loop. Measure what changes.

That's what the companies pulling ahead are doing — not replacing their entire operation overnight, but systematically closing loops that were previously open.