The fake egg problem

Hi, and happy Tuesday.

Just a year after coining the term “vibe coding”, AI leader Andrej Karpathy has again hit the zeitgeist by popularizing another familiar concept with a new frame - coining it “auto-research”. 

If you've felt anxious that you or your organization is falling behind on AI - but couldn't quite articulate what "behind" even means when you're in an industry with real protections - such as physical assets - this post is for you.

🔔 We've built two interactive applications to simulate these ideas. 

See these ideas in action here: AI Optimization Loop Simulations

Before we get to “auto-research”, we need to talk about a bird.

In the 1950s, a biologist named Niko Tinbergen built a fake egg.

It was bigger than a real egg. More colorful. Exaggerated in every way. He placed it next to a mother bird's actual eggs to see what would happen.

She abandoned her real eggs and sat on the fake one.

The bird's instinct was simple: sit on the biggest, most colorful egg. That rule had worked for millions of years as a way of ensuring the survival of her species.

Bigger and brighter really did mean healthier offspring - in nature.

But Tinbergen's plaster egg broke the rule. It was bigger and brighter than anything nature could produce. And the bird's instinct didn't say "wait, that's suspicious." It said "that one. Definitely that one."

Biologists called this a supernormal stimulus - an artificial thing that triggers a stronger response than the real thing it copies.

Every optimization loop at scale has already produced its own fake egg.

You already know social media companies measure engagement: dwell time on each post, clicks, shares. Their algorithms test what keeps you scrolling, keep what works, toss what doesn't. The result? They often push hostile, emotionally charged content. Users don't even prefer the posts shown to them. The platforms optimize for addiction over preference.

But this optimization incentive exists everywhere. 

For example, every company is intrinsically optimizing for repeat purchase - how often you buy again - against what the company sells. Even these loops can have unintended consequences. 

• Ultra-processed foods engineered for addictive repeat consumption. UK childhood obesity has risen 700% in 30 years. 

• Pharmaceutical companies optimizing for prescription volume led to the opioid crisis - drugs were marketed against physician adoption rates, not patient outcomes

• Plastics producers optimized for unit cost and versatility - microplastics are now found in human blood

These unintended consequences often leave companies scrambling for solutions. In fact - some of the projects we undertake at PreScouter are to help clients find new solutions that don’t have these negative side-effects - whether that is carbon emission reduction, sustainable packaging that preserves shelf life or even lower calories.

But, until now, these loops were bottlenecked by supply. 

Social media could only optimize over content that humans uploaded. The food industry could only test recipes that food scientists had time to formulate. Pharma could only test as many compounds as chemists could synthesize. Mining could only evaluate as many site configurations as engineers could model.

That bottleneck is now gone.

An AI agent can generate the supply - the content, the copy, the product variation, the ad creative, the pricing structure, the onboarding flow - and test it. The optimization loop no longer waits for humans to produce the raw material it runs on.

Think about what that means for your industry. 

• A CPG company generating and testing thousands of packaging designs against purchase intent - finding the combination of color, shape, and label copy that maximizes impulse buys regardless of nutritional value.

• A food manufacturer cycling through flavor formulations optimized against craveability scores. 

• An energy company testing rate structures against customer retention while obscuring total cost.

• A chemicals company optimizing safety data sheet language against regulatory pass rates. 

• A life sciences firm testing physician outreach messages against prescription volume.

All these loops just require a goal to optimize against and a machine that can run the loop. This machine - the AI - generates new variations, tests them against the goal, learns from the results and produces the next set of variations in the loop.

Andrej Karpathy’s auto-research is this loop pattern applied to optimizing an AI algorithm: 

• An AI agent creates variations of the algorithm it is improving

• It tests the algorithm’s output against a goal outcome

• Where a variation improves the goal outcome, it uses that as the base for the next set of variations it creates

It’s a continuous optimization loop.

For entirely digital loops - such as algorithm optimization, social media feeds, e-commerce ads, email sequences, dynamic pricing, chatbot scripts - the cycle has already been running fast. Our inboxes are full of spam, social media feeds are full of “slop” and even shopping on Amazon can be addictive.

Now, with AI driving supply, the loop accelerates. 

Decades ago, a well-crafted marketing email was a competitive advantage. Now AI generates thousands of variants overnight, and open rates are collapsing industry-wide because every inbox is flooded with competent-but-generic copy..

Value moves from running the loop to designing the loop.

But the pattern doesn't only apply to digital. Evolution itself is an optimization loop that ran for billions of years. Resistance to antibiotics is an optimization loop. Corporate strategy is an optimization loop - just one that runs quarterly instead of hourly.

The difference now is that AI compresses any loop where you can define a goal outcome even when the loop itself is slow. A clinical trial still takes years. A factory run still takes days. A crop cycle still takes a season. AI can't speed up the physics. But it can generate hundreds of variations to test per cycle instead of the three or four a human team would design. 

A manufacturer that used to test five production line configurations per quarter can now test fifty. A life sciences company that used to enter a trial with two formulation candidates can enter with twenty. The clock doesn't change. The number of bets per tick does.

Why “falling behind” is a risk.

Optimization loops coumpound.

The company that figures out how to apply this pattern to the loops that drive their industry doesn't just get a one-time efficiency gain - they get faster at getting faster. Every cycle produces learning that improves the next cycle. A competitor who starts six months later isn't six months behind. They're hundreds of cycles behind. 

And here's what makes this worth getting right rather than getting scared of: the same pattern that produces fake eggs can produce real ones. A drug that works better with fewer side effects is a real egg. A manufacturing process that cuts waste and defects simultaneously is a real egg. A marketing campaign that actually gets people to the value faster is a real egg. The difference is never the loop. It's always the goal metric you measure against.

The question to discuss with your colleagues:

Where could you use this pattern today? Look for processes with clear inputs and measurable outputs. Prioritize those with faster feedback loops or where the cost of creating and testing each variation is coming down. Idea generation. Formulation testing, packaging design, route optimization, supplier qualification, clinical trial design, maintenance scheduling, claims processing, demand forecasting, safety protocol iteration, customer complaint resolution workflows. These are loops waiting to be designed.

Tinbergen's bird didn't make a mistake. Her instincts were perfectly tuned - for a world where nothing could out-optimize nature.

That world is over.

The companies that win the next decade will be those that can create loops that optimize for real eggs.

🔔 We've built two interactive applications to simulate these ideas. 

See these ideas in action here: AI Optimization Loop Simulations

Let us know if we should do more of these types of simulations.

Best,

Dino