Karpathy Joins Anthropic: Why This Is Important For AI Visibility

Executive Summary

Andrej Karpathy just left OpenAI and Tesla to lead pre-training research at Anthropic. His job is to make Claude smarter about what it learns, which means Claude will get pickier about which businesses it trusts and recommends. If your business isn’t already well-represented across the web, you risk becoming invisible in AI answers.

The window to fix this is closing fast. Run the AIBAI Diagnostic to see what Claude, ChatGPT, Gemini, and Perplexity are currently saying about your business.

Introduction

On May 19, 2026, the most respected AI researcher alive, Andrej Karpathy, chose Anthropic. As a co-founder of OpenAI, Karpathy built Tesla’s Full Self-Driving training system from scratch. He could fund a startup, advise anyone in tech, etc. And yet, he took a research role at Anthropic instead.

Most coverage views this move as a talent war story: OpenAI loses a legend, Anthropic wins a name, and the AI race keeps going. Although that read is correct, it misses the point entirely for anyone running a business in 2026. The key thing is that Karpathy is going to make the AI smarter about what it learns. That sounds abstract until you realize the filter deciding which businesses show up in AI answers just got a major upgrade. If you’re not already on the right side of that filter, read more to find out how to increase your AI visibility.

What Is Pre-Training and Why It Determines Your AI Visibility

Pre-training is where a model like Claude gets its basic knowledge of the world. This knowledge is extracted from a massive learning process that happens before Claude ever talks to a single user, i.e. the model’s education. Everything Claude knows when you open a chat window was shaped during pre-training.

This is where Sediment Engineering lives at its deepest level. The content that survives the pre-training filter becomes the ground that AI answers are built on. A law firm with rich, structured, AI-readable content layered across its website, its YouTube channel, its directory listings, and its blog is building sediment. On the other hand, a firm with a twelve-page website and a blog that hasn’t been touched since 2023 is only sitting on the surface.

Karpathy understands this layer at a mechanical level. For isntance, he built the training system that taught a fleet of cars to see the road. He knows how large-scale learning works under real conditions, and where data quality creates leverage versus where it creates noise. And now, he’s building a loop where Claude helps design its own training, with each pass through that loop producing a smarter sieve. It will be a smarter sieve that raises the bar for everything that wants to get through it.

How Karpathy’s Pre-Training Research Loop Affects AI Search Rankings

Karpathy is leading this new team, whose job is to use Claude to make AI research faster. This helps Claude get smarter, faster, and informs every model of Claude following it.

A smarter AI filter works better by catching weak content right away. So, businesses businesses built on weak content will not appear. This is what I call the Velocity Paradox.. I see it in almost every AEO conversation I have. What happens is business owners are moving fast by posting content, updating pages, and running ads. However, they’re moving fast in the wrong direction. They’re still trying to please a search crawler that’s fading away, whereas the real competition has shifted to AI answers. Most importantly, Karpathy’s work speeds this shift up.

Rankfossil — AEO Key Word

A Rankfossil is an old search strategy. It worked well back when Google ruled search, but it doesn’t work anymore for AI answer engines. It’s still around. But it doesn’t help you compete. The businesses that are still using Rankfossils are falling behind.

What Is the Dark Pipeline and Why Karpathy’s Hire Makes It Worse

The Dark Pipeline was already flowing before Karpathy showed up. Your prospects, especially if you run a professional services firm, have been asking Claude, ChatGPT, Gemini, and Perplexity for recommendations for over a year. Then, the prospect reads the answer, makes a call, and either puts your firm on a list you never knew existed or cuts you before you knew the game had started. And all this happens without leaving a trace in your analytics.

This is what I call the Dark Pipeline: revenue you can’t see because it doesn’t produce a trackable signal. As Claude’s pre-training gets smarter under Karpathy’s direction, the answers Claude produces become more trusted. A more trusted answer engine means the Dark Pipeline carries more weight in how buyers decide. Thus, being absent from those answers costs more every quarter.

Some relevant statistics

What Is a Ghost RFP and How AI Pre-Training Changes Who Gets Found

For example, a general counsel at a mid-size company needs outside legal help. She opens Claude and asks about litigation firms in her region with the right practice area. Then, Claude answers, naming firms with confidence to go on her short list. The counsel then sends three emails, none of which your firm gets. This is what I call the Ghost RFP.

This is happening today in your market. The Ghost RFP doesn’t care about your bounce rate or your Google ranking. Rather, it cares about how Claude has encoded your firm’s authority, practice areas, and credibility signals. And with Karpathy, they’re about to make Claude significantly better at forming those encodings. That makes the Ghost RFP both more powerful and more selective about who survives it.

What Is the Hallucination Tax and Why Smarter AI Models Raise the Cost

Usually, we assume a smarter model means fewer hallucinations, which means less risk for their business. However, this is only partially true because a smarter model is also more confident in what it knows. When your firm isn’t well represented in the training data, a smarter Claude does not hedge. It either omits you in favor of a competitor it actually knows, or it fills the gap with something it made up and delivers it with the same confidence it uses for real facts.

The Hallucination Tax is the total cost of being omitted, underrepresented, or described inaccurately across every AI-powered evaluation that is mentioning your category. And as Claude’s pre-training improves, the businesses that the model cannot find or cannot accurately describe pay more in lost pipelines.

So, if you don’t know what Claude currently says about your business when asked a category query, you’re in the pre-ghost. Despite decent Google ranking and website traffic, in the AI answer layer, your prospects are already being routed to someone else. That is the most expensive place to be, because you cannot measure the loss and cannot make the business case for fixing it.

What Is Memory Engineering and How to Apply It to Your AEO Strategy

When Karpathy announced his move to Anthropic, he said he wanted to get back to research during what he called a formative period. That word “formative” means the patterns being set right now will shape what the model knows for its entire operational life.

Memory Engineering is exactly what you do to take advantage of that window. This method builds content and structures it in ways that match how AI systems form and retrieve knowledge. Basically, structured data that gives the model clean factual anchors. So, use answer-optimized prose that matches how your prospects ask questions.

Why the Window to Build AI Visibility Is Closing Faster Than Ever

I’ve said for two years that this window to establish category authority in AI answer engines was never infinite. Cement never stays wet! The window has been closing steadily as the major models improved and as the volume of well-built AEO content in the training pipeline increased. And now, Karpathy’s arrival at Anthropic puts the best hands in the business on the machine that sets the concrete. Karpathy chose Anthropic because he believes the next few years are the most important in the history of AI development.

The question for your business is whether those years find you on the Category Island or still swimming in the Category Ocean, invisible to the answer engines that are already deciding who gets the call.