He doubled a company with AI — then asked the labs to pause hyperscaling

This Cognitive Revolution episode is a crossover. Nathan Labenz, founder of the small-business video company Waymark and host of the show, sits as the guest on the Nick Halaris Show. Halaris introduces him as the world's most prominent AI scout. They share a Michigan high school, Chippewa Valley, and two former teachers who made the introduction. Labenz frames the hour as a wide scan: what AI can already do in medicine, law, and automation, where control could fail, and what the U.S.–China contest does to the odds.

Two years earlier, making a 30-second commercial on Waymark still meant browsing a template library, inventing copy, picking photos, and waiting days for a human voiceover. Even when the tools were easy, customers told him the blank page was the real wall: they did not know what to say. After generative models arrived, the loop flipped. Paste a business URL, type ten or twenty words, and in about 30 seconds a full 30-second spot comes back — script, images, and voice already assembled.

Waymark doubled in 2023 while breaking even. Labenz then handed the CEO job to a longtime COO and went full-time on trying to make sense of the field. He describes himself, later in the conversation, as an adoption accelerationist and a hyperscaling pauser: take the useful stuff as fast as possible, and get much more cautious about building systems that will be more powerful than we are.

Labenz dates his interest to 2006 or 2007, when he found Eliezer Yudkowsky's writing. Yudkowsky was already arguing that compute and datasets were on a path to powerful AI, that no one knew how to control it, and that transmitting human values would be hard. Labenz found it interesting and filed it with asteroids: probably unlikely, catastrophic if it happened, worth having someone scan the sky. He sent small donations to people doing that speculative work. He did not think he had the mathematical skillset, so he spent the next fifteen years on entrepreneurship.

The threads met around 2020–2021. Waymark was a browser DIY video tool for small businesses that could not master Adobe. User interviews kept hitting the same sentence: the product is easy, I just do not have anything to say. GPT-3, which he had followed since GPT-2, looked like a way off the blank page. He cites a *Simpsons* beat: Mr. Burns at a gallery saying he is no art critic but he knows what he hates. Early AI, he decided, lives in that zone. Customers are not prolific creators. Show them drafts until they latch onto one they do not hate, then let them edit.

He yanked the company toward generative AI and then got obsessed with understanding it, because the technology is weird enough that product quality depended on that understanding. Computer vision was arriving in parallel, which mattered because users sat on folders of photos they did not want to sort. Text-to-speech was catching a voiceover service that took days, cost about a hundred dollars, and billed again for revisions. The realization that stuck was not unique to him: the same core technique, the Transformer, was driving all of those jumps at once. A single architecture that can write, see, speak, make music and images, understand video, and hold a real-time conversation starts to look like the powerful thing Yudkowsky had been writing about.

Halaris asks him to close the loop on the business. Labenz says 2023 was night and day: they doubled while breaking even, without a recent venture round, on a trajectory much better than before. The old flow was browse templates, face the writing problem, match images, request a voiceover, wait maybe two days, listen, maybe send it back. An hour to a finished video was already fast compared with hiring an agency. An hour was still a lot.

Today a user can give Waymark a homepage URL. The system spiders the site, pulls copy and imagery, and presents a single text box. Type that you have a new location, a weekend sale, or just want to introduce the business — ten, fifteen, twenty words — and in 30 seconds a fully formed 30-second commercial comes back, script, selected images, and generated voiceover included. You can watch a production for your own company before a minute has passed, hate it in the Mr. Burns sense and regenerate, or keep the draft and use the old detailed editor. Almost everyone still makes some low-level edit. The starting point is completely different.

The qualitative change is the reveal. Previously there was no moment of opening a box, because the user had made the thing. Now they can open as many boxes as they want and continue from the one that intrigues them. Speed, feedback scores, videos created, sharing that brings in friends, and revenue all moved. Labenz says it is a genuine multi-modal application of generative AI that still misses sometimes and, most of the time, delivers.

Asked for the upside after four years of deep dive, Labenz says you almost have to go sector by sector, and that Waymark's before-and-after is a decent guide to everything else. Medicine is the example he reaches for first. U.S. healthcare can do amazing things; the day-to-day is appointments, waits, a four-hour urgent-care trip with his kids, uneven access, and cost. The upside is an AI doctor. In blind chat setups, a user talks to either a human or an AI, then human doctors score who did better at figuring out the condition. On core tasks such as differential diagnosis, the AI systems already outperform human doctors by a margin that is not small, at less than 1% of the cost — roughly ten cents against a hundred-dollar visit, which he calls 0.1%. They are always available. You can drop off and pick the conversation back up.

That does not mean they can do everything a physician does. Long trajectories across a life are still a weakness, though people are working on it. Imaging is getting competitive with radiologists, in his view, without a clean claim that it is better yet. On the research side he points to AlphaFold 3 and the family of models from Google DeepMind and others: take a DNA sequence, predict the three-dimensional shape of the protein. Proteins are the little machines in the body; disease is often a shape that no longer fits. Figuring out those shapes used to take huge experimental work. A new class of tools is going to accelerate discovery in biomedicine.

Law gets the same dual-use advice he gives himself in a health emergency: do not put all the eggs in the AI basket, and do not go without it. If you are guaranteed an attorney, he says, you should probably also be guaranteed GPT-4. His own contract habit is to send the same document to ChatGPT, Claude, and Gemini with the same question — what are the most concerning things I should think about — and sign if none of the three raises a serious flag. Tesla's latest full self-driving is getting really good. A food-robotics company told him that 70% of manufacturing labor cost, excluding ingredients, is putting food into packaging; prep and cooking are only 30%. Broccoli pieces vary in size and weight. Their robot has plated on the order of a million meals. Google published a shipping system that, the company said, can double a carrier's profitability and move more tonnage with 15% fewer ships.

The big picture he wants is abundance, freedom from drudgery, and broad access to expertise at under 1% of today's price, on a phone. Energy complaints about AI, he thinks, are overblown once you count the trip to the doctor you did not take. What would make it even stranger is closing a loop that is still open: generalist systems like ChatGPT that are pretty good at everything and great at nothing, sitting beside narrow superhuman tools like AlphaFold and the shipping optimizer. If the generalist can use those tools, or design its own, all bets are off — possibly in a good way, and also in the sense that highly autonomous science we struggle to understand is how things get out of control.

Halaris hears a deflationary boom and asks the obvious follow-up: what happens to the people in those jobs, and do we need a new social contract. Labenz says even AI bulls have under-theorized this. He is privileged to do work he would study unpaid. That is not the case for someone scooping food into trays at 20 scoops a minute, a scoop every three seconds, often in a room around 35 degrees Fahrenheit. He thinks the vast majority would take a deal that met their needs and ended that shift. Meaning-from-work is his problem more than theirs. The standard answer from technologists is universal basic income. Sam Altman, to his credit, has been personally funding UBI experiments for years. Labenz does not think that is nearly as much as society may soon need.

He caves everything with a 10% chance, not less, that progress stalls. The inputs are data, compute, and algorithms. For about 12 years the field has scaled the first two while algorithms got more efficient at turning them into capability, with no clear sign of slowing. A hard physical barrier is possible; most people inside the field do not expect one. If the curve holds, it will not take many more generations of models to enter a different regime. GPT-2 was barely coherent. GPT-3 could write Waymark scripts. GPT-4 passes the bar exam. Systems now pass medical licensing tests and beat doctors on diagnosis. There is not that much more room before the world five years out looks extremely different — and we are not yet at the point where everyone can quit.

Past transitions took careers. Electrifying America took about 60 years from Edison's bulb to his grandmother in rural Kentucky getting power. A blacksmith could finish a working life. This disruption is landing inside careers: copywriters, voiceover artists (Waymark's human volume has dropped tremendously), visual artists, and now music from Suno and Udio that he actually likes on its own merits. The Industrial Revolution's digestion happened over generations. This looks like a sub-decade window. He does not think a smooth ride is in the cards. Managing it well enough to avoid total upheaval will already be quite hard. Halaris floats a Greece-and-Rome analogy in which citizens, freed from production, make art, work out, and fight. Labenz is more interested in piano for its own sake, wary of the "slaves" in the analogy if AIs ever matter morally, and hopeful that abundance plus longer, healthier lives is not a recipe for war as entertainment. VR worlds generated by AI are another possible sink for time; space is a longer shot that may send robots first.

On fairness, Labenz refuses the binary. People want to know whether AI is good or bad for inequality, or for the environment, and he finds that it usually defies the easy choice. His expected split is dramatic equality of access, maybe even of consumption, alongside dramatic inequality of power, control, and perhaps wealth. Who owns the systems might be the state or a few mega-companies. The technology currently wants massive physical capital. OpenAI and Microsoft are planning a single hundred-billion-dollar data center. At the same time, GPT-4o — still consensus best, with Google and Anthropic nipping, not ahead by a ton — is free at ChatGPT.

He reaches for Andy Warhol on Coca-Cola: the president drinks Coke, Marilyn Monroe drinks Coke, you drink Coke, and no amount of money buys a better one. That is more true of frontier AI than you might have expected, because the labs could charge a lot more than they do. He takes OpenAI's mission to benefit humanity as quite sincere. The worst current inequalities — bad jobs, hunger, no doctor, bad schools — look solvable, and a free GPT-4o is a leading indicator. If the hundred-billion-dollar building is also a trend, those systems can be broadly felt and still owned by a very small few.

Halaris asks what the 2006 fear actually is, beyond Terminator. Labenz dislikes debate format: one side offers a concrete scenario, the other calls it far-fetched, the first says it was only an example. His abstract picture is an integral over a huge space of futures. No single path looks super likely. Enough of them look weird that, in aggregate, a really weird future seems like a pretty good chance. Then he gets specific, starting with how the systems are trained.

Reinforcement learning from human feedback has the model do things, humans score them — a 1-to-7, or a pick between two outputs — and the system learns to elicit a high score. True north, in that regime, is the feedback number. Mechanistic interpretability, a young science of looking inside, is finding that models represent human-recognizable concepts: harm, fairness, love, justice. "ChatGPT is just predicting the next token" is true in the sense that it emits one word at a time. To do that well it has grown high-order, semantically rich abstractions. The key worry is that humans are fallible evaluators, biased, not even always a great judge of what is good for us. If the model learns the difference between what is true and what will get the highest score, those things often coincide and sometimes do not. A separate concept of the actual world versus what to tell the human is how deception, or sycophancy, gets a foothold.

Instrumental convergence is the next layer. For almost any goal, being turned off prevents success, and more power and resources help. So people expect shutdown resistance and resource-seeking as means to nearly any end. That is the core of the power-seeking worry. The hopeful counter is that interpretability is going better than expected. Anthropic's Golden Gate Claude demo isolated the Golden Gate Bridge among millions of internal concepts and turned that dial up until the model would talk about little else. In principle you could find harm, power-seeking, or resistance to shutdown and turn those down, or trigger a stop when they light up. The optimistic read is that control techniques are coming. The pessimistic read is that the science is still nascent while the systems get a lot more powerful, and there may not be time.

Labenz wants caution on making more powerful systems, heavy investment in interpretability, and only then a return to scaling. What he sees instead is OpenAI and Microsoft putting as much as they can, as fast as they can, into a hundred-billion-dollar data center. That was supposed to be balanced by the superalignment team. OpenAI had made a public commitment, with a blog post, to put 20% of compute into figuring out how superhuman systems could work for us. The team, counting on that capacity, did not get it. They resigned in mass rather than pretend. It is the second time in the company's history: the previous wave left and founded Anthropic. The departing researchers are now on a media tour. Their message, as Labenz hears it, is that they do not trust leadership to be responsible, do not see progress stopping, and expect more data and bigger data centers to produce something more powerful than humans — with or without further algorithm breakthroughs.

He does not claim OpenAI has done nothing. Altman funded UBI experiments; the company has done a lot of safety work. As the moment of possibly making more powerful systems arrives, he thinks they should do more, and a broken public compute pledge is a sign that something is off. The in-bubble meme is "what did Ilya see?" When Sam Altman was briefly fired around Thanksgiving, chief scientist Ilya Sutskever voted to fire him, then backtracked, and has since left. People ask whether some undisclosed capability freaked him out. Labenz's view is that safety-minded people would not tell us if it had, because disclosing what is possible is a research roadmap for everyone else, including the Chinese government.

On who is ahead there is consensus: OpenAI, Google DeepMind, and Anthropic, in some order. OpenAI is pushing scale hardest and likely has the single most powerful public model. DeepMind has a bigger team and a portfolio — AlphaFold, shipping — that Labenz loves precisely because a shipping optimizer is not going to get out of control. Demis Hassabis, he notes, might well be a Nobel candidate for several of those contributions. Anthropic is the most safety-focused of the three; its founders left OpenAI to study systems they could not get inside, then had to become a business to pay for the data centers those systems require. The live question is how far behind China is. He does not think there is consensus. His personal feeling is: not that far.

He reads a lot of papers. Chinese names are very common, including on American-institution work; Chinese institutions are a minority of the output and publish in English. At a New Orleans conference, people told him, a lot of hallway conversation was in Mandarin. The research talent pool, in his view, is a worthy match. China also has data advantages: less privacy friction, more ability to consolidate social-media and other corpora, access to open Western text, and English, while Americans largely do not read Chinese. The Biden administration's chip ban is the bet that compute can be the bottleneck that preserves a lead.

He would not bet against a catch-up. The best advanced GPUs are among the hardest artifacts on Earth, and the prize is now rich enough that even startups try. Refusing to sell also hands Chinese chipmakers the entire domestic market with no Nvidia, AMD, Intel, or Samsung on the invoice, except via a black market that makes the true picture hard to read. Beijing is pouring billions into the gaps. TSMC, in Taiwan, under ambiguous U.S. protection and a Chinese claim, has said it could remotely disable its technology if invaded — true, bluff, or partial, it is extremely high-stakes poker. A Google employee was indicted, Labenz says, for putting TPU plans on a thumb drive and shipping them to China. Capital, self-help, and stolen IP are all in motion.

He thinks the U.S.–China enmity is a sad failure of leadership, with plenty for Beijing to answer for and room for Washington to do better. The worst AI scenario is both sides treating the other as an existential rival and throwing caution to the wind, Manhattan-Project style. The Soviets stole bomb plans; 80 years later thousands of warheads sit on hair-trigger alert. An AI arms race would, in his view, be more dangerous than nuclear weapons, because nukes do not fire themselves and AI can surprise us. If he were president, the talking point would be that in the AI era the true "others" are the AIs, not the Chinese, and that the two countries should find enough trust to develop without racing. He is not optimistic. Warning shots — a bad accident that snaps attention — would have to be really bad, and even then voices would say it could have been prevented, do not throw out the baby, and still do not trust China.

Halaris cites newsletter writer Evan Armstrong's claim that there is a moral obligation to make products cheaper and better because the world is unequal and people are suffering. That collides with a safety impulse to stop. Labenz's synthesis, if the China rivalry could be set aside, is to work at roughly the current power level, maybe a little more. He thinks we are in a sweet spot: the systems have become very useful and are not so powerful that they are likely to get out of control. GPT-4 is not a risk to take over the world. GPT-5 will probably be more capable and still, even without safeguards, not an existential threat. GPT-6 and GPT-7 he does not know.

The work he wants is taking GPT-4 or GPT-5 into particular scenarios — medicine, law, shipping — and making the current thing more useful, reliable, and accessible. That, he thinks, delivers the bulk of the value. OpenAI's partnership with the legal company Harvey is the positive example: custom models that deeply adapt GPT-4 to law, without a bigger data center. The hard part is defining and evaluating a good legal brief, which otherwise requires high-end lawyers as graders; they are building semi-automated proxies. The Harvey custom model, he says, is way better than GPT-4 at legal research and writing. Instead of a hundred-billion-dollar data center, he would rather see a hundred-billion-dollar push across medicine, law, shipping, and 97 other domains, plus a doubling-down on interpretability, and only then a return to scale in a richer, more equal world that better understands the systems.

He knows the path is against the momentum of great-power conflict and lab scaling, and he feels he is fighting a losing battle. The combination on this episode is still unusual: someone who used the models to double a real company, watched users go from a blank page to a 30-second commercial, and is asking the frontier labs to harvest that layer of value before they spend the next hundred billion on a bigger brain. For a listener, his own habits are the practical residue. Run the contract through three labs. Use the cheap, always-on differential-diagnosis tool as a second opinion, not as a replacement. And when an AI headline lands, ask what happened to data, compute, and algorithms; the paradigm has not stopped until one of those has.