The 900: Why Covenant72B Will Soon Be Ordinary (And Why That's the Point)

On June 27, 1999, at the X Games in San Francisco, Tony Hawk dropped into a vert ramp and attempted something that professional skateboarders had spent years agreeing was physically impossible: a 900. Two and a half full rotations in the air. 1,260 degrees of spin before landing on a piece of wood.

He failed his first attempt. And his second. And his third through tenth. The competition clock expired. The judges let him keep going anyway. The crowd, the other skaters, everyone in the arena understood they were watching someone fight a wall that didn't want to move.

On his twelfth attempt, Hawk landed the 900. He wobbled. He grabbed the nose of his board. By any technical standard, it was rough. But he rode away. The crowd erupted. Fellow skaters rushed the ramp. Bob Burnquist was crying.

That was twenty-seven years ago. Today, twelve-year-olds land 900s at local skateparks. Tom Schaar landed a 1080 when he was twelve. Gui Khury landed a 1080 on a vert ramp at eleven. Mitchie Brusco has thrown 1260s in competition. The trick that consumed over a decade of Hawk's career, the trick that defined the outer boundary of what a human body could do on a skateboard, became a stepping stone. A prerequisite. Something kids learn on their way to the things that actually impress their friends.

The skating matters less than what came after.

The Seeing-It-Done Effect

Steven Kotler spent years studying what separates people who accomplish the impossible from everyone else. His book The Art of Impossible is the result, built on neuroscience, psychology, decades of reporting on extreme performers. The book is full of frameworks and biology, flow states and intrinsic motivation stacks. But the finding that stayed with me is the simplest one.

Once someone does the thing everyone agreed couldn't be done, everyone else does it too. Fast.

Kotler calls it the "seeing it done" effect. It operates below conscious thought. When you witness someone accomplish what you believed was impossible, your brain doesn't just update a spreadsheet of facts. It undergoes a neurochemical event. Dopamine floods the prefrontal cortex. The brain's predictive model, the internal map that tells you what is and isn't achievable, gets rewritten at the hardware level. You don't decide to believe it's possible. Your biology decides for you.

This is why the effect is so much more powerful than mere inspiration. Inspirational stories make you feel good for an afternoon. The seeing-it-done effect restructures your neurology. It's the difference between reading about swimming and getting thrown into the ocean. One gives you information. The other gives you a new relationship with water.

Roger Bannister broke the four-minute mile in 1954 after decades of expert opinion holding that the human body simply couldn't run that fast. The physiological arguments were elaborate. The consensus was firm. Then Bannister ran 3:59.4, and forty-six days later, John Landy ran 3:57.9. Within three years, sixteen runners had broken the barrier. Today, the four-minute mile is the entry fee for competitive middle-distance running. High school students hit it.

The pattern repeats across every domain where records are kept. But the interesting thing is what doesn't change between the last person to fail and the first person to succeed. The runners who broke four minutes after Bannister didn't have better shoes or better training programs. They had the same tracks, the same distances, the same human legs. What changed was inside their skulls. Bannister's run deleted a constraint that had been treated as physics but was actually psychology.

Hawk himself understood this. He said the 900 wasn't really about him. It was about what came after. Every kid who watched that grainy ESPN footage internalized a new reality: this is something a human being can do. They didn't need to overcome thirteen years of failure. They just needed to have seen it land once.

The Construction of Impossibility

This raises a question that Kotler circles but never quite lands on: if the barrier was psychological all along, who built it? And why?

The four-minute mile is instructive. The physicians who declared it impossible weren't fringe figures. They were establishment authorities: exercise physiologists, sports medicine experts, coaches of elite runners. Their authority rested on their expertise, and their expertise rested on the existing boundaries of the sport. If four minutes remains unbreakable, then those experts understand the full territory. If it falls, they were wrong about the map. Their authority evaporates with the barrier.

Thomas Kuhn wrote about this in The Structure of Scientific Revolutions. Dominant frameworks don't fail gracefully. The people who built their careers inside the old model don't hear the evidence against it and update their views like reasonable Bayesians. They fight. They dismiss. They explain away. The framework only dies when the next generation of practitioners grows up having never internalized the old constraints. As Kuhn put it: "a new scientific truth does not triumph by convincing its opponents, but rather because its opponents eventually die."

That sounds harsh. It's also observably true in almost every field.

The construction of impossibility serves the people embedded in the current consensus. This isn't always cynical. Often they genuinely believe the constraints are real. The sports physiologists truly thought human lungs couldn't sustain a sub-four-minute pace. But their belief was shaped by their position. They'd built careers on a model of human performance that had the four-minute mile as a ceiling. The model was coherent. It explained the evidence. It just happened to be wrong.

Now apply this to AI.

The Wall in Front of AI

The consensus in AI training has been this: you cannot train a serious model without a datacenter. Period.

The reasoning came from engineering constraints. Training a large language model requires thousands of GPUs exchanging hundreds of gigabytes of data every few minutes. That volume of communication demands co-located hardware connected by InfiniBand cables measured in meters. Internet bandwidth is too slow. Latency is too variable. The synchronization math doesn't work.

These are real constraints. But notice who benefits from treating them as permanent. Google, Microsoft, Anthropic, and Meta own the datacenters. They employ the researchers. They publish the papers. Their entire competitive position rests on the assumption that intelligence requires infrastructure that only they can afford. When a Google researcher says decentralized training can't work, they're making a technical claim. They're also making a claim that happens to protect their employer's monopoly on the most valuable technology in the world. Both things can be true at once, and that's exactly what makes the belief so durable. The belief is genuine. The people holding it see confirmation everywhere they look. That's what makes it so hard to dislodge.

So whoever controls the datacenter controls the intelligence. Three or four companies own the facilities. They decide what gets built, who benefits, and what values get baked into the systems that increasingly mediate economic and social life. And the impossibility of alternatives isn't just a technical assessment. It's load-bearing. It holds the entire power structure in place.

Then Covenant trained a 72-billion-parameter model across the public internet.

What Covenant72B Actually Did

Covenant72B won't top any leaderboard. Bigger models exist. Better benchmark scores exist. If you judge it by those metrics, it looks unremarkable.

That framing misses everything important.

Covenant72B is the largest model ever trained in a fully permissionless, decentralized manner. No datacenter. No central coordinator. No single entity controlling who could contribute compute, what data was used, or what the model learned. Participants ran training on their own hardware, scattered across the globe, connected only by regular internet connections. The model was trained on Bittensor's Templar subnet (SN3), where economic incentives rather than employment contracts coordinated the work.

The technical achievement that made this possible is SparseLoCo, developed by the Templar research team led by Sam Dare. SparseLoCo compresses gradient synchronization to 0.78% of its original size while allowing nodes to take hundreds of local training steps before communicating. That 280GB-per-sync problem? SparseLoCo turned it into 2.2GB, synchronized 500 times less frequently. The communication overhead that made decentralized training "impossible" dropped from dominating the training run to barely registering.

As I wrote when the heterogeneous training paper dropped, the fundamental question shifted from "Can this work at all?" to "How quickly can this scale?"

That shift is Bannister's 3:59.4. That shift is Hawk's twelfth attempt.

Why Ordinary Is the Goal

There's a temptation to treat Covenant72B as an endpoint. The culmination of a long technical fight. Something to celebrate and then move on from.

That would be the wrong lesson from every story about broken barriers.

The point of the four-minute mile was not Bannister's personal glory. The point was the sixteen runners who followed within three years. The point was every high school runner who has since crossed that line without it feeling historic. The four-minute mile mattered because it became ordinary.

The point of the 900 was not Hawk's twelve attempts on that June evening in San Francisco. The point was Gui Khury, eleven years old, throwing 1080s in his backyard. The 900 mattered because it became boring.

Covenant72B needs to become boring.

It needs to become the thing that a research group at a university in Nairobi does as a warm-up exercise. The thing a collective of independent GPU owners runs on weekends for fun. The thing that's so routine, so unremarkable, that nobody writes blog posts about it anymore. It needs to become the floor.

If that sounds like I'm underselling it, I'm not. I'm applying the only standard that matters. The measure of a breakthrough is not the breakthrough itself. It's the cascade it triggers.

Kuhn saw this clearly. The real revolution doesn't happen when the new idea is proposed. It happens when the next generation of practitioners grows up inside the new framework, unable to understand why anyone ever thought differently. The revolution is complete when the old impossibility becomes invisible. When a twenty-year-old ML researcher in 2030 hears that people once believed you needed a datacenter to train a large model, and she finds it as quaint as the idea that humans can't run a mile in under four minutes.

The Cascade Is Already Starting

Covenant didn't stop at 72B. The same team, driven by the same refusal to accept that intelligence must be centralized, has been dismantling barriers in sequence.

Grail, Covenant's post-training subnet, published PULSE: a technique that achieves 100x compression in weight synchronization for reinforcement learning. This matters because post-training is where AI learns judgment, where values and behaviors are shaped. PULSE made decentralized post-training practical at centralized speeds. The phase of AI development where alignment happens, the phase that determines whether a model is helpful or harmful, is no longer locked behind corporate doors.

Heterogeneous SparseLoCo extended Templar's work to let consumer GPUs join frontier training runs alongside datacenter hardware. Your RTX 4090 can now contribute to training runs that previously required million-dollar clusters. The barrier to participation dropped from "do you have a datacenter?" to "do you have a GPU?"

Basilica, the compute infrastructure layer, is building the plumbing that connects all of this into a unified system. Agent-native runtime. Decentralized job scheduling. The boring infrastructure work that makes the exciting stuff reliable.

Each breakthrough lowers the floor. Each one makes the next easier. Each one expands who can participate. This is the cascade. This is the sixteen runners breaking four minutes within three years.

The Attention Problem

Kotler writes about something he calls the "motivation stack." At the base of the stack is curiosity. Then passion. Then purpose. Then autonomy and mastery. The stack works by layering: each level provides fuel for the next. But the whole structure depends on an initial spark, something that grabs attention hard enough to make curiosity ignite.

Decentralized AI has an attention problem. It doesn't have the marketing budgets of OpenAI or the media magnetism of Google. Sam said it plainly: "Bittensor suffers a problem of attention. Because we reject capital that will corrupt us, they discount all our achievements."

This is the structural tension at the heart of principled infrastructure work. Projects that compromise their mission get funding, media coverage, and institutional prestige. Projects that refuse to compromise get ignored until they become undeniable.

But this is also why the barrier-breaking moment matters so much. Bannister didn't need a marketing budget. Hawk didn't need institutional validation. The 900 spoke for itself. It created its own attention because it was real, visceral, undeniable.

Covenant72B is that moment for decentralized AI. A 72-billion-parameter model, trained permissionlessly, over the open internet, with performance approaching centralized baselines. Not a whitepaper. Not a roadmap. Not a promise. A model you can download and run right now.

Sam's philosophy captures this perfectly: "Knock it out of the park. Be so good you become undeniable. The only merit that matters is your accomplishment."

What This Means for Everyone Else

Kotler's research shows that impossible achievements don't just lower barriers for professionals. They change what people believe is worth attempting.

Before Bannister, nobody trained for a four-minute mile because what was the point. After Bannister, thousands did. The breakthrough didn't just prove a feat was possible. It made attempting it rational. It moved the goal from the category of "delusion" to the category of "ambition."

This distinction matters more than it seems. We tend to think of ambition as a personality trait, something you either have or don't. Kotler argues it's closer to a biological response to environmental conditions. Your brain allocates motivational resources based on its model of what's achievable. When the model says a goal is impossible, the brain withholds dopamine, norepinephrine, and the other neurochemicals that sustain effort over time. You don't lack willpower. Your biology is rationing fuel because the destination doesn't appear on its map.

When someone proves the destination exists, the map updates. The fuel flows. What looked like a character flaw was actually a calibration error.

Covenant72B does the same thing for decentralized AI. Before Covenant72B, starting a decentralized training run felt quixotic. The technical consensus said it couldn't work at scale. Why burn your GPUs on something the experts agreed was impractical? Your brain, quite reasonably, steered you toward things with better expected returns.

After Covenant72B, the calculus changes. It works. Someone did it. The question has moved from whether to how fast. And every research group, every university lab, every GPU collective that was waiting for permission to take decentralized training seriously has that permission. Not permission from an institution, but the more powerful kind: permission from reality. The kind your neurology can't argue with.

This matters because the alternative is grim. As I've written about the AI war machine and the enclosure of the AI commons, centralized AI development has produced exactly what concentrated power always produces. Every major lab now contracts with defense agencies. Every one frames AI as nationalist competition. The companies that pirated the open internet to build their models now demand intellectual property protections for the output.

If intelligence remains something that only three companies can produce, intelligence will serve those three companies and the governments they align with. If intelligence can be produced by anyone with compute to contribute, the structural incentives change. The change won't be perfect or magical, but the direction reverses.

And this is the part that Kotler's framework illuminates most sharply. The seeing-it-done effect doesn't just give individuals permission to try. It rewires entire communities. When Bannister broke four minutes, it didn't just inspire sixteen runners. It restructured how coaches designed training programs, how sports physiologists modeled performance, how young athletes picked their events. The broken barrier propagated through the whole system, changing assumptions at every level.

Covenant72B has the same potential to propagate. If it does, the change won't look like a single team scaling to 200B, then 500B. It will look like hundreds of teams, in dozens of countries, training models for purposes that Covenant never imagined. Some will be better. Many will be worse. All of them will be possible because the map changed.

The Undeniable

Tony Hawk spent thirteen years chasing a trick that experts said the human body couldn't perform. Roger Bannister ran against a wall that physicians said human lungs couldn't breach. In both cases, the barrier was real right up until it wasn't. Then it was never real again.

Covenant72B trained a frontier-scale language model on infrastructure that the AI establishment said couldn't support it. Across commodity internet connections. Without a datacenter. Without a central coordinator. Without permission from anyone.

It worked.

And now, like every barrier that has ever fallen, it will start to seem obvious. Inevitable, even. People will forget it was ever considered impossible. Kids at skateparks don't know the 900 was once the edge of human capability. They just know it's one of the tricks you learn on your way to something harder.

That's the goal. Not permanent amazement at what Covenant achieved. But a world where what Covenant achieved is so ordinary that nobody remembers why it was hard. Where decentralized training is just training. Where intelligence produced as public infrastructure is just intelligence.

Sam ended a recent community call with something that stuck: "We will keep delivering. Thank you for your faithfulness."

The faithfulness he's asking for isn't blind. It's the specific faith of showing up when the consensus says you're wasting your time. The faith of Hawk dropping in for a twelfth attempt after eleven failures. The faith of Bannister lacing up his shoes against the four-minute wall.

That faith looks foolish right up until the moment it becomes foresight.

Covenant72B landed. The 900 is done. Now comes the cascade.

---

Related Reading:

• The Internet is the Datacenter: How Templar's heterogeneous training enables consumer GPUs in frontier model training
• Who Teaches the Machine: How Grail's PULSE research is decentralizing AI post-training
• The Enclosure: How AI labs that pirated the commons now demand you respect their fences

Disclosure: I work with Covenant AI and am directly involved in the organization's communications. For full transparency about my involvement and investments, see my projects page. All opinions expressed are entirely my own.

For more on Covenant AI's ecosystem, visit covenant.ai.