A letter from zeb

You are working with zeb because we have built this company around AI as an environment, and that structure produces the highest quality of work, in the fastest time, at the lowest cost.

The old way to sell expensive work was to claim privileged access to expertise the client did not have. We do not make that claim. The thing we have that the rest of the industry does not is structural: every part of how zeb intakes, executes, and ships work has been rebuilt around AI as the environment we operate inside, rather than a tool we operate on top of. That structure compresses time, cost, and quality in a way wrapper firms cannot match without rebuilding themselves. We will not pretend our advantage is anything other than what it is.

Before we describe what we built, we want to describe what we are, because the structure did not come from a slide deck or a strategy offsite. It came from a prior conviction about how a company like this has to work at the level of its biology.

That is not a metaphor. It is the design.

Consider the grasshopper. A solitary grasshopper is a cautious, independent creature: distinct coloring, distinct behavior, distinct neural wiring. Crowd enough of them into proximity and something measurable changes: serotonin spikes, the nervous system reorganizes, and the individual becomes part of a swarm capable of covering continents. The individuals did not disappear. The swarm is made entirely of them. But the swarm does things no individual chose to do and no individual could do alone. The emergent behavior is real, not rhetorical: it is a phase change with a measurable biological trigger.

Siddhartha Mukherjee, writing about cancer in The Emperor of All Maladies, describes something structurally identical in malignant cells. The properties that make cancer so difficult to kill (adaptability, resource acquisition, resistance to signals that tell normal cells to stop) are not bugs in the biological system. They are the biological system operating without a governor. The resilience is native. The evolutionary pace is real. We are not interested in the pathology. We are interested in what the pathology reveals about the substrate it runs on; that a system built from first principles, without vestigial structure, without layers that do not produce, is extraordinarily hard to stop.

Every person at zeb is genuinely solitary: our own motivations, curiosities, needs. We are not held together by hierarchy or mandate. We are held together by a shared trait: the love of the problem and the drive to produce an outcome. When that density reaches threshold, the phase change happens. We become something the client cannot build internally and the GSI cannot replicate at scale, not because we are smarter, but because the organization itself behaves like a living system. Signaling pathways replace reporting lines. Reward signals replace performance reviews. Vestigial functions are removed the way evolution removes them, not by committee, but because they stop producing. What remains is an organization where everything has a purpose because nothing without a purpose survives.

Before, we had practices. Practice heads ran them. Each engineer was specialized, because specialization was what made an engineer a subject matter expert. Research lived in its own arm of the company: innovation existed, but it was somebody else's responsibility. When a project came in, staffing was top-down: which specializations does this need, who is available, assemble the team. Sales sat with sales. HR sat with HR. PMO sat with PMO. The structure looked like every other firm in our industry, because it was modeled on every other firm in our industry.

After, we have two teams in engineering: research and applied research. No specializations. People join the projects they want to join. We took the idea from Valve, that the best work comes from people who chose the work, and we took it seriously. We track engineers across competencies, breadth and depth, and we incentivize them to take projects that make them uncomfortable. Curiosity is the input we measure for. Discomfort is the medium it travels through.

We removed middle management. The flat structure is not a perk. It is a load-bearing decision. In a company where creativity is the unit of production, layers of people whose job is to summarize the work for other people are friction we cannot afford. Extreme ownership is the default, not because we tell people to own things, but because there is no one else to own them. Sales, ops, PMO: these functions still exist, but the people in them lead with technical fluency. The base prerequisite is that you can create the work, not that you can talk about it.

We mean this literally. A client hires a firm to produce an artifact. The firm puts the request to a model. The model produces the artifact. The firm packages the artifact and delivers it. The client pays the firm. The intermediation layer added almost nothing the client could not have done. The firm's value is friction: the friction of finding the right model, writing the right prompt, formatting the right output. AI has made that friction trivially small.

We don't say this to be cruel. We say it because if you do not see it clearly, nothing else we built makes sense.

The same diagnosis applies upstream of services. Specialization stopped being scarce the moment the generalist became a genius. The companies whose moat was we have specialists who write the specialized code never had a moat: they had a delay. The GSIs, the long tail of vertical SaaS: most of them are running on infrastructure built for a world where lack of access was the moat. That friction is gone, and the moat went with it. What looks like industry health right now is the lag between the moat collapsing and the customers noticing.

Fine-tuning was supposed to be an extension. It is not. The generalist models are getting good enough at nuance that the case for specialist models is thinning every quarter. The companies betting on specialization as their structural answer are betting on a curve that is bending against them.

The arc of software has always pointed in one direction. Easier execution. Higher abstractions. More leverage per engineer. Forty years of accelerators, libraries, frameworks, compilers, package managers, IDEs: all of it pointed at the same horizon. Every step in that direction made the previous moat smaller. We have been walking toward this for a long time. AI is the largest step yet, but it is a step in a direction the industry was already walking.

The firms treating AI as a tool inside their existing structure are treating an over forty-year deadline as a feature request. We will adopt this where it helps [us]. That posture made sense for compilers. It made sense for cloud. It does not make sense for the layer of abstraction we are now operating at, because at this layer the tool is the environment. You do not adopt an environment. You move into it.

We borrowed an image from shipping. Malcolm McLean did not invent the shipping container. The container existed in various forms before he got to it. What McLean invented was the ecosystem: the ships designed for containers, the cranes that handled them, the trucks that received them, the ports that processed them. Pre-container, port specialization was the moat. Each port had its own dockworkers, its own loading techniques, its own way of doing things. Post-container, that whole structure became dead weight overnight. The ports that restructured around the container survived. The ones that did not, didn't.

The container was not the cause of the change. It was the deadline on a change that the global shipping industry had been delaying for decades.

Picture a steel mill. Raw context goes in one end. Finished outcome comes out the other, near-instantaneously. That is the shape of how zeb runs.

The mill has two halves. The first is the Foundry: execution. Substrate is the system of record at the engineering layer. Variants run inside it. Solution definitions resolve into shipped artifacts. The Foundry is what makes us cheaper and faster than anyone doing the same work the old way. It is the part of the company that the client interacts with.

The second is the Forge: research. The engineers who, in our old structure, would have been writing specialized code now operate at the layer of raw primitives. They extend what Substrate is capable of, ahead of any business case demanding it. They do not have specializations. They have curiosity. The Forge is what makes the next year's Foundry better than this year's. It is the part of the company that the client never sees, and never needs to.

Together they are the moat. Separate, neither would be.

The principle running through both halves is one we keep coming back to: you can outsource thinking, but you cannot outsource understanding. Every artifact we ship is grounded in a human who understood what the outcome had to be. Substrate does the production. The human carries the understanding. When reality changes, and reality always changes, the human carries the work forward, and Substrate learns through the harnessing engine we built underneath. The model does not get smarter on its own inside our walls. It gets smarter because we built the structure that lets it learn from understanding it could not have generated for itself.

Here is what we mean. The wrapper firms operate as vendors: distinct from the client, billing for distance, performing the role of the outside expert who has been brought in. There is a whole vocabulary built around this posture. Engagement. Stakeholder. Deliverable. Stand-up. The vocabulary exists because vendors have to keep reminding everyone they are vendors, otherwise the artificial separation collapses.

We do not want the artificial separation. We want the work to look like it came from inside your company, because in every way that matters, it did. We learn the cultural fluency, the language, the operating cadence, the way decisions get made, and then we work inside it. Not on top of it. Not adjacent to it. Inside.

The metaphor that works for us here is the mirror. Not in the cosmetic sense but in the structural sense. A mirror does not add anything to the room. It does not change the room. It returns the room to itself with enough fidelity that, briefly, you cannot tell which side is which. That is the posture. The understanding we ground the work in is not generic understanding of the problem space. It is specific understanding of the company we are working with. You can outsource thinking, but you cannot outsource understanding. We said it earlier about Substrate and the Forge, and we mean it the same way about the client.

Models innovate within the bounds of their training. They produce derivatives of induced experience: next-best-token across what they have already seen. This is not a complaint. It is a description. A new primitive (a genuinely new idea, a previously uninstantiated concept, an unprecedented framing) is something the model could not have generated until a human created it. After the human creates it, the model can learn from it forever. But the primitive is upstream of every derivative the model will ever make.

A human can be transformed by a single experience. A model cannot. That asymmetry is the human moat in any business operating in an AI environment, and it is the only moat we know of that compounds rather than depreciates.

The figures who shaped our thinking here did not come from our industry. They came from mathematics, physics, biology, computer science, and what they shared was not genius in the conventional sense but a specific kind of intellectual posture: the refusal to optimize against a known objective to the exclusion of the unknown ones surrounding it.

Richard Feynman, asked to explain why magnets repel, refused to give a satisfying answer, not because he didn't understand magnets but because he understood that a satisfying answer at the wrong level of abstraction is a finished answer, and finished answers stop being useful the moment the problem changes shape. His diagrams were not a finished theory. They were a notation that let physicists think about problems Feynman himself never tackled. "I would rather have questions that can't be answered than answers that can't be questioned." He treated his own ignorance as data.

Margaret Dayhoff, in the early 1960s, built the first biological database when bioinformatics wasn't a word and biologists viewed computers with suspicion. No single researcher could hold all the relevant comparisons in their head, so she built the infrastructure that let the field hold them collectively. Her PAM substitution matrices improved as more data flowed in. She didn't try to solve protein evolution. She built something that let the field solve it incrementally over decades. Her tools got smarter as the world got more data.

Fei-Fei Li saw what the computer vision field had stopped seeing: that the bottleneck wasn't the algorithms, it was the data. This was not an obvious read at the time. Through the 2000s, the dominant conviction inside computer vision was that the path forward was cleverer models: better feature engineering, more sophisticated architectures, tighter mathematical formulations. The field was optimizing hard against a known objective and not asking whether the objective was correctly defined. Li, coming from cognitive science rather than from inside the vision field, was asking a different question: how do children actually develop visual understanding? The answer pointed at volume and variety of labeled experience, not at the cleverness of the inference engine running on top of it. The bottleneck was not the algorithm. It was the substrate the algorithm was learning from. So she built ImageNet (three years of work, more than fourteen million labeled images, a dataset so large that the vision field initially questioned whether it was worth building) and released it. The field's existing algorithms, run against that substrate, suddenly started working. Geoffrey Hinton's team won the 2012 ImageNet competition with a convolutional neural network that cut the error rate nearly in half. Deep learning's takeoff happened on her runway. She did not invent the algorithm that won. She built the ground it ran on.

Srinivasa Ramanujan had almost no formal training when he began mailing results to G. H. Hardy from Madras (the city where zeb was born) in 1913. The results were correct: theorems of extraordinary depth that Hardy and his colleagues could verify but often could not initially derive. Ramanujan had not worked backward from axioms. He had worked forward from intuition so compressed it preceded formalization entirely. "An equation for me has no meaning unless it expresses a thought of God," which is one way of describing a relationship with mathematical truth so direct that the formal apparatus for understanding it had to be built afterward, by others, over decades. He produced primitives. The framework for explaining why they were true came after.

What all four of them share: naïveté about the destination field was not their handicap. It was their methodological advantage. Dayhoff applied computing to biology before biologists would. Li approached vision through cognitive science. Feynman barged into Caltech biology labs because he found it interesting. Ramanujan had no established school of thought to constrain what he was allowed to find. In each case, the lack of formal grounding in the destination field was the reason they could see the obvious thing the natives had stopped seeing.

The Forge exists for this reason. We do not put engineers in the Forge to do work that Substrate could do for itself, slower. We put them there to do work no model can yet do at all: to operate at the layer where new primitives come from, so that the next iteration of Substrate has primitives to derive from that did not exist before we put them there.

You probably know the scene we are about to describe. The Hot Gates. Leonidas has brought three hundred Spartans. The Arcadians arrive with a thousand men, and their commander is outraged that Leonidas brought so few. Leonidas walks the Arcadian line and asks each man his profession. Potter. Sculptor. Blacksmith. Then he turns to his three hundred and asks the same question. They roar back: Aroo.

Leonidas turns to the Arcadian and says: I have brought more soldiers than you.

We come back to that scene a lot.

The argument it makes is the argument we are trying to make about zeb and the GSIs. A 1,500-person zeb structured correctly outcompetes a 700,000-person GSI with the same model access. Not because we are smarter than the people at the GSI, some of them are extraordinary. The argument is structural, not personal. Every person at zeb is professionally the work. Most people at the GSI are professionally something adjacent to the work: managers, account leads, partners, ops staff, salespeople who do not build, analysts who do not ship. The thousand Arcadians are professionally something else, and they fight when called. The three hundred Spartans are professionally Spartans, and they are in the fight even when there is no fight.

The GSIs are not failing because they lack the tools, reader. They have the same tools. They are failing because they are too large, too profitable, and too complacent to restructure around the tools. The music is still playing for them. It has been playing for so long that they have forgotten the music can stop. It will stop without warning, and they will not know until it has.

In 218 BCE, Hannibal Barca took 50,000 men, 9,000 cavalry, and 37 war elephants over the Alps in winter. He was attacking Rome from the north, the direction Rome was not defended from, because the Alps were thought to be uncrossable in winter and so Rome had not bothered defending against an attack from above them. He lost nearly half his force in the crossing. He arrived in Italy with an army that could not resupply, could not retreat, and could not be reinforced. Then he won three of the largest battlefield victories in ancient history before Rome figured out how to fight him.

And once they weren't, Rome's northern defenses were facing the wrong direction permanently.

We have been thinking about Hannibal a lot.

The industry's defenses are pointed at the directions it expected to be attacked from: quality, expertise, headcount, brand, billing rate. These are the moats it has been reinforcing for decades. None of them are the route we took. Substrate, the Foundry, the Forge, our embodiment posture, the 100% outcome guarantee, Aroo: these are not better versions of what the industry has. They are an attack from a direction the industry does not defend.

By the time the wrapper firms understand the structural argument we have made in this letter, their defenses will already be facing the wrong way. Rebuilding takes longer than the deadline allows. We say this without smugness, reader, because we are aware that we could be wrong and that the cost of being wrong is real. But we are not running a parallel old-zeb in the background. The route we took is the only route we are on. It is irreversible without dismantling the company, which we will not do, which is the point.

We want to say something about why any of this matters beyond the company doing it.

Individual potential is always built on collective infrastructure. That is not a political statement. It is an observation about how capability actually works, one that holds up across 2,400 years of philosophy and across every empirical case we have looked at closely.

Plato's Allegory of the Cave ends not with the freed prisoner staying in the sunlight. It ends with his obligation to descend, to go back into the dark with what was learned in the sun. The capability is not legitimized by its existence. It is legitimized by whether it reaches the people who are still in the cave. Simone de Beauvoir sharpens this from a different direction: "To will oneself free is also to will others free." Freedom is not a private possession. A world of constrained people offers only a degraded version of freedom even to the unconstrained.

The observation holds regardless of ideology: whoever owns the infrastructure shapes the consciousness, options, and agency of everyone standing on it. We are building frontier capability at a moment when frontier capability is concentrating, narrowing toward fewer hands, fewer companies, fewer decisions made by fewer people. That is not an accident of technology. It is a choice being made continuously by the people who own the substrate.

The Nordic social democrats got the operationalization right without requiring collective ownership of the means of production. Tage Erlander's folkhemmet (the people's home) held that the legitimacy of a modern state rests on whether it functions as a home for all its members. The philosophical move was accepting markets while socializing the floor: education, healthcare, retraining, childcare. The question was never markets vs. no markets. The question was what counts as floor?

That is the question we carry into the work. If AI literacy, capable models, and the infrastructure to build with them are treated as floor (like roads, like literacy), the equalizing potential is structural. If they are treated as premium goods, the concentrating dynamic wins by default. We are not neutral on this. zeb exists at the infrastructure layer, and we are aware of what that means.

We measure success by whether the work we shipped made the client more capable of doing it themselves next time. We do not optimize engagement design for dependency. The ideal outcome is that the client cannot tell where their organization ends and zeb begins, and the next ideal outcome is that they need us less, because we built them something that keeps learning.

We took our posture from Marcus Aurelius. The Meditations are the proof that you can run a hard job at scale without performing it. The conviction of this letter is downstream of him. We declare what zeb is. The work either ships or it does not. That is the only register we are allowed to operate in, and it is the register he wrote in for himself, every morning, for twenty years. He also gave us the hive: "That which is not good for the bee-hive cannot be good for the bee." Individual flourishing is not sacrificed to the collective. It is constituted by it.

We took the long horizon from Charlie Munger and Warren Buffett. Sixty-four years of holding position, refusing to act when nothing was worth acting on, ignoring narrative. The 100% outcome guarantee is only credible from a company that can hold position without panicking. They are the proof that the holding is what compounds.

We took the small-team asymmetry from Renaissance Technologies. Jim Simons built the best-performing investment vehicle in history with a hiring posture that refused industry pedigree and refused to scale headcount. A small number of unusually capable people working with the right tools will outproduce a large number of average people working with the best tools. That ratio is the only one that matters to us.

We took the advantageous-divergence posture from Shankh Mitra and Welltower. He reframed his company as an operating company in a real estate wrapper, a category-replacement move. We are doing the same thing, in a different industry, for the same reason. The category needed replacing. Once that landed, there was no hedged version of either company running in parallel.

We took the non-negotiable from Anthropic. They have held lines that cost them short-term ground because the lines were the thing that earned them the right to exist in the first place. Our 100% outcome guarantee is the same kind of line. It is expensive. It restricts what we will sign. Anthropic is the proof that a company can hold an inconvenient line and still win.

We took the builder-only structure from Stripe, Shopify, and Ramp. Three companies that scaled without ever installing a political layer above the builders. The building of the product is what every role is in service of. There are problems to solve, and the company is organized as a flat-as-possible system for solving them. We are not them. But we are organized on the same premise.

The scientists in the Generative Unknown section are also in this gallery. Feynman for treating ignorance as data and rigor as a discipline rather than a performance. Margaret Dayhoff for building infrastructure over discoveries, for understanding that the substrate others learn from is the highest-leverage contribution. Fei-Fei Li for the naïveté that let her see what the field had stopped seeing. Ramanujan for the primitive that precedes the proof, for the reminder that output can be correct before anyone can fully explain why, and that the framework for understanding it sometimes has to be built by the people who come after.

We also took from Barbara McClintock, who spent decades observing corn genetics in patient solitude before the field caught up to what she had found. From Claude Shannon, who moved between cryptography and information theory not despite their distance but because of it, the proof that following genuine interest across boundaries is more generative than disciplined optimization within one. From Lynn Margulis, who was rejected for years on endosymbiosis and was correct, and who held provisional knowledge long enough to become the next generation's consensus. From Stewart Brand, who understood that the tool is the discovery, because the tool is what makes future discoveries cheap.

And from the philosophers whose argument we tried to carry into our own practice in the section above. Plato, for the obligation to descend. De Beauvoir, for the relational nature of freedom. Sartre, for the insistence that radical freedom is hollow without the material conditions that let people actually exercise it. And Tage Erlander, for the proof that the floor can be raised without dismantling the ceiling, that the question has always been what counts as floor, and that the answer is a choice.

Finally, we are grateful to Vivek and RK, the cofounders of AVASOFT, for having the courage to begin the journey that led to zeb and this moment in time.

Everything in this letter is downstream of one thing.

The structural advantage, the living system, the wrapper trap, the forcing function, the Foundry, the embodiment, the generative unknown, Aroo, the trade-offs, the route, the floor and the ceiling, all of it is an argument for why the work has to ship a certain way. None of it excuses the work not shipping. The Stoic posture we inherited from Marcus Aurelius is the floor of the company, not the ceiling: no flattery, no complaint, no consolation that anything outside the work could fix what is difficult about it.

We wrote this letter, reader, because we wanted the thinking on the record before the work moved past it. We wanted the people on the ride with us (clients, team, partners, friends) to see the shape of what we were doing while it was still possible to see it whole. The shape will get less visible from here. The work will get harder. That is fine. That is what we signed up for.