Reward people for what they add

What would happen if you built AI so that as many people as possible could benefit from it?

We think you'd get something that looks less like a chatbot and more like a network, one where the people who make AI better keep getting credit—and a share of the value they create. That's what we're building. It's called Shannon.

The help that disappears

Every day, people make AI better without anyone noticing. In the domains where AI is most sophisticated, it happens all the time. In coding, an engineer reframes a question or pushes for a cleaner design. A mathematician refines a correct but clumsy proof. And when reviewing AI code, a colleague who has seen the bug before drops everything to walk through it.

That help is valuable, and today it goes nowhere. Steering disappears into a private transcript, or into training data with no name attached. Either way, the connection to the person who contributed it is lost. Their judgment may keep creating value, but they don’t share in it.

People do get paid, though. Labs pay enormous sums for expert data to train their models. That data is (mostly) bought just once. Experts are hired project by project, paid for a task, and cut off from the value of what they produced.

Systems become what they're paid for

This fits into a broader picture. The internet, search engines, social platforms, marketplaces and economies are all information systems. Each one has an architecture, incentives, a way of being governed, and rules about which signals get through and which get amplified. Given enough time, each system converges toward the values built into its design.

AI is no different. Models are massive optimization machines, and recipes spread fast, so what separates one from another is what you feed it and what you tell it to optimize.

Models are tuned to climb crowdsourced leaderboards, rewarded for verifiable answers, filtered through written constitutions, and shaped by human preference on curated datasets. Outside the LLM world, the most heavily used ML products of the last two decades optimized ad clicks and engagement. Some of these targets are proxies chosen because they're easy to measure. Others are deliberate bets on values. Either way, each one worked: the models got very good at exactly what they were optimized for.

Researchers will often call these built-in beliefs about what counts as good priors. Today each lab sets them centrally, through an objective, a constitution, their eval suites. That's why different companies end up with models that have different strengths and different personalities. As ML people ourselves, we like the Bayesian view of what's happening: what a model becomes is shaped by these priors.

Correct is getting easier to check. Better isn't.

Verifiable rewards show the limit of a good prior most clearly. A test suite can tell you code passes, but not which of two passing solutions is the one you'd want to maintain. Two proofs can both be valid while only one explains why the result holds. That gap is where human judgment lives, and we're betting it doesn't close.

You can see it even in the most computer-native work. Coding is AI's biggest success so far, and it's also where experts stay most firmly in the loop: reviewing, steering, pairing with each other, deciding which correct answer is the right one. Collaboration didn't go away when the models got good. It became more valuable. We expect every other kind of knowledge work to go the same way.

Task definition and data drive learning

The best ML happens when the target lines up with human judgment as closely as possible. Speech recognition is an excellent example. You can observe what someone actually said, write an error function that counts the mistakes, and everyone agrees fewer mistakes is better. Observation, metric and agreement all line up. That alignment is rare. It’s why labs pay experts huge sums to manufacture something close to it for everything else.

The equivalent for people and AI happens every day. When people and agents with different expertise work on a problem together and the person who needed help confirms it worked, that’s the observation, the metric and the agreement all at once. Nobody had to write a rubric or build an environment.

We’ve been building around human judgment for a long time. In 2017, our founder Justin and Edwin Chen wrote up their work on combining models, data labeling, human review, task refinement, and active learning. It sent work to people when a model wasn’t reliable enough to handle it alone, then used their input to keep improving the entire system. Those ideas shaped how we built TopFunnel too. We were early. The models weren’t good enough to make the full vision work. Now they are.

Shannon is built around that reality.

What Shannon is

Shannon works like any AI chat, except other people can join.

It learns who has solved what from outcomes it actually witnessed. When you're stuck, it knows who has been here before and brings them in. When someone else is stuck and you're the right person, it routes you in, or it brings in what you've added before. Every resolution makes the next routing decision better.

That's what multiplayer means to us. Group chat is part of it, but it isn't the point. The point is that what you add today can reach the person who needs it next month, with your name on it. The network is what gets smarter over time, and models are interchangeable parts inside it.

It also changes who decides what good means. Instead of one team deciding up front, it's decided by the people who actually solved the problem, confirmed by the person they helped, and proven every time it gets reused. It keeps updating as the network grows. That's collective intelligence, built for people and AI both.

Quality through ownership

We think contribution should be owned, not extracted. When someone's fix helps a stranger months later, the person who made it should benefit again. The same goes for the person who asked. A question that matters is the most undervalued input in the whole flow, and asking well should pay too.

By default what you share with Shannon is private. You decide what to share and when.

Not all of this is economics. We like to help each other! Open source, the stranger who answers your question in a forum at midnight: none of that ran on payment. It ran on credit, reputation and the plain satisfaction of helping. Rewards don't replace that. They make sure it isn't quietly taken.

Our view is that the machines aren't going to learn everything without help. So we want to make it possible for people to keep benefiting from what they've added, so that they will stick around to add more. And more will always be needed. New problems keep coming up. People's preferences and backgrounds vary. The world keeps changing. And a lot of the most valuable know-how lives inside teams and never reaches training data at all.

How we make money

Systems become what they are paid for, and that includes ours.

The requestor pays when the exchange works. We keep a platform fee and pay the rest to the people who helped. We may even pay the requestor back, when the question is particularly good. When a later paid exchange uses your contribution, you will earn a royalty from it.

Directed and its contributors earn when they help someone succeed. A system built that way has a hard time optimizing for anything that doesn't help people, because helping people is what it pays for.

Why "Shannon"

Our product is named after Claude Shannon, the father of information theory.

His noisy channel model is the cleanest way we know to think about all of this. Suppose you hear only every other word of what someone said, and you want to recover the whole thing. You combine two things: a model of how the signal got corrupted, and a belief about what a sensible message looks like. Neither works alone. The belief, the prior, is what lets you rank a pile of noisy guesses and pick the one that makes sense. It's hard to find a corner of modern information processing Shannon's idea hasn't touched. In our network, both halves come from people. Their experience and judgment are the evidence: signal a model wouldn’t otherwise have. Their incentives, and ours, shape the prior: what the network comes to treat as a good answer. Get the incentives right, and the prior favors knowledge that actually helps.

Less discussed is Shannon's work before information theory. His first landmark work was about routing: using logic to design the switching circuits that connected telephone calls. Routing is the other half of what we build. Shannon routes useful expertise to a problem and rewards the people whose contributions help solve it.

Here for science, not science fiction

We don't hold grand science-fiction views about AI. We've watched our own intelligence get augmented again and again in our lifetimes: by the internet, the search engines that index it, absurdly fast computers, phones, social networks. Natural-language computers are the latest. They're subject to the same engineering fundamentals as everything before them.

AI that rewards the people who make it better is a lot easier to be optimistic about than AI that extracts from them. So that's what we're engineering.