We believe the majority of economic value created by AI will come from long-horizon agents that learn from real-world outcomes.
We believe the majority of economic value created by AI will come from long-horizon agents that learn from real-world outcomes.
The most valuable work inside companies already happens over long horizons. Developing a drug, allocating capital, underwriting a loan, or operating a supply chain all require sequences of judgement based in tacit knowledge. These processes are split across people and teams because no individual can hold all of the context, expertise, and responsibility required to own the entire outcome.
The most valuable work inside companies already happens over long horizons. Developing a drug, allocating capital, underwriting a loan, or operating a supply chain all require sequences of judgement based in tacit knowledge. These processes are split across people and teams because no individual can hold all of the context, expertise, and responsibility required to own the entire outcome.
AI has been applied within these systems one task at a time, augmenting the ways that humans already work, through copilots and assistants. These are helpful, but designing AI around human workflows is like asking for faster horses. We think the larger opportunity is to redesign work around the outcome itself without the constraints of humans. This requires moving beyond prescriptive processes and allowing models to learn the best way to achieve the objective.
AI has been applied within these systems one task at a time, augmenting the ways that humans already work, through copilots and assistants. These are helpful, but designing AI around human workflows is like asking for faster horses. We think the larger opportunity is to redesign work around the outcome itself without the constraints of humans. This requires moving beyond prescriptive processes and allowing models to learn the best way to achieve the objective.
Long horizon autonomy can’t be engineered through a harness
Long horizon autonomy can’t be engineered through a harness
Tesla’s self-driving project Autopilot began as a harness atop general models for object detection, segmentation, and physics predictions. Their approach was to wrap these models with logic to define how to drive.
Tesla’s self-driving project Autopilot began as a harness atop general models for object detection, segmentation, and physics predictions. Their approach was to wrap these models with logic to define how to drive.
After 300k lines of C++ harness engineering, their system was only capable of highway lane assist, because of the deep complexity and long tail of situations that exist in the real world. Realizing they had reached the limit of rigid logic, they replaced the harness with a single end-to-end model which learned from experience. They deployed the model into test cars, it made mistakes and then learned to fix them, and eventually it became good enough to unlock full self-driving.
After 300k lines of C++ harness engineering, their system was only capable of highway lane assist, because of the deep complexity and long tail of situations that exist in the real world. Realizing they had reached the limit of rigid logic, they replaced the harness with a single end-to-end model which learned from experience. They deployed the model into test cars, it made mistakes and then learned to fix them, and eventually it became good enough to unlock full self-driving.
Harnesses are powerful when we know the behavior we want to encode. They can give models tools, context, workflows, constraints, and known heuristics. But their capability is bounded by what can be specified in prompts and unambiguous SOPs.
Harnesses are powerful when we know the behavior we want to encode. They can give models tools, context, workflows, constraints, and known heuristics. But their capability is bounded by what can be specified in prompts and unambiguous SOPs.
The limitations of prompting are even more noticeable as tasks stretch into longer horizons. Intermediate states—document searches, test results, user feedback—change the course of the workflow and send the model down branches of the decision tree that could not have been anticipated in advance.
The limitations of prompting are even more noticeable as tasks stretch into longer horizons. Intermediate states—document searches, test results, user feedback—change the course of the workflow and send the model down branches of the decision tree that could not have been anticipated in advance.
And as we deploy AI towards more valuable objectives like scientific discovery and market research, most of the path will be unknown. A system responsible for outcomes therefore has to be capable of discovering smarter decisions through its lived experience.
And as we deploy AI towards more valuable objectives like scientific discovery and market research, most of the path will be unknown. A system responsible for outcomes therefore has to be capable of discovering smarter decisions through its lived experience.
The most valuable experience is locked inside companies
The most valuable experience is locked inside companies
Every action taken in the real world can produce new evidence. A biological assay reveals some property of the cell at question. A product launch reveals how customers behave. A lending decision eventually leads to repayment or default.
Every action taken in the real world can produce new evidence. A biological assay reveals some property of the cell at question. A product launch reveals how customers behave. A lending decision eventually leads to repayment or default.
Some domains, like math and coding, are legible to centralized model training. For these tasks, approaches can be fully tried and tested in-silico. But in the real economy, the most valuable learnings come from the real world.
Some domains, like math and coding, are legible to centralized model training. For these tasks, approaches can be fully tried and tested in-silico. But in the real economy, the most valuable learnings come from the real world.
Ideally, foundation models would learn directly from this experience and outcomes. But this data is out of reach for the labs. OpenAI’s CFO addressed this data gap earlier in the year: “I think ninety-five percent of the world’s information sits behind corporate firewalls.”
Ideally, foundation models would learn directly from this experience and outcomes. But this data is out of reach for the labs. OpenAI’s CFO addressed this data gap earlier in the year: “I think ninety-five percent of the world’s information sits behind corporate firewalls.”
To circumvent these walls, frontier labs are plowing billions of dollars into shallow proxies for the training signal. Expert-data markets recruit junior associates, scientists, and other young professionals to label data on the side, a mirage of the expertise that veteran employees learn from decades of experience. But these data farms can never reproduce the full context in which the expert judgement developed, and so the models cannot learn the judgement itself.
To circumvent these walls, frontier labs are plowing billions of dollars into shallow proxies for the training signal. Expert-data markets recruit junior associates, scientists, and other young professionals to label data on the side, a mirage of the expertise that veteran employees learn from decades of experience. But these data farms can never reproduce the full context in which the expert judgement developed, and so the models cannot learn the judgement itself.
Take a moment to realize how insane that is. We’ve somehow decided that it’s so hard to train language models that our only option is a game of data-telephone. Real world outcomes are observed by a company’s veteran employees, who train junior employees on menial projects, who then absentmindedly label simulated tasks in fake companies, which get aggregated into training mixtures and hopefully teach the model the most inoffensive capabilities to pass benchmarks without holding any real opinions about how work should be done.
Take a moment to realize how insane that is. We’ve somehow decided that it’s so hard to train language models that our only option is a game of data-telephone. Real world outcomes are observed by a company’s veteran employees, who train junior employees on menial projects, who then absentmindedly label simulated tasks in fake companies, which get aggregated into training mixtures and hopefully teach the model the most inoffensive capabilities to pass benchmarks without holding any real opinions about how work should be done.
The right path, as we’ve done for every other ML cycle, is to learn from real data. For long-horizon autonomy, we have to shift the locus of training from the frontier labs to the organizations which own the proprietary data flywheels. We have to bring the training to the data, inside the secure environment.
The right path, as we’ve done for every other ML cycle, is to learn from real data. For long-horizon autonomy, we have to shift the locus of training from the frontier labs to the organizations which own the proprietary data flywheels. We have to bring the training to the data, inside the secure environment.
Training turns experience into compounding capability
Training turns experience into compounding capability
After enough examples are captured, patterns of success start to emerge. These patterns are often sub-lingual, where useful behavior emerges statistically and solidifies as judgement.
After enough examples are captured, patterns of success start to emerge. These patterns are often sub-lingual, where useful behavior emerges statistically and solidifies as judgement.
Training makes these lessons persistent and generalizable. And once the foundations are learned, the model can develop increasing nuance in its decisions.
Training makes these lessons persistent and generalizable. And once the foundations are learned, the model can develop increasing nuance in its decisions.
In the same way that a new tennis player can’t learn subtle techniques to improve their topspin before learning to volley, models need to reach base levels of capability before developing true expertise in long horizon tasks.
In the same way that a new tennis player can’t learn subtle techniques to improve their topspin before learning to volley, models need to reach base levels of capability before developing true expertise in long horizon tasks.
In training, this shows up in the “cold-start problem” where models can be stuck in cycles of early missteps, so every rollout is deemed a failure and no signal is generated. But when training delivers a curriculum of mastery, the learnings and capabilities begin to compound.
In training, this shows up in the “cold-start problem” where models can be stuck in cycles of early missteps, so every rollout is deemed a failure and no signal is generated. But when training delivers a curriculum of mastery, the learnings and capabilities begin to compound.
In long horizon agents, this looks like increasing spans of work that the model can handle over time, until the entire lifecycle of a drug program, investment decision, or product plan is executed end to end.
In long horizon agents, this looks like increasing spans of work that the model can handle over time, until the entire lifecycle of a drug program, investment decision, or product plan is executed end to end.
Learning at machine scale will exceed human capability
Learning at machine scale will exceed human capability
Human expertise is bounded by the amount of experience a person can accumulate and process. Even the best experts see only a fraction of the decisions and outcomes generated across an industry.
Human expertise is bounded by the amount of experience a person can accumulate and process. Even the best experts see only a fraction of the decisions and outcomes generated across an industry.
Autonomous systems can break these constraints and learn across millions of examples. They can accumulate more experience, retain more of what they learn, and discover strategies that no individual human could have discovered in their career.
Autonomous systems can break these constraints and learn across millions of examples. They can accumulate more experience, retain more of what they learn, and discover strategies that no individual human could have discovered in their career.
We see no reason to assume that human performance is the upper bound for economically important tasks. And we expect that super-human capabilities will be critical to unlocking the abundance that AI is supposed to foster.
We see no reason to assume that human performance is the upper bound for economically important tasks. And we expect that super-human capabilities will be critical to unlocking the abundance that AI is supposed to foster.
Every autonomous company will need a training loop
Every autonomous company will need a training loop
When a company lets their proprietary intelligence compound at superhuman rates, it becomes impossible to compete with. Therefore we expect the market leaders in the next era of companies will be those that tightly couple training with lived experience. These company-wide training loops are inevitable.
When a company lets their proprietary intelligence compound at superhuman rates, it becomes impossible to compete with. Therefore we expect the market leaders in the next era of companies will be those that tightly couple training with lived experience. These company-wide training loops are inevitable.
TrainLoop is building the training layer for the autonomous company
TrainLoop is building the training layer for the autonomous company
The future will be owned by autonomous companies that learn from their own experience. For long horizon agents, this means that training and data flywheels are coupled and capabilities compound over time.
The future will be owned by autonomous companies that learn from their own experience. For long horizon agents, this means that training and data flywheels are coupled and capabilities compound over time.
TrainLoop gives companies the infrastructure to unlock their proprietary data, connect decisions to real-world outcomes, learn from the lived experience of their employees and agents, and train models to compound against their most valuable objectives.
TrainLoop gives companies the infrastructure to unlock their proprietary data, connect decisions to real-world outcomes, learn from the lived experience of their employees and agents, and train models to compound against their most valuable objectives.
We believe this leads to a world where organizations can operate at a scale beyond human capability, unlocking the abundance we all strive for while enabling a renaissance of human creativity.
We believe this leads to a world where organizations can operate at a scale beyond human capability, unlocking the abundance we all strive for while enabling a renaissance of human creativity.