Who is TypeSafe AI? The RLHF pioneer behind Jev

Why a model that does not talk is worth tracking

TypeSafe AI is a model company founded by Diogo Almeida, a researcher who helped build reinforcement learning from human feedback (RLHF) at OpenAI, the training method that taught large language models to produce output humans prefer reading. Almeida spent years making models better at sounding like people. TypeSafe AI is his argument that sounding like people is the wrong goal for most of what automation needs to do.

The company spent roughly two years in stealth and emerged having raised $40 million. Its core claim is that LLMs optimise for human language, and computers speak a different language. The practical consequence of that gap, in Almeida’s framing, is that every time an agent calls a tool, queries a database, or hands off to another process, it is translating through a medium built for human readers. TypeSafe AI is building a model that skips that translation.

For leaders tracking where agentic AI architecture is heading, the TypeSafe thesis is worth understanding on its own terms, not because it is certain to win, but because the background of its founder gives it more foundation than most stealth-era pitches carry.

The RLHF background and what it implies

Almeida’s work on RLHF at OpenAI was about shaping model behaviour through human preference signals. Annotators rated outputs, those ratings trained a reward model, and the reward model guided further training. The result is a model that is measurably better at producing text humans find clear, helpful, and appropriate.

That is genuinely useful when the consumer of the output is a human. When the consumer is a downstream system parsing a JSON payload or routing a task through an agentic AI orchestration layer, human-preference optimisation adds friction without adding value. The model learns to write well. The pipeline has to then parse what it wrote.

Almeida’s position is that this is an architectural mismatch, not a gap that prompt engineering closes permanently. The fix is a model trained against a different objective from the start, one where the reward signal comes from whether the output executes correctly in a machine context, not whether a human annotator preferred it.

What System One means and where Jevons fits

TypeSafe AI calls its model category System One, borrowed from Daniel Kahneman’s two-system model of cognition. Kahneman’s System 1 is fast and automatic; System 2 is slow and deliberate. The company’s use of the name maps fast, low-overhead task execution to System 1 and positions chat-style reasoning models as the System 2 equivalent: useful for complex deliberation, expensive for routine operations.

The Jevons reference, embedded in the company name itself, points at Jevons paradox: the observation that efficiency gains in resource use tend to increase total consumption rather than reduce it. Applied here, the argument is that making agents faster and cheaper to run will expand the number of tasks routed through them, not compress it. The market for execution-optimised models grows as adoption grows.

That framing aligns with where enterprise deployments are already heading. Salesforce reported Agentforce handling 3 million support conversations with Agentforce ARR reaching $800 million, up 169% year over year. JPMorgan has more than 450 agentic AI use cases running in production daily. At that volume, each percentage point of efficiency in model inference compounds into real operational cost. A model that does not need to produce readable prose to complete a task is, at minimum, cheaper per call.

flowchart TD
    A[Agent receives task] --> B{Output consumer?}
    B -->|Human reads response| C[Chat model: optimise for human language]
    B -->|System parses output| D[Execution model: optimise for structured output]
    C --> E[Natural language response]
    D --> F[Typed action or structured payload]
    F --> G[Downstream system executes]
    E --> H[Human acts or re-prompts]

What the vendor landscape looks like around this bet

TypeSafe AI is not proposing that chat models disappear. The argument is more specific: for the portion of AI agents for automation where no human is in the loop between model call and system action, a model trained on machine-readable output objectives should outperform one trained on human preference.

That portion is growing. Klarna’s AI agent handled the workload equivalent of 853 employees by Q3 2025, saving $60 million. Morgan Stanley’s DevGen.AI reviewed 9 million lines of legacy code, saving 280,000 developer hours. In both cases, the model’s output was consumed by a system or a structured review process, not a person reading a chat thread. The natural language a chat model produces in those contexts is overhead.

The current tooling partially addresses this through structured output modes, function calling, and typed schemas layered on top of existing models. Vendors including, by way of category example, Prefactor have built governance and routing layers that work with these structured outputs. TypeSafe AI’s argument is that layering structure on top of a model optimised for prose is a workaround, not a solution. Whether that argument holds at production quality against fine-tuned versions of existing frontier models is the empirical question the company’s emergence from stealth will start to answer.

flowchart TD
    A[Task enters pipeline] --> B[Orchestration layer]
    B --> C{Model selection}
    C -->|Human-facing response needed| D[Chat-optimised model]
    C -->|Machine-to-machine action| E[Execution-optimised model]
    D --> F[Readable text output]
    E --> G[Structured action payload]
    G --> H[Tool call or API action]
    H --> I[Result logged to agent memory]
    F --> J[Human review or reply]

What this says about where automation AI is heading

The TypeSafe thesis implies a bifurcation in the types of AI agents and the models that power them. One branch handles tasks where human interaction is the point: customer service agents, conversational AI agents, sales and support interfaces. The other branch handles tasks where the entire loop runs between systems, and a human only sees aggregated outcomes or exception reports.

Gartner projects that 40% of enterprise applications will integrate task-specific AI agents by end of 2026, up from less than 5% in 2025. As that share rises, the overhead cost of routing machine-to-machine tasks through human-language models becomes more visible on infrastructure budgets and latency measurements.

A founder who spent years teaching models to optimise for human preference, then concluded that preference is the wrong signal for machine contexts, is making a technically informed argument. The $40 million raise indicates investors read it similarly. Whether the model performs as the thesis predicts is a separate question from whether the thesis is coherent. The thesis is coherent.

For leaders building multi-agent systems or evaluating agentic AI design patterns, TypeSafe AI is worth tracking as a signal, not because stealth-era companies always deliver, but because the category problem it is addressing, the mismatch between human-language models and machine-language pipelines, is real and measurable in production today. Understanding how AI agents work at the model level will matter more as infrastructure choices lock in over the next 18 months.

Where to start

If you are mapping your agent stack against the infrastructure decisions that TypeSafe AI’s emergence surfaces, the clearest first step is understanding where your current pipelines consume model output directly versus where a human is still in the loop. Take the agent readiness assessment to get a structured view of where your organisation sits before the vendor landscape narrows further.