TypeSafe AI CEO Diogo Almeida: Announcement of Noul Output "Leak"
TypeSafe AI CEO Diogo Almeida announced on social media that the company discovered a large-scale unauthorized leak of all customer Noul outputs, emphasizing the importance of transparency by directly providing a GitHub link for review.
The announcement specifies that the incident was discovered around 6:49 PM on Friday, with the linked content consisting of 101 values arranged from 0.00 to 1.00 in increments of 0.01; Noul is one of the primitives of its model used to return probability values between 0 and 1 for statements.
TypeSafe AI was founded in 2024 and launched its first model, Jev, from stealth mode in September 2026, completing a seed round of approximately $40 million led by DCVC; Jev is positioned as a machine-native System One model that outputs structured decisions with calibrated probabilities through primitives like Noul, Choice, and Score, rather than generating text.
Almeida previously worked at OpenAI on RLHF and InstructGPT-related projects, and the company emphasizes calibration decisions through reinforcement learning, aiming for intelligence to operate internally within software rather than in human dialogue; the announcement sparked significant interaction, with some responses focusing on output accuracy and calibration features.
The so-called leaked data only covers the entire probability space that the model might output, without involving any customer-specific inputs, keys, or proprietary data; the company has not reported any actual security incidents or customer impacts.
Funding and attention are directed towards developers and enterprise customers of structured decision models, with TypeSafe charging API fees for providing low-latency, calibrated probability outputs; teams mastering type safety and calibration interfaces benefit, as traditional text generation models face higher uncertainty costs in automated scenarios.
The announcement is presented in a serious disclosure format, with the actual content being a complete enumeration of the model output range, further highlighting the design of its product to treat probability as a first-class output rather than a post-processing result.
Source: Public Information
ABAB AI Insight
During his time at OpenAI, Diogo Almeida participated in RLHF-related work, helping to establish pathways for reinforcing model preferences through human feedback, and later left in 2024 to found TypeSafe AI, shifting focus from free text generation to structured decisions directly consumable by machines.
The capital path reflects seed funding directed towards the System One model and RLCD training methods, concentrating resources on building primitives like Noul to return calibrated probabilities, strategically targeting the demand for certainty and confidence in software automation rather than expanding conversational interactions.
This is similar to early paths embedding search or classification models into software systems, and akin to some financial risk control systems using probability scores directly as decision inputs; the current phase represents a transition from chat models to programmable intelligent components, with the output format itself constituting an interface barrier.
Essentially, this is a technological substitution: free text output is replaced by structured probabilities, mechanically achieved by limiting the output space and providing calibrated confidence, allowing downstream code to directly perform threshold control and routing, reducing prompt engineering uncertainty, while simplifying the "leak of all possible outputs" to the disclosure of known ranges.
ABAB News · Cognitive Laws
The smaller the output space, the lower the transparency cost
Calibrated probabilities replace prompt requests, with interfaces becoming barriers
Serious announcements package known universes, with attention as a product signal