Dragonfly Invests in Decision Model Company Typesafe
Dragonfly is investing in Typesafe's latest round of financing, with its product Jev serving as a decision model aimed at exchanging constrained outputs for advantages in cost, throughput, and latency.
Haseeb Qureshi pointed out that large language models are versatile but slow, costly, and unreliable, making it difficult to embed them in software at runtime, especially in financial scenarios.
The decision model sacrifices the versatility of language generation, significantly outperforming large language models in cost, speed, and reliability for tasks such as if-else, buy-sell, labeling, selection, or clearly defined state machines.
Since its launch less than a month ago, over 25% of Fortune 500 companies have adopted Jev; OpenAI, Microsoft, Amazon, and Cloudflare have since launched similar decision models.
Typesafe was founded by former OpenAI researcher Diogo Almeida, who employs a community-first strategy and emphasizes maintaining a lead over replicators in model training, services, and products.
Funding is shifting from general large models to specialized decision models, with developers and enterprises embedding runtime decisions into code, benefiting from low-cost and high-reliability specialized model providers, while general models face pressure in embedded scenarios.
Source: Public Information
ABAB AI Insight
Diogo Almeida previously participated in RLHF-related work at OpenAI and founded Typesafe in 2024, launching Jev in September 2026 after two years of stealth development, positioning it as a System 1 model that returns calibrated probability-based decisions rather than text.
The capital path focuses on constraining model capabilities to structured outputs to reduce runtime invocation costs and latency, attracting Fortune 500 companies to embed decision logic into software, with institutions like Dragonfly supporting its continuous training and service optimization through investment.
This is similar to the early ASIC versus general-purpose processor path, currently in an expansion phase transitioning from embedding everything with general large models to specialized decision models handling runtime logic, akin to specific domain chips replacing general computing.
Essentially, this represents a technological substitution and industrial chain reconstruction: large language models handle high-dimensional code generation, while decision models manage low-dimensional runtime decisions. The mechanism lies in the fact that after problem constraints, specialized models achieve orders of magnitude advantages in cost and reliability, shifting embedded intelligence from general models to specialized layers.
ABAB News · Cognitive Laws
General models write code, specialized models make decisions
Constrained outputs yield speed and cost advantages
When the market is large enough, replicators quickly follow, but leading relies on continuous optimization