TypeSafe AI announced approximately $40 million in seed funding on September 15, led by venture firm DCVC, and simultaneously emerged from stealth with its first model designed to power what the company calls "fast, structured decisions." The San Francisco startup, founded in 2024 by former researchers from OpenAI and Google Brain, opened early developer access the same day to Jev, a model the team argues represents a fundamentally different approach to machine intelligence.
The company would not disclose its valuation or name investors beyond DCVC. LinkedIn data suggests TypeSafe has between 11 and 50 employees.
A ChatGPT co-author strikes out alone
Diogo Almeida, the company's co-founder and CEO, carries notable credentials from his OpenAI tenure. He co-authored the March 2022 InstructGPT paper and appears among the contributors listed on GPT-4. In a LinkedIn post last week, Almeida described his role more bluntly: he "co-invented ChatGPT and RLHF," the reinforcement learning technique that helped make conversational AI feel natural—a self-attributed claim not independently verified.
His co-founders bring a mix of computational biology and research engineering. Erik Gafni, now CTO, previously co-founded Ravel Biotechnologies and held roles at Freenome and Invitae. Sasha Sheng, the COO, came from Meta's FAIR research engineering team and has published work at NeurIPS and ECCV.
Built for decisions, not prose
TypeSafe positions Jev as something it calls a "System One Model." The reference borrows from Daniel Kahneman's behavioral psychology—System One thinking is fast and instinctive, while System Two is deliberate and analytical. In this framing, Jev handles rapid, structured choices rather than generating open-ended text.
The model returns three types of outputs the company labels Choice, Score, and Noul, each paired with probabilities and confidence scores. Questions get evaluated in parallel against a shared input "state," according to developer documentation. Because outputs conform to predefined schemas, TypeSafe says they eliminate type errors.
Response times clock in between 70 and 500 milliseconds in the company's internal West Coast testing. TypeSafe claims the model runs "two orders of magnitude faster and more efficient" on these narrow tasks than frontier large language models, though those systems were built for broader, more flexible use cases.
Training relies on what TypeSafe calls RLCD, or Reinforcement Learning for Calibrated Decisions. Pricing sits at $0.042 per million input tokens. Output tokens are free, which the company describes as "too cheap to meter," echoing an old nuclear energy promise that never quite panned out.
Benchmarks without an auditor
TypeSafe published workflow evaluations on a dedicated microsite comparing cost and latency against unnamed LLM baselines across security incidents, invoice processing, customer service, and agent observability tasks. Consensus labels for these tests were generated by averaging outputs from OpenAI and Anthropic models, the company said. No independently verified performance or cost benchmarks have been published yet.
"TypeSafe was founded to pursue an alternative path for AI research, focused on machine-native AI," Almeida said in the announcement. "Most intelligence should eventually live inside software, running quietly in the background."
James Hardiman, a general partner at DCVC, offered the sort of endorsement venture investors tend to provide. "TypeSafe is approaching one of the biggest remaining challenges in AI," he said.
Seed rounds keep climbing
DCVC has been writing unusually large seed checks this year. In August, the firm participated in a $100 million seed round for Callosum, which focuses on AI orchestration. In January, it led an $80 million seed for Proxima, a biotech company. The escalating size of these early-stage investments reflects a broader pattern across venture capital, where investors are placing bigger bets on technical founders with elite pedigrees before products fully prove themselves in the market.
TypeSafe participated in the AWS Generative AI Accelerator cohort announced in July 2024, well before it disclosed any funding. That timing suggests the company had been building quietly for months, perhaps longer, before its public debut.
Early access requests for Jev are now open. In a blog post, Almeida framed the offering this way: "Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out."
Whether developers will embrace that paradigm remains to be seen. The AI landscape is already crowded with models claiming speed advantages or cost efficiencies. TypeSafe is betting that a narrow focus on structured decision-making, rather than general-purpose language generation, will carve out a distinct market. The approximately $40 million in seed funding buys time to find out.
