TypeSafe AI Launches Jev, a Model Built for Structured Decisions Over Text Generation

·68 views

TypeSafe AI has released Jev, its first public 'System One' model, designed to make fast, structured decisions that software can consume directly rather than generating natural language responses. The model is inspired by Daniel Kahneman's System 1 and System 2 thinking framework, positioning Jev as a quick, intuitive decision-maker suited for backend applications. Instead of parsing generated text into booleans or scores, developers define typed questions and receive structured outputs such as yes/no answers, categorical choices, or scored evaluations in parallel. Jev uses a new architecture combining a parallel sampler and a training method called Reinforcement Learning for Calibrated Decisions (RLCD) to deliver calibrated probabilities alongside each decision. The model targets use cases like routing support tickets, flagging sensitive prompts, approving agent actions, and triaging transactions where structured outputs matter more than human-readable explanations.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

RBA Questions AI Productivity Gains as Data Centre Spending Strains Economy

Reserve Bank of Australia Assistant Governor Sarah Hunter told the UBS Australasia Conference in Sydney that evidence of AI meaningfully lifting productivity or reducing employment in Australia is still limited. She noted that data centre capital spending has risen roughly tenfold in recent years, placing demand pressure on the power grid, construction sector and engineering workforce. Hunter warned this investment could fuel persistent inflation and keep interest rates high. She also flagged that shared cloud and foundation model providers create new systemic risks to financial stability.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

AI Hackathons Need Scoped Budgets, Not Shared Keys or Cards

AI teams often struggle less with ideas than with a clean way to pay for experiments, and this quickly becomes a developer-experience problem. Default workarounds like shared subscriptions or shared API keys are convenient at first but create security risks, accounting confusion, and unclear ownership. The article proposes treating experimental AI budget as a temporary, bounded capability, similar to short-lived branches or limited tokens. Contributors would receive a defined credit envelope that limits spending without exposing the sponsor's full account or credentials. The author notes that prepaid credits or gift cards can support this model, though rules vary, and that such budgets should still be paired with engineering controls like server-side secrets, input validation, usage logging, and per-user limits.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Guide Explains Using One Free IPv6 /48 Across Two VPS Providers Without Renumbering

A network operator describes how to announce a single IPv6 /48 prefix from two VPS hosts in different cities so both sites can carry traffic. The article recommends setting local-preference and AS-path prepending to make one site primary and the other a backup, while keeping ROA maxLength at /48. It also covers export filters, drain and withdrawal procedures during maintenance, and choosing between anycast and DNS failover for stateless versus stateful services. The piece is written by a leasing company's representative, who discloses that the company's IP space is featured in the walkthrough.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Python script flags failing endpoints and slow requests in fictional iPhone log demo

A DEV Community tutorial shows how to build a small Python log analyzer that runs on an iPhone using only the standard library, requiring no server or network connection. The script parses simplified, fictional access-log lines, rejecting malformed entries while keeping their line numbers visible. It counts HTTP 5xx server failures, computes a 30% server-error rate across ten parsed requests, and reports a 195 ms median latency. It also ranks the slowest requests and groups failures by endpoint, with checkout showing the most errors. The author stresses that such summaries point to where to investigate rather than revealing the underlying cause.

0 comments Read more at DEV Community