Open Source

Competition in artificial intelligence is no longer driven solely by the raw power of models. The State of Open Source AI report shows that open models are rapidly closing the gap with proprietary solutions, particularly in terms of value for money. This progress does not mean they outperform across every use case: proprietary models still retain certain advantages in long-context processing, professional knowledge and overall polish. Nevertheless, it changes how businesses make decisions, as the prestige of a frontier model is no longer enough to justify its selection. The report also stresses that a model with downloadable weights is not necessarily open source. Without training data, preparation methods, training code, provenance information and a permissive licence, its actual capacity to be audited, modified or redistributed remains limited. The main obstacle is therefore becoming operational. Open models are widely adopted by developers, yet reach production less often because of infrastructure costs, security requirements, maintenance, deployment complexity and limited support. Value is consequently shifting towards the harness, the execution layer that manages context, memory, tools, permissions, budgets, observability and revocation. The more interchangeable the model becomes, the more enduring the advantage held by the system that knows when to call it, what data to provide and which boundaries to enforce. This shift also explains growing corporate interest in optionality: owning an executable copy, deploying on private infrastructure and switching providers offer protection against price increases, access restrictions and changing terms. The challenge is therefore no longer limited to identifying the most intelligent model. It is now about building a reliable, controllable system capable of acting on evidence.

Jev does not write text, code or summaries. The model introduced by TypeSafe focuses on a single task: choosing from several answers defined at the time of the request, then indicating how confident it is in that choice. This approach resembles a classifier, with one important difference: while a spam filter remains specialized in one particular decision, Jev is designed as a general-purpose model. It can select a category, decide whether to issue a refund or escalate a case, choose an action in a game, or identify one object from several possibilities. To demonstrate the approach, TypeSafe had Jev play Doom using a text description of the situation, without direct access to the screen, making ten decisions per second. Its appeal does not lie in claims of greater intelligence than ChatGPT or Claude, but in its speed and low cost. Responses arrive within a few hundred milliseconds, output tokens are free, and one billion input tokens cost $42. This efficiency could make previously impractical uses accessible, such as evaluating 10,000 articles one by one to determine whether they are relevant to a given question. The name Jev refers to the Jevons paradox: when a resource becomes more efficient, consumption can increase because new uses emerge. TypeSafe applies this idea to artificial intelligence. Jev is therefore not intended to replace generative models, but to expand the toolkit available to developers and automated agents. Available in early access, it can be tested through an API, a Python kit and a skill designed for coding agents. Its answers always conform to the expected format, but they can still be wrong: eliminating malformed responses does not guarantee that the choice itself is correct.

OpenHuman offers a personal artificial intelligence designed to run directly on the user’s machine, in a private environment that remains under their control. The project champions a local-first approach that reduces reliance on a mandatory remote service and avoids systematically exposing interactions to third-party platforms. Users can ask questions or assign tasks to their assistant while keeping the experience centred on their own environment. This approach addresses a straightforward need: accessing a personal AI without giving up control over how it operates. OpenHuman also places a strong emphasis on simplicity. The application aims to minimise advanced configuration, complex interfaces and unnecessary technical steps. Instructions provided in the repository guide users through installation, helping them get started quickly and interact with the application directly in everyday use. Clear guidance supports the setup process, allowing users to focus on their conversations with the assistant rather than on technical complexity. The goal is therefore not to multiply features, but to make local artificial intelligence clearer and more accessible. The project also follows an open-source model. Its code is publicly available, making it possible to understand its structure, examine how it works and adapt it to specific needs where necessary. This transparency strengthens the promise of local control at the heart of the application. With permanent memory highlighted from the outset, OpenHuman aims to make the assistant a lasting presence that remains closely aligned with its user. The project is therefore built around three complementary ideas: a personal AI running locally, guided installation and an open architecture. OpenHuman seeks to provide a straightforward alternative to fully remote services, without imposing external dependency as the price of accessing artificial intelligence that is private, understandable and ready for direct use.