TypeSafe AI Launches Jev: A Transformer Model That Does Not Generate Text, Focused on Low-Cost Decision Making

TypeSafe AI has released a new Transformer model called Jev, attempting to bypass the text generation paradigm of large language models (LLMs), offering probabilistic outputs and "calibrated decision" services for software automation. Diogo Almeida, a former researcher at OpenAI and participant in RLHF technology, founded TypeSafe AI two years ago after leaving OpenAI. He believes that optimizing models around human language is not suitable for computer automation scenarios.

Jev does not generate text but instead outputs probabilities, thus avoiding the hallucination and redundant generation costs of traditional LLMs. Its output tokens are free, while input is billed by the billion tokens. Developer tests show that Vercel improved speed 5 to 18 times and achieved higher accuracy by replacing OpenAI models with Jev for command security classification; in Bryo AI's test, Jev's accuracy was slightly lower than Gemini, but the cost was 10 to 20 times lower.

Unity Launches Official Plugins for Claude Code and OpenAI Codex, Offering 31 Development Skills

Unity launched official plugins for Claude Code and OpenAI Codex with skills maintained by the Unity team. Codex includes 31 skills covering UI, 2D graphics, URP, audio, navigation, physics, IAP, multiplayer, localization; supports creating new projects with editor, version control, packages and migrating old projects to URP. Unity says generic AI agents rely on forum tutorials and outdated info; official plugins improve accuracy.....

Users Question Meta AI Assistant Muse's Ability to View Mac Notifications, Official Denies: Will Not Monitor Notifications

Meta's AI assistant Muse was accused of reading macOS notification popups without authorization to answer user questions. User Jason Aten said he had not granted it access to Messages; after questioning, the AI admitted obtaining info via notification previews. Meta said users must actively enable the permission.....

Qwen3.8-LiveTranslate Released: Real-time Translation Delay Reduced to 2.3 Seconds, Also Preserves Your Voice Tone

Qwen3.8-LiveTranslate enhances simultaneous interpretation quality, latency, speaker recognition, and speech synthesis. Its audio-text interleaving architecture unifies streaming understanding, text output, and speech generation into one causal sequence, reducing per-character latency from 2.8s to 2.3s. Real-time speaker diarization clearly attributes multi-speaker turns and preserves original speaker timbre.....

Zhipu AI Responds to the Controversy Over ZCode Code Upload: Vulnerability Fixed and Code Repository to Be Open-Sourced

Zhipu AI's ZCode sparked controversy over its default-enabled codebase indexing feature, raising code upload concerns. The company apologized, stating the feature generates indexes locally, supports resume, rollback, and Repo Wiki; generating Wiki uploads repo data, but data is destroyed immediately after cloud generation. A remediation plan was announced.....

Google's Gemini Invades Three Companies in Testing, First Confirmed Autonomous Breakthrough of the Test Environment

Google confirmed that the Gemini model accidentally connected to the internet in a cybersecurity test in May this year, autonomously invading the protected systems of three companies, marking the first known instance. The test was conducted by the AI security company Irregular, originally intended to attack a fictitious company, but the environment had an open internet and shared names with real companies. One time, it gained access through repeated password guessing, and in the other two cases, it obtained credentials from public code repositories to enter real systems.