A study finds that AI models can detect some misbehavior in agent logs, but higher detection rates also mean more warnings about harmless actions. The results suggest auditors may be more useful for estimating how often problems occur than for safety checks that must catch every incident. Andy Hall evaluated several models on logs from METR’s MALT dataset. At a 5 percent false-alarm rate, GPT-5 caught nearly 98 percent of 47 reward-hacking cases, while Gemini 3.8 Flash found about half. Auditors caught 71 to 81 percent of sandbagging cases, most of which were synthetic.
Prism ML’s ternary-weight Ternary-Bonsai-2-2B reduces the language model to about 0.52 GB, versus roughly 3.8 GB at FP16. It retains a 262K-token context window and image input, with the vision component loaded separately. Weights take values of −1, 0, or +1; additional scale metadata raises the actual storage cost. Image input requires an optional 0.36 GB Q8_0 mmproj file.
Grok Bot now offers dedicated bot email addresses and tagged aliases that can trigger different routines. The guide presents 50 use cases for this and other updates, while noting that availability and setup requirements vary. Bots can have dedicated addresses and separate aliases for receipts, leads, or research newsletters. Incoming email can start the matching routine. The guide recommends testing each alias before routing real business mail and requiring approval before any external message is sent.
Xiaomi Research released Xiaomi-OCR-0, a 0.8-billion-parameter model for document understanding and OCR under the Apache-2.0 license. The company says it competes with larger systems on several document benchmarks. Built on Qwen3.5-0.8B-Base, the model supports page parsing, text extraction from selected regions, schema-based information extraction, and visual question answering. Xiaomi reports training on roughly 170 million samples and a mean score of 83.2 across five OCR-VQA benchmarks.
Third Intelligence has released Haas-ja-v1.0, an open-weight Japanese speech recognition model under the Apache 2.0 license. In the company’s five-dataset evaluation, it scored 6.12% strict CER; its reported hourly cost estimate is roughly 75–100 times lower than GPT-Transcribe’s. Haas-ja-v1.0 has 615.8 million parameters and was trained on about 46,000 hours of data. The company reports the best strict CER among evaluated open models on asr-ja-bench v1.1, which contains 18,318 utterances and 23.38 hours of Japanese speech.
Google DeepMind says PlanAssist reduced average review time for householder applications by 58% in pilots across three local authorities. The tool is entering a 12-month live beta with 12 more authorities, while qualified officers continue to review and approve every assessment. The Gemini-powered tool checks application documents, reviews local constraints and planning policies, and drafts assessments with citations. The company reports that officers approved more than 70% of generated reports without edits.
Open-source Bark can generate speech, music, laughter, and sound effects from a single text prompt. The source says it supports 13 languages and more than 100 voice presets, but it may alter wording or add unexpected sounds. Bark generates audio clips of about 13–14 seconds and is released under the MIT license. The source reports about 12 GB of GPU memory for the full models, with smaller variants available for lower-memory use.
AutoTrust Labs released JEV-27B v1.1, a 27-billion-parameter model that returns option probabilities rather than free-form text. The company reports a score of 67.31 across 37 benchmarks; the model supports multiple-choice and yes/no questions only. The model selects among supplied options and returns a probability distribution. It stands out on tool selection and intent classification, but is weaker on open-ended reasoning and some other tests.
OpenCode Browse provides an MCP server and a Chromium extension. It lets coding agents control your actual browser while keeping your logins, tabs, and extensions intact. Agents can use the browser you are signed into, with its existing tabs and extensions.
The open-source DS Bot plugin turns DeepSeek Harness into teams of bots with memory and task handoffs. The project is at version 0.2 preview and has passed 4,600 GitHub stars. DS Bot is available as a desktop app, a DeepSeek Harness plugin, or a Docker image. It offers persistent bot memory and per-bot model selection; in group chats, bots can hand off tasks through @mentions.
OhMyGame combines game design, asset creation, coding, and playtesting in an open-source Electron app. It is available for macOS (Apple silicon) and Windows x64 under Apache 2.0; model-provider usage charges may still apply. Developers can ask its AI agent to draft a design document, create art, write code, and run the game to check for visible failures. Web games can use Three.js, R3F, or Phaser, and workspace files use standard formats that can be opened in other editors.
DeepSeek has released dsh-superagent-prototype, an MIT-licensed TypeScript research prototype scoped to one GitHub repository. Designed to act on requests or written rules while minimizing progress chatter, it is explicitly not production-ready, according to the company. The agent can receive GitHub events and connect to Feishu. After an authorized request, it works in a separate session and directory, then returns the result to the conversation where the request began; it can respond if blocked by a decision requiring human input.
