Astra for Law
Astra for Law is a product by OpenAI designed to assist legal professionals with AI-powered tools. It aims to streamline legal research and analysis using advanced AI technology. (openai.com)
Show HN: AutoBot – live voice control for long-running AI work
github.com12 points3 comments
AutoBot is an open-source harness that wraps native ChatGPT on a Mac to run long, multi-application workflows reliably and privately. It claims substantive benchmark wins - 18.5% higher task completion than a published OpenAI Sol Max baseline, #1 on the AssistantBench hidden-test leaderboard (50.70% accuracy), and strong OSWorld 2.0 results (32.41% binary, 64.28% partial accuracy) - by using a self-improving operating contract rather than a new model. Hard workflows become upgrades to the harness: AutoBot repairs and validates its own changes, preserves hierarchical memory on disk with task-specific retrieval and nightly consolidation, and keeps persistent task graphs, atomic checkpoints and independent supervision so work can be resumed or verified without retraining the underlying model. Local CPU orchestration, compiled context, and reusable proofs reduce repeated inference and direct model budget to difficult judgments.
Operational specifics emphasize safety, provenance and audience control. Privacy zones separate personal, family, work and shared context; every external write must pass account/destination/content checks and a signed-in first-party route; and completion requires an independent validator label and destination readback. Communication tones are stored as compact, channel-specific profiles to prevent cross-contamination. AutoBot bundles a desktop ChatGPT app with an "Autobot Workspace" (AGENTS.md operating contract, goals, skills, plugins, Computer Use policy) and explicitly depends on the host ChatGPT plan, region and permission limits rather than replacing the model.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
More in AI
Canto: A speech model built for the real world
Canto is a new speech model by Wispr AI designed for real-world dictation, performing well with background noise and varied conditions. It achieved the lowest word error rate among tested models in real-world scenarios, outperforming competitors like Google and OpenAI. (wisprflow.ai)
Tests of the New AI Siri
Apple's new AI Siri, launched with OS 27, can access and interpret data across Apple apps to assist with personal queries, replacing many traditional searches. It demonstrated impressive capabilities in tests, though it still has some limitations. (pogueman.substack.com)
Towards Self-Driving Codebases
Agents can perform complex code migrations and rewrites with proper guardrails, but most software development still relies on human engineers. Future self-driving codebases will focus on enabling engineers to concentrate on high-level ideas, with automation handling routine tasks. (blog.detail.dev)
4-Bit Rotational Quantization: -45% RAM, <1% recall drop vs. TurboQuant
Weaviate's 4-bit rotational quantization reduces memory usage by 45% with less than 1% recall loss, improving vector search efficiency. Enhancements like SIMD support for Fast Walsh-Hadamard Transforms significantly boost encoding speed across hardware platforms. (weaviate.io)
Palantir's Karp: AI needs to have 'reasonable guidelines,'
Palantir CEO Alex Karp urged for reasonable guidelines and regulations for artificial intelligence to address its inherent risks. He also suggested that AI labs might need to be nationalized to ensure accountability for harmful technology development. (cnbc.com)