Nous Research highlights growing buzz around Hermes Agent, with observers calling it 'insanely good' and betting the team can build a large company if execution holds.
Willison notes ChatGPT Work (mobile/web mode) ships an integrated browser with screenshots and can deploy web apps straight to Cloudflare Workers ('ChatGPT Sites'), with discoverability the main remaining gap.
Nous Research promotes DeepSeek V4 Flash at 90% off via novita_labs, claiming it outperforms a top model on Terminal-Bench 2.1 while being over 1,000x cheaper, with other models 20-50% off — a fresh salvo in inference pricing.
Adds a future-state value estimator on top of VLA backbones to improve RL post-training for robotic manipulation, beating single-frame value-based baselines.
Sequences heterogeneous tasks to reduce interference and catastrophic forgetting in shared low-rank adapter fine-tuning, improving transfer and retention over naive shared-LoRA.
Evaluates agents on following a user-defined schema with source evidence grounding for enterprise document workflows.
A benchmark where geometry, timing, resources and rule interactions all matter, turning decisions into measurable outcomes to stress-test models beyond typical game benchmarks.
Verifies candidate answers via tool-flow rather than natural-language reflection only, improving answer verification and self-correction over reflection-only baselines.
A lightweight CLI that serves several local LLMs on Apple hardware via the MLX stack.
A self-hostable real-time avatar for on-premises talking-face and avatar apps without cloud dependencies.
An open-source system giving agents a shared company-wide knowledge base to consult before acting.
A lightweight, single-binary job queue designed to drop into any stack with minimal setup.
A web tool that checks whether major AI assistants recommend your brand in their responses, for marketers and founders tracking AI referral visibility.
Source:
