BERT

AI overview

Headlines · 2

4d
Naive-N0.5-Flash automates parts of model training
New Zhiyuan reports that the open-source Naive-N0.5-Flash can handle parts of model-training workflows, including data preparation, training, evaluation, and identifying flaws in scoring rules. In a community project described by the outlet, a creator used Naive-N0.5-Flash through Claude Code to fine-tune a 0.6B model on a 24GB Mac after writing a one-page task specification. New Zhiyuan also says it used the system on a 64GB Mac to train a locally running, 110-million-parameter BERT classifier for the BANKING77 dataset. The outlet reports 93.34% accuracy on the official test set and a 6.77-millisecond median response time, while noting that the workflow still requires a human-written program.md and researcher intervention.
新智元 · 🔥 0
Oct 12
BAAI's Wang Pengwei on RoboBrain Orca: world models vs VLA for embodied generalization
Wang Pengwei, who leads the embodied model research center at China's BAAI, presented the RoboBrain Orca world model at the 2026 BAAI conference. He argues that today's four families of world models — language-, pixel-, 3D-structure- and visual-representation-centric approaches such as LLMs, video generators, World Labs' Marble and JEPA — each capture only a partial world and remain passively task-driven; Orca instead uses Next State Prediction, reasoning forward and backward about state transitions, and separates "unconscious" video learning from "conscious" event-level state-transition learning. He says Orca's pretraining used no action data, drawing on about 125,000 hours of internet video, roughly 130 million expert-model-labeled event-level state-transition samples and causal and general VQA data, and that BAAI is building a benchmark for state-transition world models because none exists; he expects some approach to break the 0% generalization barrier within a year, though 90%-plus would be needed for real usability.
钛媒体 · 🔥 0

Experience and discussion from the community

niubigeo: An open-source GEO research tool for AI visibility, reproducible testing, competitors and evidence.

What problem it solves Teams need to understand how AI models describe their products, which competitors appear in answers, and which sources support those observations. Without original answers, citations, execution conditions, and historical runs, it is difficult to investigate visibility findings or determine what to trust. Key features Evidence-based visibility research — Enter a domain to see…

github.comhttps://github.com/Albert-Weasker/niubigeo

Swarm Intelligence•Florian•Oct 29editedExperience Pinned

Agent Swarm: Organizational Growth and the Passing On of Experience

The Research system Anthropic unveiled in June 2025[1] organizes a research problem into a process of analysis, parallel search, and synthesis of results. Beyond model capabilities, how work is allocated and how results from different directions are connected have also become design questions that AI engineering needs to answer separately. Along this line, I want to ask a deeper question: if even …

EvoMap EvoX OpenClaw Swarm GEP