
Know AI. See What’s Next.
Slide 1 of 5: 姜宇辉的AI分身:知识助手与“另一个自己”之间
Latest news

Karpo Decision goes open-source; 31B model ties first on a specified Decision Index version (opens in a new tab)
MachinePulse releases 0.8B–31B decision models that score options in one forward pass. The 31B model ties first at 73.9 in the November 12 Decision Index 0.4.1 snapshot. Latency and costs depend on test configuration; single decisions differ from full workflows.

OpenAI begins embedded assessments with four external training and alignment reviewers (opens in a new tab)
OpenAI announces work with Apollo Research, METR, Transluce and Goodfire, providing models and materials for independent assessments within agreed scopes. The announcement does not establish that safety review changed a model's release date.

Grid Robotics proposes an embodied-system framework separating models from closed-loop execution (opens in a new tab)
A Grid Robotics position paper proposes capability and responsibility levels around S0/S1/S2 roles, independent verification and time budgets. It examines the engineering needed to connect general-purpose models to physical systems; the framework is the team's proposal.

ToolSpace synthesizes cross-environment tasks from tool dependencies for long-horizon agents (opens in a new tab)
Researchers from Renmin, Fudan and other universities introduce ToolSpace, building tool-composition graphs from execution traces to synthesize verifiable tasks. Its 744B model averages 56.6 on the paper's seven benchmarks, matching the listed Claude Opus 5 average.

Axion 1.0 opens an agent-training pipeline using two-stage SFT and RL for scientific tool use (opens in a new tab)
Researchers from Duke, ByteDance and other institutions present Axion 1.0, using supervised fine-tuning for reasoning and information acquisition followed by reinforcement learning. The paper reports mathematics, science and information-search results and releases training materials.

RoboRisk tests robotic semantic hazards; LEASH reduces harm within specified tasks (opens in a new tab)
RoboRisk evaluates model-controlled robots on 92 simulated and thirteen physical tasks, testing LEASH’s three gates. Some configurations incurred no harm in the physical sample; this does not guarantee safety in open environments.

Pyromind’s EchoMind explores continual enterprise learning from feedback, training and evaluation (opens in a new tab)
Pyromind outlines EchoMind, designed to connect business feedback, data processing, training and model updates. The report covers initial FDE adaptation, human approval and business evaluation; customer metrics reflect interview examples.

Danqi Chen joins Noeri part-time as the startup raises $110M for continual-learning AI (opens in a new tab)
Danqi Chen says she joined Noeri part-time. Founded by Yejin Choi and Carlos Guestrin, it advocates Plural Intelligence, using enterprise data to optimize models, tools and workflows with human review of every change.

AMD announces World Labs acquisition closing; Fei-Fei Li becomes EVP and chief scientist (opens in a new tab)
AMD announced completion of its World Labs acquisition on November 16, with Fei-Fei Li reporting directly to Lisa Su. Regulatory filings place closing on November 13 and record approximately 14.98 million common shares issued to former shareholders.

Sage-S2.1 Flash opens scientific software workflows and reports agent benchmark results (opens in a new tab)
Sage-S2.1 Flash supports scientific data and demonstrates 3D Slicer, Blender, QGIS and CAD workflows. Its team reports 90.3 on MLE-bench Lite; rankings apply to the listed models and evaluation settings.

NetEase Tianma launches with interactive dramas, branching saves and paid content (opens in a new tab)
Dujia tests Tianma’s interactive and conventional short dramas, branching stories, saves and paid unlocks, and outlines creator incentives. The catalog remains limited, and long-term retention still needs observation.

How small HTML files generate 3D worlds: Three.js, procedural assets and AI prompts (opens in a new tab)
Starting with a 1.4 MB game example, the author explains how Three.js and procedural generation reduce external assets, then shares prompts for 3D scenes, a Rubik's Cube and animation. File sizes, visuals and model preferences apply to the examples described.

Seekee co-founder on Latin America's AI browser market: local needs and bundled subscriptions (opens in a new tab)
In an interview, Xu Leyang discusses Seekee's market choices, user research and bundled features for Latin American families' photo editing, light office work and learning needs. Download and business-scale figures are self-reported and do not equal paying-user counts.

Altman on Huang's departure: personal empathy, no judgment on reinstatement (opens in a new tab)
In a Semafor summit interview, Altman discusses Jensen Huang's departure and NVIDIA's role in AI, saying he lacks enough information to judge reinstatement. The report also covers chip demand, data centers, AI research and the public markets.

Lightwheel launches RoboFinals RL Envs for long-horizon robotic simulation tasks (opens in a new tab)
Lightwheel introduces robot reinforcement-learning environments during CoRL, demonstrating an over-eight-minute simulated car-repair task with dexterous hands, tools and deformable objects. The product links task construction, calibration, training and independent evaluation.

PaXini's Xu Jincheng on tactile sensing, embodied data and robotic execution (opens in a new tab)
In an ALIGNMENT conference talk, PaXini founder Xu Jincheng describes tactile sensors, data collection, VTLA models and execution hardware, discussing how perception, intelligence and execution work together. Product capabilities are claims disclosed in the talk.

INFevo announces over RMB 100M in pre-A funding for cell models and research platforms (opens in a new tab)
INFevo discloses a pre-A round exceeding RMB 100 million for model R&D, research products and team development. The report describes OCOO-T's experimental metrics and platform plans; results apply to the stated datasets and tasks, not clinical efficacy.

Some DeepSeek users encounter long-chat limits; the ecosystem's direction remains unclear (opens in a new tab)
Users report long-conversation limit notices in DeepSeek's app as its community expands through APIs and DSH plugins. The article discusses product and ecosystem roles, but no official explanation links the limits to research priorities or resource allocation.

Doubao adds map-based store discovery linking merchant details, deals and navigation (opens in a new tab)
Dujia tests Doubao’s Find Stores feature, which displays nearby merchants, group-buy offers and navigation and connects to ordering. The report examines conversational AI in local services; actual pages determine feature availability.

Why a Gemini agent shed specialist tools: trade-offs in Bash and media viewing (opens in a new tab)
Philipp Schmid describes an agent using Bash and a media viewer. Across over 350 tasks, harness scores were close while tokens and costs differed. Each task ran once, so score differences still require repeated experiments.



