AI FRONTIER
Signals worth understanding before they become obvious.
Mage-VL-Cowork: A Full Training Recipe to Learn Hybrid Computer Use From Evolved Harnesses
The paper presents Mage-VL-Cowork, a 4B hybrid computer-use agent trained from a general multimodally aligned checkpoint through an end-to-end, progressive five-stage curriculum across midtraining, SFT, and RL. Its Mage-CUA-Forge pipeline synthesizes 293,806 desktop and web tasks across 6,915 skills using reusable, ski...
- Mage-VL-Cowork is a 4B hybrid CUA trained from a general multimodally aligned checkpoint.
- Training uses an end-to-end, progressive five-stage curriculum across midtraining, SFT, and RL.
- Mage-CUA-Forge synthesizes 293,806 desktop and web tasks across 6,915 skills.
- The paper reports Mage-VL-Cowork-4B reaches 66.0% on OSWorld-Verified and states it achieves state-of-the-art performance at its parameter size.
For teams building on-device agents, the recipe shows that reusable skill-conditioned task synthesis plus an evolved teacher harness can unify GUI interaction with code and terminal execution in a single 4B policy.
mistralai/Mistral-Large-4-1T-A52B: 1T-Parameter Natively Multimodal MoE Released
Mistral released Mistral Large 4, a 1-trillion-parameter natively multimodal Mixture-of-Experts model with 52 billion active parameters per token, a 1M context length, and support for 160+ languages under the Mistral Ecosystem Open-Weight License.
- ~1T total / ~52B active parameters, with 128 routed experts (4 active per token) and 1 shared expert
- Supports a 1M context length and natively multimodal text and image input with a Pixtral-style vision encoder
- Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own datacenters in Europe
- Reports 48% on the Artificial Analysis Coding Agent Index, effectively tied with leading open-weight model GLM-5.3 while using 47% fewer tokens per task on average
For builders of multimodal and agentic workflows, this model offers a 1M context and native function calling in open weights, but the 1T parameter scale requires multi-GPU tensor-parallel deployment and MaaS commercial use needs a separate ...
AutoHebb: Self-Modifying Neural Networks that Zero-Shot Infer and Continually Learn
The paper introduces autonomous Hebbian networks (AutoHebb), a class of online learning feedforward neural networks that combine external feedback with internally generated activity to produce a global learning signal (neuromodulator), which gates a three-factor Hebbian local learning rule that updates plastic weights....
- The paper reports that AutoHebb networks succeed on zero-shot inference (ZI) tasks.
- The paper reports emergent, untrained ZI and continual learning (CL) abilities.
- The paper reports that AutoHebb admits reverse-engineering into explicit algorithms.
For AI builders, searching fixed weights by meta-learning while updating plastic weights through a three-factor Hebbian rule offers an analyzable alternative path to online continual learning and zero-shot inference.
DyCraft: An Agentic System for Long-Horizon 4D Scene Generation
DyCraft decomposes animation generation into steps: it first expands the user's prompt into a detailed plan and builds a static scene, then creates a 3D storyboard of key animation poses so a coding agent can author local motion between verified start and end states.
- The paper evaluates DyCraft on 24 long-horizon, multi-shot animation tasks covering humanoid robots, human and animal characters, and physical mechanisms.
- Compared with a native GPT-6 agent in Codex and prior agentic baselines, DyCraft substantially reduces physical violations and generates higher-quality animated 3D scenes that are more often preferred by human raters.
- The paper reports that frontier coding agents equipped with general-purpose harnesses frequently produce objects that interpenetrate, lack physical support, or move unnaturally.
- DyCraft uses 3D storyboarding to decompose the problem into pairs of verified start and end states, paired with a collection of motion authoring tools.
For builders of 3D and animation pipelines, DyCraft indicates that splitting long-horizon generation into independently verifiable start and end states curbs physical violations better than having an agent script the full animation in one p...
OpenAI introduces LifeVal, a benchmark for AI on everyday tasks
OpenAI introduces LifeVal, a benchmark measuring how well AI helps people complete everyday tasks, from finding childcare to managing finances and government paperwork.
- LifeVal is a benchmark introduced by OpenAI.
- It measures how well AI helps people complete everyday tasks.
- Example tasks include finding childcare, managing finances, and government paperwork.
For AI builders, LifeVal shifts evaluation toward real-life tasks such as childcare, finances, and government paperwork, signaling where models need targeted improvement.
Qodo builds context into AI code review with Codex
Qodo brings company knowledge and coding standards into Codex, helping developers catch P0 issues across repositories before opening a pull request.
- Qodo brings company knowledge and coding standards into Codex.
- The integration helps developers catch P0 issues across repositories before opening a pull request.
For AI builders, injecting organization-specific context into code review agents shifts review from generic checks toward catching high-severity issues before the PR.
66-year-old Yann LeCun's new paper challenges Scaling Law: 109M parameters beat 72B
Yann LeCun published a new paper reporting that a 109M-parameter model outperformed a 72B-parameter model, challenging the Scaling Law assumption that larger models necessarily understand the world better.
- The paper reports that a 109M-parameter model outperformed a 72B-parameter model.
- The paper was published by 66-year-old Yann LeCun.
- The paper challenges Scaling Law, arguing that larger models do not necessarily understand the world better.
For AI builders, this paper suggests parameter scale is not the sole determinant of world understanding, and architecture and training objectives may matter more.