Shaun Johnson, Yann LeCun and Oriol Vinyals launch a $100M AUM technical and go-to-market-focused venture firm writing $1-5M checks in applications, robotics, infrastructure and core intelligence.
Ant Group's model ships with hybrid-linear attention and 1/64 sparsity, trained on 10,000+ interactive environments, with LMSYS shipping day-0 SGLang support that cuts TTFT by 60-80%+ on long agent runs.
A wireless foundation model pretrained to model multipath propagation for channel estimation, beam prediction and localization, showing strong generalization across wireless sensing tasks.
Shows that roughly 1M-parameter transformers can benefit from a simple Chain-of-Thought form, enabling detailed analysis of step-by-step reasoning at a scale far smaller than frontier models.
An adaptive agentic red-team that breaks Deep RL-based autonomous cyber defense systems, which are typically only evaluated against static heuristic attackers.
Analyzes how privileged-information-conditioned self-teachers provide dense per-token supervision, exposing failure modes of compute-efficient self-distillation as an alternative to verifiable-reward RL.
A training-free sparse attention method using hashing on binary principal components to accelerate long-context decoding without KV-cache reprocessing cost.
Moonshot AI's open-weights Kimi K3 is touted as reaching frontier-class multimodal and agentic performance, with Together AI leading or tying for #1 on OCRBench, MMMU Pro Vision and DeepSWE.
Highlighted by Together AI, DeepSeek's smaller V4 Flash outpaced its larger V4 Pro Preview (72.1) while using roughly one-fifth the total parameters, continuing the V4 line's extreme parameter efficiency on agentic and coding benchmarks.
A quantized NVFP4 build of DeepSeek V4 Flash 0731 adds MoonViT vision support, targeting lower-footprint multimodal inference for the V4 Flash line.
The tech distributor and robotics integrator bring Physical AI to industrial environments via quadruped robots with AMD-powered edge computing for real-time monitoring.
Shares rose after Elon Musk said SpaceX will exclusively buy Nvidia AI chips, reinforcing Nvidia's position as sole AI silicon supplier to another major customer.
Fueled by Autodesk, the Smart Industrialized Construction Lab will train students on construction processes using robots to address housing shortages and labor declines.
Sapiom ships a cost-aware model router, an Agent Studio, and a Runtime with typed step graphs and full traces, letting agents pay for search, scraping and model calls mid-run.
A generative UI framework that lets coding agents in Claude Code, Codex and Gemini CLI render interactive interfaces from natural language.
An open-source Model Context Protocol gateway for enterprises bundling built-in OAuth 2.0 and role-based access control to centralize and secure MCP tool access.
An MCP aggregator that consolidates a large set of Model Context Protocol tools into a single unified interface of two tools, reducing tool-discovery burden on agents.
The llm CLI and library gained reasoning traces, OpenAI Responses API support, server-side tools, improved logging and broader LLM compatibility.
Source:
