Founded by xAI co-founder Igor Babuschkin, River AI raised $1.1B across seed and Series A led by General Catalyst and AMP PBC, with Nvidia, AMD Ventures, Y Combinator and Temasek participating, to build an open full-stack AI platform for personally-owned agents.
Smart Analytics Global reports Chinese manufacturers accounted for over 97% of worldwide humanoid shipments in the first half, with AgiBot at 44% and Unitree at 31% of a 1.91-million-unit market up 274% year on year.
Lightcap, who spent nearly eight years at OpenAI, announced he is leaving to start something new, calling the decision bittersweet and planning to stay involved for a few weeks. Sam Altman responded warmly, saying they will work together on what's next.
Across 247,694 instruction lifetimes in 1,867 repos, agentic prompts grow +226% over their lifetime, with older instructions increasingly unlikely to be deleted. Comments that encode latent reasoning remove 99.3% of excess instructions and improve real-world instruction-following by up to 23.1%.
The first benchmark evaluating agents across full data-science workflows in real computer environments finds the strongest agent (Claude-4.6-Sonnet) reaches only 56.70% task success, while all open-source agents stay below 1%.
Generates financial reasoning benchmarks from executable computation trees grounded in real tables, with answers correct by construction. Accuracy falls up to 51% with reasoning depth, and lightweight LoRA fine-tuning lifts it from 81.1% to 85.6%.
A lightweight LoRA running alongside a frozen LLM restructures the residual stream to localize hallucinations in one pass, reaching 0.966 AUROC and >80% localization accuracy for tool-calling, acting as a negligible-latency guardrail in ReAct loops.
A 200-task benchmark across ten everyday-life domains, each a scripted multi-week timeline with silent changes requiring proactive re-inspection, finds all seven frontier models score low — current agents are far from real-life assistance.
An agentic dev environment with institutional memory that connects to your services, ownership, docs and architectural decisions, and can run Claude Code, Codex and other agents in one place.
Supports ChatGPT, ChatGPT Work and Codex with project and browser workflow integration, shipping as .deb or .rpm for x64 and ARM64 across Ubuntu 24.04/26.04 LTS, Debian 13 and Fedora 43/44.
A new coding agent launched on Product Hunt, claiming significant speedups over Claude Code and Codex on everyday tasks.
AI Agents now work on the canvas to design, write, analyze and organize sites, with the option to connect your own AI like Claude Code or Codex, plus branching for safer idea testing.
The automation platform launched 3B, a secure environment for running AI agents, apps and automations.
Source:
