A platform for researchers to publish and debate how increasingly capable AI should be built, governed and used, with first essays covering reasoning transparency, economic policy and human flourishing.
The round was led by Vista Equity Partners, Cambium Capital and Emergence Capital, valuing the Trinity-family open-weight model developer at roughly $1 billion.
Koa is fine-tuned from the open-weight Nemotron-3-Super-120B, converting Agentforce workflow descriptions into multi-turn RL training data via GRPO.
A comparison-oracle method that extracts directional information from preference pairs without optimizing a differentiable preference loss, with a convergence guarantee for its offline scheme.
Introduces the PanoCaps dataset and the panoptic grounded captioning task, requiring VLMs to describe both foreground objects and background regions with pixel-level grounding.
Converts scientific code into programmable environments for agents, guided by expert-defined cases and acceptance criteria to ease the 'scientific experience bottleneck'.
HOPE is a second-order expert-pruning objective that provably minimizes an upper bound on pruning error, beating REAP across three frontier MoE models.
A deployed MoE model learns from current user-provided facts and corrections instead of relying only on static pretraining, by generating weights from live interaction data.
A report on China's push to field humanoid robots for defense, drawing on the 2026 World Humanoid Robot Games, highlights growing military interest in embodied AI.
AMD detailed its AI compute roadmap spanning gigawatt-scale data center factories down to on-device agentic PCs, highlighting the Ryzen AI Max PRO platform with up to 192GB of unified memory.
Speaking at TechCrunch Disrupt 2026, Karpas argued physical AI still lacks a ChatGPT-style breakthrough, positioning Nvidia's robotics stack as the compute layer waiting for that inflection.
MLCommons published datacenter results, with Nebius reporting its GB300 NVL72 server performance on the DeepSeek R1 closed inference benchmark.
A single conversation that blends voice, text and visuals for customer support, letting customers switch mediums mid-conversation instead of using separate channels.
An open-source e-ink picture frame that listens for birdsong and draws each bird as a stylized 1800s-engraving illustration; it topped Hacker News with over 1,600 points.
An entire app (UI, schema and local SQLite data) packed into one portable .capsule file that runs offline, and that can be generated or updated from AI prompts.
An MIT-licensed, Python-native engine for model-driven synthetic test data and PII-aware pseudonymization, with MCP/IDE integration for regulated environments.
Led by Spark Capital, the round backs Rune's RELIC modular compute deployed at renewable sites, pairing AI compute with stranded energy.
Led by Eclipse, with Craft Ventures and Westly Group, and angels including Meta CTO Andrew Bosworth, backing the ex-Meta genAI VP's physical AI world-model startup.
Led by Xora Innovation (Temasek) and founded by ex-Infosys CEO Vishal Sikka, the extension brings the AI strategy firm's total to $85M.
Free means $0 per token today, not unlimited or unconditional: both platforms rate-limit free use and require a free API key, some OpenRouter free endpoints may train on your prompts unless you opt out in account settings, and the roster changes daily. Verified against each provider's own pricing data at generation time — but pricing changes without notice, so confirm it's still free on the provider's own page before you rely on it.
Source:
