hot events · 2026-05-08

This is Ai2's tech report published on the Hugging Face blog. Both sources covering it are pulling from the same official post, so there's no independent third-party take yet. EMO tackles a known MoE problem: in theory, each token only activates a few experts, but in practice, a single task ends up firing nearly all of them because experts specialize in low-level patterns like punctuation rather than high-level domains like math or code. EMO's approach is to let modular structure emerge during pretraining without relying on human-labeled domain categories. I'd take the "12.5% of experts" claim with a grain of salt for now. The paper compares EMO against a standard MoE with the same architecture, and the standard one degrades badly when you only use a subset of experts—EMO degrades less. But the blog post shows trend charts, not specific benchmark numbers. What's missing: exact performance drops per task, whether 12.5% is the sweet spot, and whether this modularity holds at larger scales.

Two outlets report RadixArk’s $100M seed, both anchored on the SGLang team. Their angles split between “open AI infrastructure” and the unusual Nvidia-AMD-Intel investor lineup. The available body is only a WeChat verification page, so valuation, lead investor, product scope, and shipping timeline are not disclosed. I don’t buy the “next-generation infra” label on its own. The stronger signal is that SGLang already has developer credibility in inference serving, KV cache work, and agent workloads. That puts RadixArk in the same pressure zone as vLLM, TensorRT-LLM, and Triton. If all three chip vendors are actually on the cap table, the bar is brutal: this cannot stay a framework story; it has to show reproducible cross-GPU performance wins.

OpenAI launched openai-cli, so you can now call GPT models straight from the terminal without installing the Python SDK or writing curl commands. Two sources covered this, but the WeChat article from jiqizhixin is behind a CAPTCHA wall — we only have the headline. The other source, aihot, has a similar headline, which suggests both are working off the same official announcement or GitHub release. I'd take this with a grain of salt for now. We don't know which models are supported, how billing works, or whether there's rate limit control. If it's just a thin wrapper around the API, it's genuinely useful for quick prototyping in the terminal, but you'd still want the SDK for production. Anthropic and Google both shipped CLI tools earlier, so this feels more like OpenAI catching up than breaking new ground. Still missing: the GitHub repo isn't linked in the coverage we have, so no visibility into stars, issues, or community reaction. Also unclear whether it matches the existing Python/Node SDKs feature-for-feature. Wait for the official docs before judging how good this actually is.

Both sources put Anthropic near a $1T valuation, but the chain appears to rest on the FT headline; the accessible body gives no revenue number, terms, or investor names. AIhot pushes “tens of billions this summer” and an OpenAI-flip angle, while FT’s visible framing is narrower: surging revenue and a deal being weighed. I don’t buy the excitement around “overtaking OpenAI.” Claude has real pull with developers, especially around Sonnet, coding workflows, agents, and the safety-heavy enterprise pitch. But a $1T mark demands repeatable, high-margin revenue, not just API usage spikes. OpenAI still has ChatGPT subscriptions and consumer distribution. Anthropic has to prove the enterprise contract base can carry the valuation.