SpaceX-owned Cursor told the House Homeland Security and China Select Committees it will explore phasing out People's Republic of China-developed open-weight models over 'geopolitical risk,' CNBC reported Friday. Cursor had built its Composer 2 coding model on Moonshot AI's Kimi K2.5.
AI news for Saturday, November 21, 2026
The Daily AI Espresso — the links the most-followed people in AI actually shared, curated every morning. Edited by Alexis · Live updates →
Six developments: one new today and five from yesterday's alerts.
Pick the topics, companies and people you track. We’ll follow what leading AI experts are reading and sharing, filter the noise, and send your focused edition weekly—or sooner when enough important news breaks (up to three times per week).
Choose my signals →The rest of today’s expert-filtered signal, ranked for consequence, utility, surprise, and range.
Thinking Machines Lab released Inkling 2 Mini on November 20, a 26.4B-parameter MoE model (3.4B active) whose middle layers loop three times per token. The 20GB NVFP4/MXFP8 checkpoint runs on a laptop and leads open-weights models under 100B on DeepSWE v1.1 at 48.2%.
Stanford's Ravi Vakil wrote Friday that mathematician Alessandro Maria Masullo drafted a proposed proof of the Generalized Riemann Hypothesis titled 'The Music of the Primes,' entirely human, no AI. Brian Conrad reviewed it for months and found it 'so far looking good'; Vakil later said problems surfaced from AI-assisted checks.
Amazon Threat Intelligence said November 20 that a suspected Chinese-nexus actor used the open-source Cairn agentic framework to autonomously find and exploit a zero-day in a medical-imaging application, instructing the agents only to 'obtain shell access.' The attackers targeted East Asian universities between September 8 and 10.
NYU's Tristan Buckmaster and Anthropic's Levent Alpoge posted a new AI-aided proof November 20 claiming finite-time blowup for the 3D Navier-Stokes equations with no external force, which many mathematicians consider the real version. They say it closes a controversial loophole in OpenAI's September solution.
President Trump said at a White House press conference Friday that US companies will be banned from building data centers abroad and will face a 'tremendous tax' if they do, CNBC reported. The announcement surprised his own staff and lands as Alphabet, Amazon, Microsoft, Meta and Oracle already operate large overseas footprints.
Tighter US chip export controls that took effect Thursday are disrupting the semiconductor supply chain more broadly than expected, reaching mobile and consumer parts and hitting developers outside China, Nikkei Asia reported. TSMC has paused certain shipments, with executives saying the impact is wider than first anticipated.
Useful releases, methods, and workflows worth trying.
Kyutai released AURA on November 20, an audio-driven video model that generates 576p talking-head video chunk by chunk at 20.7 fps on a single GPU in the 1.3B version, and 720p at 9.7 fps in the 5B version. Only the current audio chunk is needed to start generating, enabling low-latency voice avatars.
OpenAI made its GPT-Live-1.1 voice model available in the API on November 20 and introduced a standby mode that bills idle time at $0.02 per minute versus $0.05 per minute for active inference, the company said on its developer community forum. Developers can set idle timeouts from 5 to 60 seconds on both GPT-Live-1 and 1.1.
That’s today’s shot. — Alexis · AI Weekly · Sponsor this newsletter
