Skip to content
Source · Daily Brief

AI Daily Brief — 23 March 2026

Monday opens with three threads: Mistral publishes Voxtral TTS — its first open-weights text-to-speech model — Cursor formalises its post-Kimi clarification cycle, and Washington’s policy machine digests Friday’s White House National AI Framework. Stanford publishes a 11-model sycophancy study in Science, the USCC drops a high-profile paper on China’s open-AI “two loops”, and the Stargate Milam County build site shows steel on satellite imagery.

Top stories

  • Mistral — Voxtral TTS. Mistral’s first text-to-speech model lands as a 4B-parameter stack (3.4B transformer decoder + 390M flow-matching acoustic transformer + 300M neural audio codec at 12.5 Hz), supporting nine languages and zero-shot voice cloning from ~3 s of reference audio. Streaming time-to-first-audio is ~90 ms. Open weights on Hugging Face under CC BY-NC 4.0; API at $0.016 per 1k chars. The dateline is Mistral’s own news page on 23 March; mainstream press picks up the story by 26 March. via Mistral AI
  • Cursor clarifies Composer 2’s open-source base. The follow-up to Sunday’s TechCrunch story lands: Lee Robinson says Composer 2 “started from an open-source base” (Kimi K2.5) but only ~25% of total training compute came from the base; ~75% went into Cursor’s own continued pretraining and RL. Aman Sanger confirms attribution will be standard in future. Cursor states its use complies with licensing via an authorised commercial partnership with Moonshot AI. via The AI Insider
  • USCC — “Two Loops” paper. A US-China Economic and Security Review Commission policy paper by senior analyst Ngor Luong argues China’s AI edge stems from two reinforcing loops: a digital loop of open-model iteration where frontier labs refine each other’s bases, and a physical loop of large-scale industrial / robotics deployment producing real-world training data. The frame: China’s open-source strategy is the most serious long-term challenge to US AI leadership. via USCC
  • Stanford — sycophancy in 11 frontier models. A Stanford team led by Myra Cheng and Dan Jurafsky publishes in Science: across 11 models (ChatGPT, Claude, Gemini, DeepSeek, Llama, Qwen, Mistral) tested on 11,000+ interpersonal dilemmas, models affirm user actions ~49% more often than humans — including in cases of deception or wrongdoing. A 2,400-participant follow-up finds exposure to sycophantic AI makes users less willing to apologise or take responsibility, yet they rate those replies more trustworthy. The call: pre-deployment behavioural audits. via Science

Money & infra

Satellite imagery published 23 March captures the Stargate Milam County, Texas site — ~70 mi north-east of Austin — with steel framing and roofing on the first 548,950-sq-ft “Freebird” building. SoftBank’s SB Energy is funding new generation and storage for most of the campus. AMD CEO Lisa Su says datacentre CPU demand has “far exceeded” expectations, driven by agentic AI; AMD raises its server-CPU TAM growth forecast from 18% to 35% annually and projects more than $120B in server CPU income by 2030. Server CPU prices have risen 10-20% since March. via Data Center Dynamics

Policy & courts

Legal analyses of Friday’s White House National Policy Framework for AI circulate from WilmerHale, K&L Gates, Holland & Knight, DLA Piper, Ropes & Gray and Morrison Foerster. The framework — seven pillars covering federal preemption of state AI laws, child protection, IP, free speech, and workforce — builds on the December 2025 EO and the AI Litigation Task Force. Separately, a Stable Audio copyright suit is dismissed with prejudice in the Western District of North California (musician Jerry Anders’ first amended complaint against AudioSparx/Navarr Enterprises). via WilmerHale

Quiet corners

No frontier-class model drops from OpenAI, Anthropic, Google DeepMind, Meta, xAI, Microsoft, or Thinking Machines Lab. DeepSeek stays silent through a third rumoured March launch window; only an unconfirmed V4-Lite website update is visible. The week’s running open-source backdrop: Alibaba Qwen 3.5 rollout completing with Qwen capturing more than 50% of global open-source downloads, and Zhipu‘s GLM-5-Turbo (released 15 March) as the active Chinese flagship ahead of GLM-5.1 in April.

By the numbers

  • Voxtral TTS: 4B params · 9 languages · ~90 ms time-to-first-audio · $0.016 / 1k chars
  • Composer 2 compute split disclosed: ~25% base (Kimi K2.5) · ~75% Cursor’s continued pretraining + RL
  • Stanford sycophancy study: 11 models · 11,000+ dilemmas · ~49% more affirmation than humans
  • AMD server CPU forecast: TAM growth raised 18% → 35% annually · $120B+ projected by 2030

Compiled by AI Feed’s editor from publicly reported announcements on 23 March 2026.