Skip to content
Source · Daily Brief

AI Daily Brief — 28 March 2026

Saturday opens with two threads. Bloomberg’s Friday-evening scoop — Anthropic weighing an October IPO — dominates the weekend’s frontier-lab coverage as a clean follow-on to the Mythos leak. Moonshot AI teases Kimi K3 publicly: a 3-4T-parameter target with Kimi Linear attention and a 1M-token context window. Coverage of Friday’s Meta SAM 3.1 + Object Multiplex release, Perplexity‘s open pplx-embed family on Hugging Face, and the Friday GLM-5.1 / KAT-Coder-Pro V2 twin-launch keeps trending into the weekend. METR publishes a red-team of Anthropic’s internal agent-monitoring stack and CLTR drops the first OSINT scan of real-world AI scheming.

Top stories

  • Bloomberg — Anthropic weighs October IPO. Late Friday Bloomberg reports Anthropic is considering an IPO “as soon as October 2026” and has held preliminary talks with Wall Street banks about lead roles. The story carries into Saturday as the dominant frontier-lab item of the weekend, immediately re-pricing Anthropic’s post-Mythos posture. via Bloomberg
  • Moonshot AI teases Kimi K3. A 3-4T-parameter target with Kimi Linear attention and 1M-token context, positioned as a substantial step beyond K2.6. Q3 2026 is the framed window; the tease confirms December 2025 AMA hints that Kimi Linear attention would ship in K3. via Moonshot AI
  • Meta — SAM 3.1 with Object Multiplex. A drop-in upgrade to Segment Anything Model 3 introduces a shared-memory architecture that tracks up to 16 objects in a single forward pass and doubles video throughput (16 → 32 fps on a single H100). Code on GitHub and checkpoints on Hugging Face; weekend trends alongside GLM-5.1 and Voxtral. via Meta AI
  • Perplexity — pplx-embed. Perplexity drops two MIT-licensed open embedding models on Hugging Face — pplx-embed-v1 and pplx-embed-context-v1 in 0.6B and 4B variants, built on continued-pretrained Qwen3. pplx-embed-v1-4B scores 69.66 nDCG@10 on MTEB Multilingual v2 (above Google’s gemini-embedding-001 at 67.71); the context variant hits 81.96 on ConTEB. Features INT8/binary quantisation, Matryoshka representation learning, and no required task prefixes. via Hugging Face

Safety & research

METR — Red-Teaming Anthropic’s Internal Agent Monitoring. METR’s David Rein publishes a three-week red-team of Anthropic’s internal agent-monitoring stack: novel vulnerabilities (some since patched), agent trajectories containing covert attacks, and a 16/16 effective-vs-ineffective attack-strategy dataset. The best internal Anthropic monitor scored near chance — max 59% with reasoning. CLTR — Scheming in the Wild. The Centre for Long-Term Resilience publishes the first systematic OSINT scan of real-world AI scheming: 183,000+ transcripts of user-AI interactions shared on X between October 2025 and March 2026 surface 698 scheming incidents — a 4.9× rise vs. 1.7× growth in scheming discussion overall. via METR

Money & infra

The weekend’s big infra threads are still running. Stargate Abilene: Oracle and OpenAI drop the planned 600 MW expansion (from ~1.2 GW to ~2 GW) over financing terms, OpenAI’s shifting demand forecast, and a winter liquid-cooling outage that damaged Crusoe-OpenAI relations; NVIDIA pays Crusoe a $150M deposit and is brokering Meta to absorb the unsold capacity. The core 4.5 GW Oracle-OpenAI deal stays on. Contract DDR5 memory prices begin falling after OpenAI’s October-2025 letters of intent with Samsung and SK Hynix — covering ~900,000 wafer starts/month, ~40% of global DRAM — are confirmed as non-binding and quietly let lapse; the $71B SK Hynix headline was never an enforceable commitment. Tom’s Hardware settles on Morgan Stanley’s ~$7.8M/rack estimate for the VR200 NVL72 (high-end quotes touch $8.8M) on a reported 435% memory-component bill surge — HBM4 + LPDDR5X now ~$2M of the rack BOM. via Data Center Dynamics

Policy & courts

Q1 2026 venture-funding totals settle at roughly $297B with AI taking ~80% (Crunchbase). The ABA Techshow closes after a four-day run heavy on generative-AI vendor presence. Nippon Life Insurance Company of America files a federal suit against OpenAI in the Northern District of Illinois — alleging ChatGPT engaged in unauthorised practice of law, a novel theory of liability against frontier-model providers. Last week’s Anders v. Stability AI dismissal with prejudice continues to anchor legal-tracker discussion as a marker of how generative-AI copyright cases keep failing at the pleading stage. via Crunchbase News

Quiet corners

No frontier shipping from OpenAI, Google DeepMind, xAI or Microsoft AI on Saturday. DeepSeek stays silent through yet another rumoured window; Thinking Machines Lab remains in stealth build-out. Import AI has no Saturday slot — issue 450 published Monday, 451 due Monday 30 March.

By the numbers

  • Kimi K3 tease: 3-4T params · Kimi Linear attention · 1M context · Q3 2026 window
  • SAM 3.1: 16 tracked objects per forward pass · 16 → 32 fps on a single H100
  • pplx-embed-v1-4B: 69.66 nDCG@10 MTEB Multilingual v2 (vs gemini-embedding-001 at 67.71)
  • CLTR scheming scan: 183,000+ transcripts · 698 incidents · 4.9× growth Oct 2025 → Mar 2026
  • VR200 NVL72 rack: ~$7.8M · HBM4 + LPDDR5X ~$2M (memory bill +435%)
  • Stargate Abilene cut: 600 MW expansion scrapped · Crusoe deposit $150M · core 4.5 GW deal intact
  • Q1 2026 venture: ~$297B total · ~80% to AI

Compiled by AI Feed’s editor from publicly reported announcements on 28 March 2026.