Newsletter Archive

NVIDIA's Nemotron-3-Super-120B arrives with 120B params (12B active), 1M context window, and LatentMoE architecture—alongside fresh releases from Sarvam, Cohere, and others spanning OCR, transcription, code-editing, and math theorem proving. A month of niche, application-focused models rather than headline-grabbing frontier models.

Interconnects · 5 min read Repos

South Korean AI chip startup Rebellions raised $400M to compete with Nvidia's dominance, backed by local institutional investors. The funding signals a broader regional push to reduce dependence on U.S. semiconductor suppliers for AI infrastructure.

SiliconANGLE · 6 min read Hardware

ScaleOps closed a $130M funding round to automate real-time infrastructure management, targeting GPU scarcity and ballooning AI cloud costs. The platform dynamically optimizes resource allocation to reduce spending.

TechCrunch · 5 min read Industry

A California judge temporarily blocked the Pentagon from designating Anthropic as a supply chain risk and ordering agencies to stop using Claude, ruling the government sidestepped proper contract dispute procedures and used inflammatory social media to fuel a "culture war" rather than follow established law.

MIT Tech Review · 5 min read Policy

Mistral launched Voxtral, a 4B open-weights TTS model achieving 68.4% win rate vs ElevenLabs Flash v2.5, combining autoregressive semantic tokens with flow-matching for acoustic generation—a technique borrowed from image generation. The episode covers architecture, real-time voice agent deployment, and enterprise applications.

Latent Space · 50 min read Tools

David Sacks steps back from his White House AI czar role but retains major influence as co-chair of Trump's science council—a move that removes ethics constraints while keeping a powerful Silicon Valley voice steering federal AI policy as GOP skepticism grows.

Axios · 3 min read Policy

Microsoft released Harrier-OSS-v1, a family of three open-source multilingual embedding models (270M, 600M, 27B parameters) that hit state-of-the-art on Multilingual MTEB v2. Unlike traditional BERT-style encoders, they use decoder-only architectures with last-token pooling, enabling developers to swap in modern LLM-based embeddings across 100+ languages.

MarkTechPost · 4 min read Tools

EU AI Act Article 50 II requires dual transparency labeling (human + machine-readable) for AI-generated content by August 2026, but researchers show this collides with fundamental constraints of current LLMs: watermarks learned as spurious features during training, non-deterministic outputs break provenance tracking, and no cross-platform marking standards exist. Compliance cannot be solved via post-hoc labels—it demands architectural redesign.

arXiv AI · 3 min (abstract) Policy