Live Feed
Meta Deploys MTIA 400: 6 Petaflops Custom Silicon Targets Nvidia for GenAI Inference
Meta has completed testing on the MTIA 400 and is deploying it across data centers now, targeting generative AI inference workloads — image generation, video synthesis, text response — that previously ran on Nvidia GPUs. At 6 petaflops FP8 and 288GB HBM, it's the first custom chip to directly challenge Nvidia on frontier model serving, not just recommendation systems.
China Makes Pre-Development AI Ethics Approval Mandatory, Effective Immediately
Ten Chinese government agencies jointly issued binding rules on April 3 requiring every company, university, and research institution engaged in AI development to establish internal ethics committees and obtain clearance before work begins. High-risk projects face mandatory government-led expert review under a quasi-administrative approval process.
Cursor 3 Rebuilds the IDE from Scratch Around Agents — Composer 2 Beats Claude Opus 4.6 on Terminal-Bench
Cursor shipped a ground-up rebuild of its editor on April 2 with a dedicated Agents Window for parallel agent orchestration across local and cloud environments. Its proprietary Composer 2 model scores 61.7 on Terminal-Bench 2.0, ahead of Claude Opus 4.6 (58.0) at 86% lower token cost than its predecessor.
Google Gemma 4 Ships Four Open Models Under Apache 2.0 — 31B Hits Arena ELO 1452
Google DeepMind releases Gemma 4 on April 2 with four open-weight models (E2B, E4B, 26B MoE, 31B Dense) under Apache 2.0. The 31B variant scores 1452 on Chatbot Arena — third among all open models globally — and 89.2% on AIME 2026 math, up from 20.8% in Gemma 3.
Microsoft Launches Three In-House AI Models — Breaking Its OpenAI Dependency
MAI-Transcribe-1 beats Whisper on all 25 tested languages; MAI-Voice-1 generates 60 seconds of audio per second; MAI-Image-2 debuted at #3 on the Arena.ai image leaderboard. All three are live on Microsoft Foundry as of April 2, 2026.
Arcee AI Ships Trinity Large Thinking Under Apache 2.0 — 400B Sparse MoE at $0.85/M Output
Arcee AI has released Trinity Large Thinking, a 400B sparse mixture-of-experts reasoning model with only 13B active weights per token, under Apache 2.0. At $0.85 per million output tokens, it posts 94.7% on Tau2-Bench Telecom and undercuts Claude Opus 4.6 by 98.9% per output token.
OpenAI Kills Sora at $1M/Day Loss — Google Fills the Gap with Veo 3.1 Lite at $0.05/Second
OpenAI shuts down Sora on April 26, cancelling its $1B Disney deal. The same week, Google launches Veo 3.1 Lite at half the price of its previous cheapest tier — a direct play for the developer market OpenAI is abandoning.