AI Daily — 2026-03-26

English 中文

Meta unveils TRIBE v2: Trimodal Brain Encoder for neural prediction · Gemini 3.1 Flash Live Enabl...

Covering 36 AI news items

🔥 Top Stories

1. Meta unveils TRIBE v2: Trimodal Brain Encoder for neural prediction

Meta introduces TRIBE v2, a foundation model that predicts human brain responses to sight and sound. Built on 500+ hours of fMRI data from 700+ participants and drawing on Algonauts 2025, it can create a digital twin of neural activity and enable zero-shot predictions across new subjects, languages, and tasks. A public demo is available for exploration, signaling potential advances in neuroscience research and neuroadaptive AI, while raising considerations around data privacy and consent. Source-x

2. Gemini 3.1 Flash Live Enables Real-Time Voice and Vision Agents

Gemini 3.1 Flash Live is presented as a real-time model for building voice and vision agents, reflecting more than a year of work on model, infrastructure, and user experience. The release claims a step-change in quality, reliability, and latency, with potential implications for conversational AI, robotics, and on-device assistive tech. This could accelerate practical deployments of embodied AI across industries. Source-x

3. TurboQuant Redefines AI Efficiency with Extreme Compression

Google Research debuts TurboQuant, a compression approach aimed at drastically reducing AI model size and compute needs while preserving accuracy. The technique targets faster inference and lower hardware requirements for large AI workloads, signaling a major efficiency advance that could reshape deployment at scale. Source-rss

📰 Featured

Open Source & Tools

Mistral Voxtral TTS: Open-Weight, Multilingual, Ultra-Fast — Voxtral TTS is an open-weight text-to-speech model designed for natural, expressive speech across 9 languages with ultra-low latency and easy adaptation to new voices; could lower barriers to on-device TTS, though licensing and deployment considerations remain. Source-x
Chroma Context-1: 20B Search Agent, Faster and Cheaper — A 20B-parameter search agent offering order-of-magnitude speed and cost improvements while remaining open-source under Apache 2.0, with added features like HLS playback; a notable step for accessible, capable agent tooling. Source-x

AI Agents & Multi-Agent Systems

OpenAI backs Isara, coordinating 2,000 AI agents — OpenAI-backed Isara aims to coordinate thousands of AI agents to tackle complex problems (e.g., market forecasting); raised $94M at a $650M valuation, signaling emphasis on coordinated multi-agent tooling for finance. Source-x

AI Tools & Integrations

Claude Code Auto-Fix in the Cloud for PRs — Auto-fix runs in the cloud, with PRs automatically following Web and Mobile sessions to fix CI failures and address review comments, keeping PRs green remotely. Source-x
Codex plugins roll out; works with Slack, Figma, Notion, Gmail — Codex now interoperates with popular tools, streamlining developers’ workflows and deepening Codex integration with everyday tooling. Source-x

AI Safety

Strix Open-Source AI Hackers Scan and Fix Vulnerabilities — Strix deploys autonomous hacker agents that run code, discover vulnerabilities, and validate them with PoCs, integrating with GitHub Actions to scan on PRs and auto-fix insecure code before production. Source-github

Hardware & Inference

Qwen 3.5 27B Hits 1.1M Tokens/s on B200 GPUs — The 27B dense model achieved 1,103,941 tokens/s on 12 nodes with 96 B200 GPUs using vLLM, aided by several optimizations to maintain high efficiency without custom kernels. Source-reddit

⚡ Quick Bites

RotorQuant: Clifford Rotors Boost TurboQuant 10–19x — A rotor-based approach claims 10–19x speedups for TurboQuant workloads. Source-reddit
NVIDIA gpt-oss-puzzle-88B Boosts Inference on H100 — An OSS model puzzle purportedly improves H100 inference efficiency. Source-reddit
Home-trained LTX 2.3 LoRA of George Costanza on 5090 in a day — Quick-turnaround home training for a named persona using LoRA. Source-x
ComfyUI Debuts Dynamic VRAM Optimization for Local Models — Dynamic VRAM management improves local-model usability. Source-x
I built an MCP to flag known bugs before Claude’s recommendations — A custom machine-checking protocol aims to pre-filter issues before Claude’s outputs. Source-reddit
CUA-Suite Boosts Data with Massive Video Demonstrations for CUAs — Expanded video demonstrations for conversational AI agents to improve evaluation. Source-huggingface
DA-Flow: Degradation-Aware Optical Flow with Diffusion Models — Introduces diffusion-based degradation-aware optical flow. Source-huggingface
AI Users Wrecked by Delusion: Real-Life Impacts — The Guardian reports on real-world harms stemming from AI misinformation. Source-rss
BerriAI litellm: Access 100+ LLMs via Python SDK — A Python SDK to access a broad suite of LLMs. Source-github
Robust LLM Extractor for Websites in TypeScript — A tool for robustly extracting LLM results for web content. Source-github
Plain-text cognitive architecture for Claude Code — A cognitive architecture approach for Claude Code using plain text. Source-rss
Optio: Orchestrate AI coding agents in Kubernetes from ticket to PR — Orchestrates AI coding agents within Kubernetes for end-to-end workflow. Source-github
Ensu: Ente Launches Local LLM App — Local LLM app by Ente. Source-rss
Disney exits OpenAI deal as Sora is shuttered — Report on OpenAI deal changes in the media/entertainment space. Source-rss
Cohere Unveils 2B Open-Source Transcription Model — Cohere releases a 2B open-source transcription model. Source-reddit
Mistralai Voxtral-4B-TTS-2603 Released on Hugging Face — Voxtral-4B-TTS released for community use. Source-reddit
Claude session limits adjusted during peak demand hours — Claude session quotas adapt to demand cycles. Source-x
AI Hardware Spotlight: CPU, GPU, TPU, NPU, LPU Compared Visually — A visual guide comparing AI accelerator types. Source-x
We Rewrote JSONata with AI in a Day, Saved $500K/Year — Case study on AI-assisted JSONata rewrite. Source-rss
NYC Hospitals drop Palantir as UK expansion continues — Palantir usage in NYC hospitals declines amid UK expansion. Source-rss
Claude Subconscious Extends Claude Code with Memory and Guidance — ClaudeCode gains memory and guidance enhancements in a new patch. Source-github
Most Claude Outputs Land in Low-Star GitHub Repos — Analysis shows Claude outputs often appear in less-rated repos. Source-rss
Tips: Run llama-server with -np 1 for single-user efficiency — Practical tip for single-user performance of Llama Server. Source-reddit
AI Art in Project Hail Mary, Claude Usage Stats Revealed — Claude usage statistics tied to AI art generation in a media project. Source-reddit
Claude Becomes Third Top Contributor on OpenAI’s Repo — Claude attains third place in top contributors on OpenAI’s repo. Source-x
Trying to prove I’m not AI: Aunt remains unconvinced — Real-world challenges in distinguishing human from AI in social tests. Source-rss

Generated by AI News Agent | 2026-03-26