Hello from The AI Night,
Today in AI:
Google Gemini Can Now Create AI Music Using Lyria-3
Meta and NVIDIA Sign Multi-Year AI Deal Spanning Data Centers and GPUs
OpenAI and Paradigm Launch EVMbench; An AI Benchmark for Smart Contract Security
Google DeepMind
Google Gemini Can Now Create AI Music Using Lyria-3
Image Source: Google blog
Here's the deal: Google DeepMind launched Lyria-3, a generative music model now in beta inside the Gemini app. Users can describe a track or upload a photo/video, and Gemini produces a 30 second song with lyrics and custom cover art.
The Breakdown:
Lyria-3 generates lyrics automatically from prompts, removing the need to write your own
Users get control over style, vocals and tempo with support for both lyrical and instrumental output
Available to all users 18+ across eight languages (English, German, Spanish, French, Hindi, Japanese, Korean, Portuguese), rolling out on desktop first, mobile in the coming days
AI Plus, Pro and Ultra subscribers get higher usage limits
All output is watermarked with SynthID. Gemini can now verify whether uploaded audio, images or video were generated by Google AI
Prompts naming specific artists are treated as broad style inspiration, not imitation. Filters check outputs against existing content
Lyria-3 also powers YouTube's Dream Track for Shorts creators, expanding beyond the U.S
The bigger picture: This puts music creation directly inside a general purpose AI assistant, lowering the barrier from "use a music tool" to "just ask Gemini." For short form creators especially, native soundtrack generation inside YouTube could change how Shorts are produced at scale
Image Source: Meta blog
Here's the deal: Meta and NVIDIA announced a multiyear strategic partnership covering on-premises, cloud, and AI infrastructure. Meta will build hyperscale data centers using millions of NVIDIA Blackwell and Rubin GPUs, deploy Arm based Grace CPUs at scale and integrate Spectrum-X Ethernet networking across its infrastructure.
The Breakdown:
This is the first large scale NVIDIA Grace only CPU deployment, focused on improving performance per watt across Meta's data centers
NVIDIA Vera CPUs are also in the pipeline, with potential large scale deployment in 2027
Meta will use GB300-based systems to create a unified architecture spanning on-premises and cloud partner deployments
NVIDIA Confidential Computing has been adopted for WhatsApp private processing, with plans to expand across Meta's portfolio
Spectrum-X Ethernet replaces or supplements existing networking to support AI-scale throughput and lower latency
The bigger picture: This signals Meta is doubling down on NVIDIA's full stack rather than relying solely on custom silicon. For the broader ecosystem, large scale Grace CPU adoption strengthens the Arm server software ecosystem and the Confidential Computing expansion could set new standards for privacy preserving AI at scale.
Image Source: OpenAI blog
Here's the deal: OpenAI and Paradigm Launch EVMbench, a benchmark that tests AI agents on detecting, patching and exploiting high severity smart contract vulnerabilities. It draws on 120 curated vulnerabilities from 40 audits, mostly from open code audit competitions plus scenarios from the Tempo blockchain.
The Breakdown:
Three capability modes: Detect (find vulnerabilities), Patch (fix them without breaking functionality), and Exploit (execute full fund draining attacks in a sandboxed environment)
GPT-5.3-Codex scores 72.2% on exploit tasks up from GPT-5's 31.9% six months prior
Detect and patch performance remains below full coverage, agents often stop after finding one issue instead of auditing exhaustively
A Rust based harness handles deployment, deterministic transaction replay and grading
OpenAI is committing $10M in API credits for cyber defense research, especially for open source and critical infrastructure
Tasks, tooling and evaluation framework are publicly released
The bigger picture: Smart contracts secure over $100B in crypto assets. As AI agents get better at code exploitation, benchmarks like EVMbench help track both offensive risk and defensive capability, pushing security teams to integrate AI assisted auditing before attackers do.
What else you need to know:
Cursor launched a plugin marketplace letting agents connect to external tools like Figma, Stripe, AWS and Linear through bundled MCP servers, skills, subagents, rules and hooks.
ElevenLabs became the first company to secure AI agent insurance through AIUC-1 certification after its voice agents passed 5,835 adversarial tests across 14 risk categories.
Enterprise AI search startup Nerve is shutting down its product in 30 days and its team is joining OpenAI to work on building search capabilities for ChatGPT at larger scale.
Mistral AI is acquiring serverless cloud platform Koyeb to strengthen its Mistral Compute offering, bringing Koyeb's team and infrastructure tech to support AI workloads across its expanding GPU capacity.
Microsoft pledged $50 billion by 2030 to bring AI infrastructure and internet access to lower income countries, noting AI adoption in developed nations is roughly double that of the global south.
That’s it for today’s edition of The AI Night.
Our goal is to cut through the noise, surface what actually changed, and explain why it matters.
If this was useful, you’ll get the same signal here tomorrow.