Cognitive Warmup. A big secret has been revealed. We finally know that AI music production company Suno has trained its models. One of the biggest in terms of what they do, here’s a sketch of the millions of songs and lyrics they found pulled from various sources. 113,879 hours of YouTube Music, 62,117 hours of Pond5 and 12,287 hours of Deezer, to name a few. It also features audio from the stock music libraries Pond5, Jamendo, Freesound, and the International Music Score Library Project, as well as podcasts via RSS feed.
Listen has faced several lawsuits from the recording industry in recent years, mostly related to using copyrighted data for training AI models. UMG, Sony, Warner Music Group and Production Music Library have all been (or still are) at odds with Listen on some level. Listen’s deepest argument so far is that it can be used for “essentially all music files of reasonable quality that are accessible on the open Internet”.
First, on Neural Dispatch
Security in an agentic future
I have charted Google’s intention to make AI relevant to the masses in India, over time and in detail. This includes real-world implementations in education, agriculture, health care, etc. This week, I’d like to go back to an announcement from I/O Connect India 2026… which talks specifically about India’s cyber security apparatus in the age of AI agents.
Reading: Google bets on India’s agentic AI future with new education, health push
“The agentic era increases that responsibility. Software can now interpret intent, use tools and take actions autonomously. This gives developers extraordinary speed, but also changes what must be secured. Security can no longer be the last checkpoint before launch. It must be part of the underlying architecture from day one, shaping how agents are built and how they interact,” Google India said in a statement.
For this, Google has started three major initiatives in India.
- Google says they are making their specialized cyber security agent, Sec-Gemini v3, available to trusted government and enterprise testers, including Flipkart. Sec-Gemini can reason over complex security data and help teams investigate incidents faster. Amidst high threat perception, this will be relevant to Indian enterprises and government service systems.
- For this, Google refers to the Big Sleep vulnerability research agent developed by Google DeepMind and Project Zero and says that it can be useful in identifying software vulnerabilities. They also talk about the CodeMender agent that can automatically write security fixes and contribute directly to open-source projects. July 2025 was a watershed moment for the Big Sleep AI agent. Sandra Joyce, vice president of Google Threat Intelligence, said at the time, “We believe this is the first time that an AI agent has been used to directly thwart attempts to exploit a vulnerability in the wild.”
- Google India says, “To help startups build securely from day one, we are open-sourcing CAPSEM (Capability Security for Agents), a secure runtime environment developed by our privacy and security research team.” To put it simply, CAPSEM is a capability-secure runtime environment (think of it as an isolated virtual machine) that securely restricts AI agents in terms of what they can access. If any agent is compromised or encounters a malicious signal, the broader system remains completely secure.
- Google also wants to be at the forefront of the conversation around an open standard for independent agents working in interconnected systems. There is a case for the Device Bound Session Credentials (DBSC) open W3C standard that ties cryptographically activated login tokens to a user’s physical device hardware, making stolen session cookies instantly useless to bad actors, and also for the Agent-to-Payments (AP2) protocol designed to make authorized, low-value agent-led financial transactions (less than $100) secure and accountable.
The latest on wired knowledge
a sign of a new beginning
Mira Muratti’s Thinking Machines Labs has finally released its first AI model, called Inkling. Murati, who was the former CTO of OpenAI (and was also briefly CEO during Sam Altman’s ouster in 2023), wrote in a post on They’re calling Inkling a mixture-of-experts transformer with 975 billion total parameters, 41B of which are active. It supports context windows of up to 1M tokens. It was pre-trained on 45 trillion tokens of text, images, audio and video. Those of you who follow my work will remember that back in September of 2024, I said that there were more chapters to be written in Mira Muratti’s legacy. To many people it might have seemed impossible at that time. It never was.
“This is the first in a family of models of different sizes: along with this we are sharing a preview of Inkling-Small, a lightweight model with 12B active parameters, trained with a uniform recipe, achieving strong performance even with low cost and latency,” the official post reads.
Inkling finds itself close to Zipu’s GLM 5.2 and OpenAI GPT-5.6 Sol in the AIME 2026 Reasoning and IFBench Chat benchmarks, but there is still a long way to go before Inkling catches up elsewhere.
“Inkling is designed to be broad. We trained it across agentic, reasoning, coding, instruction-following, factuality, vision, and audio tasks, rather than narrowly optimizing for a single domain. This breadth matters for customization and real-world use: different users need models that can adapt to very different workflows, not just excel at benchmarks,” the official post says, which sounds like an attempt to temper expectations.
China positions itself as an AI partner
Mehran Gul told me two things recently. First, people are increasingly realizing (especially after Anthropic took the Mythos model offline for a while) that they are not buyers of American technology, but rather tenants of American technology at best, who can be harvested at any time. And secondly, whereas in the past questions were asked of China about trust and credibility, the country is now presenting itself as a more reliable partner, and instead the same questions are now being asked of the US stack.
It was a good perspective ahead of the World Artificial Intelligence Conference (WAIC) 2026 in Shanghai, where Chinese Prime Minister Xi Jinping positioned China as an AI partner for the Global South.
“China is willing to work with all parties to take advantage of and address the opportunities and challenges of artificial intelligence development with a more open attitude, more practical actions, and a long-term perspective,” he said in his address, translated from Mandarin via Google Translate.
The AI battle has become quite heated in the past months. For every powerful AI model created by American companies, Chinese Frontier Labs respond with more affordable models, but with similar capabilities. According to the 2026 Stanford HAI AI Index report, the performance gap between top US and Chinese AI models has narrowed to a very small 2.7%, with the US leading in total top-tier model volume and private investment, while China leads in research volume, patents and citations.
OpenStreetMap, a service that lets developers and users access hundreds of different AI models (like Cloud, Gemini, and OpenAI’s GPT) through a single account and API key, notes in its latest Insights data that the Chinese model has overtaken the US model in token share in early June. Token is a basic unit of processed data.
OpenRouter, a US-based platform that routes traffic across hundreds of models, said Chinese-developed models overtook US models in their share of tokens – the basic unit of processed data – on the platform in early June. DeepSeek’s stake has doubled to 18%, from about 9% at the beginning of the year.
The analytics also notes that “A group of Chinese open source models, including Xiaomi, Minimax, and Tencent, have seen their token share increase over the past 6 months. This appears to have come at the expense of some major US model companies, notably Google and OpenAI.”