This is a very quick post. I had an idea as I was walking the dog this evening, and I wanted to build a functioning demo and write about it within a couple of hours.
While the post and idea started this evening, the genesis of the idea has been brewing for a while and goes back over a year to August 2024, when I wrote about being sucked into a virtual internet. WebSim has been on my mind for a while, because I loved the idea of being able to simulate my own version of the web using the browser directly and not via another web page. And a couple of weeks ago, I managed to work out how to get Puppeteer to intercept requests and respond with content generated via an LLM. — Read More
Monthly Archives: September 2025
Why Building Superintelligence Means Human Extinction (with Nate Soares)
The Voice Lives On: Moises Powers Whitney Houston’s Return to the Stage
Moises’ AI stem separation technology extracts Whitney Houston’s vocals from original recordings, enabling live orchestral performances across a seven-city tour
Whitney Houston’s voice moved generations, and through a collaboration between The Estate of Whitney E. Houston, Primary Wave Music, and Park Avenue Artists, it has now returned to the stage. The Voice of Whitney: A Symphonic Celebration, which debuted in August 2024, brings Houston’s legendary vocals to concert halls across US cities. The concert transports fans into Houston’s musical world, as live orchestras perform alongside Houston’s vocals and rare footage. Audiences experience the power of Houston’s voice in a live setting, with a breathtaking fusion of technology and artistry that celebrates her enduring legacy. — Read More
Mind the Gap: Time-of-Check to Time-of-Use Vulnerabilities in LLM-Enabled Agents
Large Language Model (LLM)-enabled agents are rapidly emerging across a wide range of applications, but their deployment introduces vulnerabilities with security implications. While prior work has examined prompt-based attacks (e.g., prompt injection) and data-oriented threats (e.g., data exfiltration), time-of-check to time-of-use (TOCTOU) remain largely unexplored in this context. TOCTOU arises when an agent validates external state (e.g., a file or API response) that is later modified before use, enabling practical attacks such as malicious configuration swaps or payload injection. In this work, we present the first study of TOCTOU vulnerabilities in LLM-enabled agents. We introduce TOCTOU-Bench, a benchmark with 66 realistic user tasks designed to evaluate this class of vulnerabilities. As countermeasures, we adapt detection and mitigation techniques from systems security to this setting and propose prompt rewriting, state integrity monitoring, and tool-fusing. Our study highlights challenges unique to agentic workflows, where we achieve up to 25% detection accuracy using automated detection methods, a 3% decrease in vulnerable plan generation, and a 95% reduction in the attack window. When combining all three approaches, we reduce the TOCTOU vulnerabilities from an executed trajectory from 12% to 8%. Our findings open a new research direction at the intersection of AI safety and systems security. — Read More
Jack Ma Returns With a Vengeance to ‘Make Alibaba Great Again’
During China’s yearslong crackdown on the tech sector, Alibaba Group Holding Ltd.’s internal messaging boards lit up with dreams to “MAGA” – Make Alibaba Great Again. Now, the company is deploying one of its most potent weapons to accomplish that mission: Jack Ma.
After vanishing from the public eye at the outset of an antitrust investigation in late 2020, China’s most recognizable entrepreneur is back on Alibaba’s campuses – and he’s more directly involved than he’s been in half a decade, according to people familiar with the company. Signs of his unseen hand are coming into sharper focus, perhaps no more so than in the company’s pivot to artificial intelligence and its declaration of war on e-commerce foes JD.com Inc. and Meituan. Ma was instrumental in Alibaba’s decision to spend as much as 50 billion yuan ($7 billion) on subsidies to beat back JD’s surprise entry to the market, said one of the people, requesting not to be named because the matter is private. — Read More
New Ultrasound Helmet Reaches Deep Inside The Brain Without Surgery
Deep-brain structures like the basal ganglia or the thalamus wield major influence on our behavior. If something goes awry, dysregulation in the deep brain may trigger neurological conditions like Parkinson’s disease or depression.
Despite the clear importance of these structures, our knowledge about them remains limited by their location, making them difficult to study and treat.
In a new study, researchers unveil a device that might offer an alternative to invasive procedures. Featuring a novel ultrasound helmet, it not only modulates deep-brain circuits without surgery, but reportedly can do so with unrivaled precision. — Read More
Read the Study
VaultGemma: The world’s most capable differentially private LLM
As AI becomes more integrated into our lives, building it with privacy at its core is a critical frontier for the field. Differential privacy (DP) offers a mathematically robust solution by adding calibrated noise to prevent memorization. However, applying DP to LLMs introduces trade-offs. Understanding these trade-offs is crucial. Applying DP noise alters traditional scaling laws — rules describing performance dynamics — by reducing training stability (the model’s ability to learn consistently without experiencing catastrophic events like loss spikes or divergence) and significantly increasing batch size (a collection of input prompts sent to the model simultaneously for processing) and computation costs.
Our new research, “Scaling Laws for Differentially Private Language Models”, conducted in partnership with Google DeepMind, establishes laws that accurately model these intricacies, providing a complete picture of the compute-privacy-utility trade-offs. Guided by this research, we’re excited to introduce VaultGemma, the largest (1B-parameters), open model trained from scratch with differential privacy. We are releasing the weights on Hugging Face and Kaggle, alongside a technical report, to advance the development of the next generation of private AI. — Read More
AI Will Not Make You Rich
Fortunes are made by entrepreneurs and investors when revolutionary technologies enable waves of innovative, investable companies. Think of the railroad, the Bessemer process, electric power, the internal combustion engine, or the microprocessor—each of which, like a stray spark in a fireworks factory, set off decades of follow-on innovations, permeated every part of society, and catapulted a new set of inventors and investors into power, influence, and wealth.
Yet some technological innovations, though societally transformative, generate little in the way of new wealth; instead, they reinforce the status quo. Fifteen years before the microprocessor, another revolutionary idea, shipping containerization, arrived at a less propitious time, when technological advancement was a Red Queen’s race, and inventors and investors were left no better off for non-stop running.
Anyone who invests in the new new thing must answer two questions: First, how much value will this innovation create? And second, who will capture it? Information and communication technology (ICT) was a revolution whose value was captured by startups and led to thousands of newly rich founders, employees, and investors. In contrast, shipping containerization was a revolution whose value was spread so thin that in the end, it made only a single founder temporarily rich and only a single investor a little bit richer.
Is generative AI more like the former or the latter? Will it be the basis of many future industrial fortunes, or a net loser for the investment community as a whole, with a few zero-sum winners here and there? — Read More
The Data Backbone of LLM Systems
QCON London: Drawing from his 8 years of experience in AI, Paul Iusztin breaks down the core components of a scalable architecture, emphasizing the importance of RAG. He shares practical patterns, including the Feature Training Inference architecture, and provides a detailed use case for creating a “Second Brain” AI assistant, covering everything from data pipelines to observability and agentic layers. — Read More
AI-Ready Data: A Technical Assessment. The Fuel and the Friction.
Most organizations operate data ecosystems built over decades of system acquisitions, custom development, and integration projects. These systems were designed for transactional processing and business reporting, not for the real-time, high-quality, semantically rich data requirements of modern AI applications.
Research shows that 50% of organizations are classified as “Beginners” in data maturity, 18% are “Dauntless” with high AI aspirations but poor data foundations, 18% are “Conservatives” with strong foundations but limited AI adoption, and only 14% are “Front Runners” achieving both data maturity and AI scale. — Read More
Rick's Cafe AI 