Welcome to Yury Akinin’s Digital Journey

I usually default to OpenAI for embeddings, but Google’s new EmbeddingGemma model is a noteworthy development. It’s not just another model; it’s a strategic move that shows real promise for improving Retrieval-Augmented Generation (RAG) pipelines, especially in on-device and edge applications. What is EmbeddingGemma? Google has released EmbeddingGemma as a lightweight, efficient, and multilingual embedding model. At just 308M parameters, it’s designed for high performance in resource-constrained environments. This isn’t just about making a smaller model; it’s about making a capable small model. ...

Physicists at the University of California, Irvine, have experimentally confirmed a new state of quantum matter, previously only theoretical. This discovery, centered on a material called hafnium pentatelluride (HfTe₅), has significant implications for developing next-generation electronics, particularly for deep space exploration. This new phase is a type of excitonic insulator—a state where electrons and the “holes” they leave behind spontaneously pair up. Uniquely, this is a spin-triplet excitonic insulator, where the paired electrons and holes have spins oriented in the same direction. When subjected to an intense magnetic field of up to 70 Teslas, the material’s ability to conduct electricity drops sharply, signaling its transition into this exotic state. ...

Two major players recently dropped new open-source models, but they represent two fundamentally different philosophies. OpenAI, the established leader, returned to the open-source scene with fanfare and its gpt-oss-20b model. Shortly after, the Chinese startup DeepSeek quietly released v3.1. While one was a media event, the other was a single tweet. The initial results from hands-on testing are starkly one-sided. Out-of-the-Box Performance: A Clear Winner When you evaluate a model as a tool to be used right now, the comparison is not even close. Across multiple practical tests, DeepSeek v3.1 consistently delivered superior results: ...

It’s a classic story at this point. We saw it recently with OpenAI’s ChatGPT, and now it’s Grok’s turn. Elon Musk’s xAI has inadvertently published hundreds of thousands of its users’ private conversations, making them fully searchable on Google. This wasn’t a sophisticated hack; it was a fundamental product design flaw. The Feature That Became a Bug The mechanism was simple and naive. When a Grok user hit the “share” button to send a conversation to a colleague or friend, the system generated a unique URL. However, instead of being a private link, this URL was made public and available for search engines to index. In effect, “sharing” meant “publishing to the open web” without any warning or disclaimer. ...

OpenAI’s proposal to encrypt AI is a commendable headline, but it sidesteps a more fundamental issue. Before we debate the complex philosophy of encrypting artificial intelligence, we should ask a simpler, more urgent question: have they patched the basic vulnerabilities in their existing systems? It’s easy to forget, but OpenAI has a history of security lapses, most notably the incident that leaked private user chat histories across the internet. This wasn’t a failure of advanced cryptography; it was a foundational security bug. They created a vulnerability and, as a result, exposed their clients’ private conversations. ...

The recent success of DeepSeek’s new AI model is more than just another headline—it’s a clear signal of a major shift in the global tech landscape. While the West has focused on restricting access to cutting-edge hardware, China has been playing a different game: achieving component independence by making good enough hardware work exceptionally well. While many are surprised that a company could develop a leading AI model without the latest NVIDIA chips, this outcome was predictable. China is strategically leveraging its core advantages: a massive domestic market and a deep pool of highly skilled, cost-effective software engineers. The core of their strategy isn’t just about building better hardware; it’s about optimizing software to extract maximum performance from the hardware they can produce domestically. ...

Quantum computing’s biggest roadblock isn’t speed; it’s stability. Qubits are notoriously fragile, easily collapsing from environmental noise. This makes scaling a reliable quantum computer an immense engineering challenge. One of the most promising solutions is topological quantum computing, which encodes information not in the state of a particle, but in the geometric “braiding” of quasiparticles called anyons. This approach is inherently more robust against decoherence. However, the leading candidates for this approach, known as Ising anyons, have a critical flaw: they aren’t “universal.” Performing computations by braiding them is like trying to type with half the keys missing from your keyboard—you can perform some operations, but not the full set required for general-purpose computing. ...

A significant strategic move is reshaping the cloud AI landscape: Oracle and Google Cloud have expanded their partnership, integrating Google’s advanced Gemini models directly into Oracle Cloud Infrastructure (OCI). This isn’t just another API integration; it’s a calculated decision that benefits both tech giants and, most importantly, their enterprise customers. What the Oracle-Google Partnership Means for Customers Effective immediately, OCI customers can access Google’s Gemini models, starting with Gemini 2.5, through the OCI Generative AI service. The key advantage here is seamless integration—businesses can use their existing Oracle Universal Credits to pay for Gemini usage, removing procurement friction and allowing them to build powerful AI agents for multimodal understanding, code generation, and workflow automation directly within their established cloud environment. ...

The success of RNA vaccines during the COVID-19 pandemic was a monumental scientific achievement. But it also highlighted a critical bottleneck: designing the delivery vehicle is as important as designing the RNA sequence itself. The process of creating the right lipid nanoparticles (LNPs) to protect and deliver RNA into our cells has been a slow, resource-intensive process of trial and error. New research from MIT, however, shows how AI is poised to break this bottleneck for good. ...

The Real Prize in Perplexity’s $34.5B Chrome Bid The tech world is buzzing about Perplexity AI’s unsolicited $34.5 billion offer for Google Chrome. On the surface, it looks like an audacious, almost impossible move from a startup valued at less than half that amount. Most analysts are dismissing it as a publicity stunt, and they aren’t entirely wrong—it’s a brilliant one. But to see it only as a PR play is to miss the fundamental shift happening in the AI landscape. This bid isn’t about buying a piece of software. It’s an attempt to acquire the single most valuable asset in the digital world: distribution. ...