标签：#an

Rivian's CEO RJ Scaringe is speaking at TechCrunch Disrupt 2026 to detail his journey, and the lessons it’s bestowed.

Every Angular application I have worked on in the last few years had the same three kinds of state: URL state — the page number, the active filter, the selected tab. Client state — what the user typed, what is expanded, what is selected. Server state — the thing you fetched, and everything that can go wrong while fetching it. And every application handled them three completely different ways. ActivatedRoute and a Router.navigate call for the first. Signals or a store for the second. A service returning an Observable , plus a loading boolean, plus an error field, plus a subscribe somewhere, for the third. None of that is wrong. It is just that the glue between them is written by hand, in every app, every time. And the glue is where the bugs live. This article is about the specific piece of that problem I could not let go of, and about the toolkit I ended up building around it. It is called craft-ng , it is in beta, and I would genuinely rather have your objections than your stars. The code I kept running into Here is the shape. I should be honest: I did not write much of it myself — I had a drawer of RxJS helpers that hid most of it. But I have read it in a lot of codebases, reviewed it in a lot of pull requests, and inherited it in a lot of projects. That turned out to matter more, because a helper that only I understand is not a solution to anything. @ Injectable () export class TaskListService { private http = inject ( HttpClient ); tasks = signal < Task [] > ([]); isLoading = signal ( false ); error = signal < string | null > ( null ); load ( done : boolean ) { this . isLoading . set ( true ); this . error . set ( null ); this . http . get < Task [] > ( `/api/tasks?done= ${ done } ` ). subscribe ({ next : ( tasks ) => { this . tasks . set ( tasks ); this . isLoading . set ( false ); }, error : ( err ) => { this . error . set ( ' Something went wrong ' ); this . isLoading . set ( false ); }, }); } } Four fields, one method, and roughly six ways to get it subtly wr

Phone and Bluetooth signals could turn roadside cameras into far richer tracking tools.

Picture a field technician standing in a basement with zero signal, trying to get an app to summarize a maintenance log and flag anything that looks like a safety issue. Or a language app that needs to correct pronunciation in real time, mid-commute, on a subway with no connectivity at all. A few years ago, both of those scenarios meant either building a degraded offline mode or just telling the user to try again later. Neither answer felt great. That's the actual reason on-device AI has become a real conversation in Android development in 2026, not because it's the trendy thing to bolt onto a feature list. Running inference locally solves specific, concrete problems: it keeps sensitive data off the network, it removes the round-trip latency of a cloud call, it works when there's no connectivity at all, and it gives you more predictable operating costs since you're not paying per-token for every user interaction. None of that means cloud AI is going away, and I'd be skeptical of anyone telling you it is. Most production apps in 2026 end up running a mix of both. But there's now a real, practical case for pushing certain workloads onto the device itself, and that's what this article is actually about - where local inference genuinely helps, where it falls short, and what it takes to build it properly in Kotlin. What Is On-Device AI? On-device AI means running a machine learning model directly on the user's phone, using the device's own CPU, GPU, or NPU, instead of sending a request to a server somewhere and waiting for a response. The model, or at least the parts of it needed for inference, lives on the device. Cloud AI still has the advantage in raw model size and reasoning depth - nobody's running a 70-billion-parameter model on a phone, at least not yet. But for narrower, well-defined tasks, on-device models have become genuinely capable, and the trade-offs are worth understanding side by side. Factor On-Device AI Cloud AI Inference location Runs on the user's dev

The problem When you type a single or double quote in Atlassian Jira or Confluence wiki, it's automatically converted to a "smart" quote 🤮 Examples “double” ‘single’ I don’t like it If you delete a smart quote and type it again, it becomes smart again, which can be disturbing. There is no option to disable smart quotes permanently. But... You can get rid of them Type Ctrl-Z immediately after they appear. It also works for the word don't : When you type apostrophe ' it looks normal, then when you type t , the apostrophe changes. Type Ctrl-Z and it reverts back to normal again.

Researchers say it took fewer than 20 prompts for a public AI tool to find a flaw (now fixed) allowing anyone on a Zoom call to hijack another participants’ device.

Evolusi Pemrosesan Data: Dari Batch ke Real-Time Transformasi infrastruktur data mendorong transisi dari pemrosesan batch statis ke arsitektur streaming. Integrasi LLM kini menjadi komponen inti sistem data terdistribusi, di mana kecepatan pemrosesan informasi menentukan relevansi dan akurasi output AI secara instan. Tantangan Latensi pada Integrasi LLM Langsung Menyematkan LLM dalam pipeline real-time memicu tantangan sinkronisasi state dan overhead komunikasi antar-node. Bottleneck utama biasanya terjadi pada transfer KV cache dan keterbatasan bandwidth memori, yang menghambat performa inferensi pada skala terdistribusi. Strategi Optimasi: TensorRT-LLM dan Arsitektur Asinkron Optimasi melalui TensorRT-LLM krusial untuk menekan latensi token-to-token melalui optimalisasi kernel CUDA dan manajemen memori yang lebih efisien. Selain itu, arsitektur asinkron seperti Pathways memungkinkan eksekusi grafik dataflow dinamis, meminimalkan idle time pada akselerator GPU/TPU. Kapan Memilih Pendekatan Batch Tradisional? Pendekatan batch tetap superior untuk tugas non-sensitif waktu. Efisiensi biaya (cost-efficiency) dan throughput tinggi menjadikan metode ini pilihan utama untuk pemrosesan dataset masif, seperti pelatihan ulang model (retraining) atau analisis historis skala besar. Masa Depan: Arsitektur Hibrida untuk Skala Besar Sistem masa depan akan mengadopsi model hibrida: inferensi kritis latensi dijalankan di edge atau buffer lokal, sementara pemrosesan berat tetap berada pada jalur batch. Strategi ini memaksimalkan data-locality dan mengoptimalkan alokasi sumber daya komputasi.

Anthropic has pledged to start marking Claude-generated text and images with machine-readable data, in an effort to comply with European rules for AI transparency. "Generated text will carry embedded watermarks, and generated files will include digitally signed provenance metadata where supported," Anthropic says on a new Claude support page. The changes are invisible to human […]

Anthropic will extend support for watermarking AI generations for older models as well.

Joby Aviation announced it was acquiring Dayton, Ohio-based defense firm Resonant Sciences in a $500 million deal, in a bid by the electric aircraft company to expand further into the military industrial complex. Joby says it expects to finance the deal with $450 million in cash and $50 million in equity. After the deal closes, […]

Why passkey apps treat Windows differently than other operating systems.

Meryem Arik discusses strategies for designing low-cost LLM inference architectures for high-volume, non-real-time workloads. She explains how software architects and engineering leaders can achieve order-of-magnitude cost reductions by making critical trade-offs across hardware, inference runtimes, speculative decoding, and smart queue reordering. By Meryem Arik

OpenAI begins testing ads in ChatGPT to support free access, with clear labeling, answer independence, strong privacy protections, and user control.

Startups and Big Ag are using Crispr gene editing to create crops that taste better and grow on a hotter planet. But will they find a market?

WordPress sitelerinde hız problemi yaşandığında genellikle ilk refleks bir cache eklentisi kurmak oluyor. Bazen gerçekten fark yaratıyor, bazen de PageSpeed puanı biraz yükselmesine rağmen site hâlâ yavaş hissettiriyor. Bunun nedeni WordPress performansının tek bir ayardan oluşmaması. Hosting, tema, eklentiler, görseller, JavaScript dosyaları ve veritabanı aynı anda sayfanın yüklenme süresini etkileyebiliyor. Bu nedenle bir siteyi hızlandırmaya başlamadan önce asıl problemin nerede olduğunu bulmak gerekiyor. WordPress projelerinde sık karşılaştığım 7 performans hatasını aşağıda topladım. Gereğinden büyük görseller kullanmak En sık karşılaştığım problemlerden biri bu. Sayfada 700 piksel genişliğinde görüntülenecek bir görselin 4000-5000 piksel olarak yüklenmesi oldukça yaygın. Özellikle yüksek çözünürlüklü stok fotoğraflar doğrudan WordPress'e yüklendiğinde tek bir görsel birkaç megabayta ulaşabiliyor. Bu da özellikle mobil bağlantılarda ciddi yük oluşturuyor. Görselleri yüklemeden önce kullanılacağı alana uygun boyuta getirmek, sıkıştırmak ve mümkün olduğunda WebP veya AVIF gibi modern formatları tercih etmek önemli. Ancak yalnızca dosya formatını değiştirmek yeterli değil. 4000 piksel genişliğindeki bir görseli WebP'ye çevirmek, görselin gereğinden büyük olduğu gerçeğini değiştirmiyor. Her problemi cache eklentisiyle çözmeye çalışmak Cache WordPress performansında önemli bir yere sahip. Fakat cache eklentisi kurmak her hız problemini ortadan kaldırmaz. Sunucu geç cevap veriyorsa, çok fazla JavaScript çalışıyorsa veya veritabanında ağır sorgular varsa cache yalnızca problemin bir bölümünü gizleyebilir. Ayrıca aynı anda birden fazla optimizasyon eklentisi kullanmak da başka sorunlara yol açabiliyor. Örneğin bir eklentide JavaScript erteleme, başka bir eklentide tekrar JavaScript optimizasyonu ve hosting panelinde üçüncü bir optimizasyon sistemi açıldığında hangi ayarın ne yaptığını takip etmek zorlaşıyor. Ben mümkün olduğunca tek bir ana cache sistemi üzerinden ilerl

Imagine working as a Data Analyst in a healthcare Non-Governmental Organization (NGO) implementing HIV and AIDS programmes across several communities. The organization has limited resources. There may not be enough funding, healthcare workers, testing kits, transport, outreach teams, or community programmes to serve every community at the same intensity. This creates an important question: How can we use data and machine learning to direct limited programme resources to communities with the greatest need? This is where Machine Learning (ML) can become valuable. Rather than distributing resources equally across all communities, an NGO can use historical programme data to identify communities experiencing greater HIV-related service gaps or higher levels of need. Resources can then be prioritized based on evidence. What Is Machine Learning? Machine Learning is a branch of Artificial Intelligence that enables computers to learn patterns from data and use those patterns to make predictions or support decisions. Instead of manually creating rules for every situation, you provide the algorithm with historical data and allow it to identify relationships within that data. For example, the NGO could have this information about different communities: Community HIV Testing Coverage ART Coverage Missed Appointments Outreach Activities Community A 85% 90% 5% High Community B 52% 61% 25% Low Community C 70% 75% 15% Medium Community D 40% 55% 32% Low Looking at this data, Community D appears to have greater programme gaps than Community A. However, in a real programme, the decision should not be based on one indicator alone. Machine learning can analyse many variables simultaneously to identify communities that may require greater attention. Why Resource Allocation Matters in HIV Programmes HIV programmes operate in environments where resources are often limited. An NGO may have: A limited number of community health workers A fixed outreach budget Limited HIV testing supplies Limi

How We Built an IoT Platform That Handles 30 Million Concurrent Connections — With a Team of 10 DGIOT is an open-source industrial IoT platform. We run 928 gateways across 16 oil fields, process 652 million data points, and maintain 99.9999% uptime. Here's the architecture that makes it possible. The Problem In 2021, we got a call from Daqing Oil Field — China's largest oil producer. They had a problem: 928 industrial gateways from different vendors 114,809 sensor points speaking 15 different protocols Data collection every 10 minutes (they needed seconds) 15-30 minute end-to-end latency (they needed <3 seconds) False alarm rate above 20% The existing system was a patchwork of vendor-specific tools, each with its own database, UI, and authentication. Operators had to log into 8 different systems just to check if a pump was overheating. They asked: "Can you unify this?" What We Built DGIOT is an Erlang/OTP-based platform that acts as a universal translator for industrial protocols. Think of it as a Rosetta Stone for machines. Modbus ─┐ OPC UA ─┤ MQTT ──┼──→ Unified Pipeline ──→ TDengine ──→ Dashboard IEC104 ─┤ A11 ──┘ The key insight: industrial protocols are just state machines . Once you model each protocol as a gen_statem FSM in Erlang, you can handle hundreds of them concurrently with almost zero overhead. The Architecture: DLAS We designed a four-layer architecture that separates concerns cleanly: Layer 1: DATA — Ingestion Parse Server (23 classes) handles device metadata, user auth, tenant isolation TDengine stores 652M time-series data points with 10:1 compression EMQX handles MQTT message routing at 1M+ msg/sec Mnesia/ETS provides in-memory caching for hot data Layer 2: LOGIC — Ontology Engine This is our secret weapon. We built a 252-entity OWL ontology that models industrial equipment: Pump ⊑ Equipment ⊓ ∃ hasPart.Bearing ⊓ ∃ measures.Pressure Bearing ⊑ Component ⊓ ∃ hasFailureMode.Overheat Overheat → triggers ( Alert ) ∧ reduces ( RemainingLife , 0.8 ) The

Introduction: The Debate Over Code Comments In the trenches of software development, a quiet but fierce debate rages: are code comments still relevant? On one side, the prevailing narrative dismisses comments as "mostly useless" —redundant, outdated, or worse, misleading. This perspective has gained traction, fueled by the rise of self-documenting practices like meaningful variable names and modular design. Developers, under the gun of tight deadlines, increasingly treat comments as an afterthought, if not a burden. The result? Comments are neglected, both in writing and reading, creating a self-fulfilling prophecy of their uselessness. But here’s the rub: this dismissive attitude is flawed. When used thoughtfully, comments are not just useful—they’re critical. The problem isn’t comments themselves but how they’re misused or ignored. Poorly written comments, lack of maintenance standards, and time constraints have deformed their purpose, turning a powerful tool into a liability. For example, a stale comment explaining a function’s behavior can lead a developer to misinterpret the code, causing bugs that cascade through the system. The mechanism here is clear: impact (misleading comment) -> internal process (developer misinterprets code) -> observable effect (bugs introduced) . The stakes are high. As software complexity grows and developer turnover accelerates, the need for clear, maintainable code becomes non-negotiable. Comments, when crafted with intent, act as a bridge between the code’s logic and the human mind, reducing cognitive load and fostering collaboration. Neglecting them risks eroding code readability, maintainability, and team productivity—a risk that materializes when a new developer inherits a poorly documented codebase and spends hours deciphering its intent. This investigation challenges the dismissive attitude toward comments, dissecting their underappreciated utility through real-world examples. By addressing misconceptions and promoting best pr

OpenAI is expanding its AI cybersecurity defense program Daybreak, and rolling out a new cyber-trained AI model with it.

Meta has been trailing competitors. Zuckerberg thinks he's found a way forward.