标签：#low

This is Lowpass by Janko Roettgers, a newsletter on the ever-evolving intersection of tech and entertainment, syndicated just for The Verge subscribers once a week. "AI is the new frontier for us," says Marc DeBevoise, who took over as the new CEO of OverDrive last week. OverDrive is best known for the ebook lending app […]

For the last few years, many enterprise AI conversations have started with the same question: “Where can we add an AI copilot?” It is an understandable starting point. Copilots are familiar. They sit inside existing tools, help users draft content, summarize information, search documents, write code, or answer questions. For teams experimenting with AI, they feel safe. But after 10 years of building mobile apps, web platforms, AI systems, internal tools, and enterprise-grade products, I have learned something that sounds simple but changes the whole strategy: The workflow is the product. Not the chatbot. Not the prompt box. Not the model. Not the dashboard. The workflow. Enterprise AI only becomes valuable when it changes how work actually moves across people, systems, approvals, decisions, and data. That is why companies now need to move beyond standalone copilots and toward AI workflow automation, enterprise AI agents, and agentic workflows that are designed around real operational outcomes. Copilots Help. Workflows Transform. An AI copilot is useful when a person needs assistance inside a task. It can draft an email, summarize a meeting, search policy documents, or help an engineer understand code. These are valuable use cases. But they usually improve a single moment of work, not the complete business process. A workflow, on the other hand, connects the full chain. For example, consider enterprise customer onboarding. A copilot may summarize the sales call. A workflow system can take that summary, extract requirements, identify missing information, create onboarding tasks, notify customer success, update the CRM, generate a kickoff plan, check billing setup, and flag delivery risks. That is a very different level of impact. AI Copilot AI Workflow Automation Assists one user Coordinates work across teams Responds when asked Triggers actions automatically Works inside a tool Connects multiple systems Improves productivity Improves operating performance Helps with

Research Context When doing low-level network programming in Assembly, you experience firsthand the immense chaos running behind the scenes of operations we solve with a single line in high-level languages (Python, C, etc.). While developing the Nested-ICMP-Communication Analysis project, specifically an Encapsulated ICMP framework, I hit exactly this kind of wall: extracting an IP address from a packet header and printing it to the screen in the correct format. Sounds simple, right? However, when x86 architecture and network protocols are involved, seeing 5.1.168.192 instead of 192.168.1.5 on your terminal is extremely common. So why does this happen, and what kind of algorithm did I develop to overcome this issue during the debugging process? Let's dive into the background. The Endianness Problem in Network Headers When you capture a packet coming over the network and read the source/destination IP address inside the sockaddr_in structure, the data arrives in Network Byte Order (Big-Endian) format. This means the most significant byte is stored at the lowest memory address. However, the x86/x64 processor architectures we use rely on Little-Endian (Host Byte Order). When the processor pulls this 4-byte IP data into a register, the reading direction is effectively reversed for our purposes. The result? A packet that arrives as 192.168.1.5 appears scrambled if we try to naively print it from memory. The inet_ntoa() function in high-level languages handles this conversion in the background. But if you are writing a custom sniffer in pure Assembly, you must do this conversion byte by byte yourself. Debugging Hell: The Problems Encountered While writing this conversion, I encountered a few critical issues that cost me hours in GDB (GNU Debugger): Register Clashes: While separating each octet (byte) of the IP address and converting it to an ASCII character (string), you must use the AX register for division operations (DIV). If you don't carefully manage your remainders

A small word that changes the rhythm of a job For as long as I have been writing Actions workflows I have been carrying a quiet workaround in my head. Want to warm a cache while the build runs? Append & to the shell command, then squint at logs that arrive out of order and pray the job doesn't exit on you. It worked, sort of. It also meant that anything more interesting than "run one thing, then the next thing" lived as folklore, hidden inside run: blocks. GitHub closed that gap this week. On June 25 the Actions changelog announced that steps inside a job can now run concurrently, marked with a new background keyword and supported by helpers to wait for them and cancel them. Until now, the changelog notes, every step in a workflow ran in sequence, with each step starting only after the previous one completed. That single rule has shaped every workflow I have ever written. It is gone, and the replacement is the kind of feature you don't notice until the day you reach for it and it's there. What the keywords actually do There are four pieces, all of them documented in the announcement. background: true is the entry point. Set it on a step and that step starts running, and the next step starts immediately. It does not block the job. wait and wait-all are the rendezvous. wait pins on one or more named background steps and pauses until they finish. wait-all is the same idea against every background step still in flight. Either way you get back into a linear flow on your terms. cancel is the cleanup. It gracefully terminates a background step when you no longer need it, which is the missing piece if you have ever tried to kill a long-running side process from inside a job and ended up shelling out to kill . parallel is the convenience wrapper. The changelog describes it as taking a group of steps and converting them into background steps with a wait placed after. For the common "fan out, then join" shape, you write one block instead of decorating five steps by hand. Where

This is Lowpass by Janko Roettgers, a newsletter on the ever-evolving intersection of tech and entertainment, syndicated just for The Verge subscribers once a week. The 2026 World Cup is breaking streaming records around the world: Brazil's CazéTV YouTube livestream of that country's opening game against Morocco surpassed 12 million concurrent viewers, a new milestone […]

Dai metadati in una lingua a 20 localizzazioni senza impazzire tra click e schermate: un flusso pratico per indie e piccoli team. Localizzare un’app non significa solo tradurre le stringhe dell’interfaccia. Una buona parte dell’acquisizione organica passa dai metadati su App Store Connect : titolo, sottotitolo, descrizione e keyword. Il problema è che, quando provi a farlo “a mano” dal pannello web, diventa subito un lavoro di pura resistenza: apri la scheda, cambi lingua, compili i campi, salvi, ripeti. Ora moltiplica per 10–20 lingue. Per molti indie (e in generale per chi ha poco tempo e zero voglia di click ripetitivi) il punto di svolta è usare ASC CLI per rendere questa attività automatizzabile, ripetibile e verificabile . Perché la localizzazione dei metadati è un caso d’uso perfetto per una CLI Dal punto di vista del flusso di lavoro, i metadati App Store hanno tre caratteristiche che li rendono ideali per l’automazione: Sono campi strutturati (title, subtitle, description, keywords): non stai “inventando” contenuti ogni volta, stai trasformando contenuti. Sono ripetitivi per lingua : la sequenza di operazioni è identica, cambia solo la locale. Sono tanti : più lingue aggiungi, più l’approccio manuale scala male (tempo, errori, incoerenze). Con una CLI, invece, il lavoro si sposta dal “fare cose” al definire un processo : prendi i metadati di partenza, generi le varianti linguistiche, applichi l’update in batch. Cosa conviene localizzare (e cosa no) In genere ha senso includere in un passaggio di localizzazione “massiva”: App name / title (attenzione ai limiti e ai trademark) Subtitle (spesso è la parte più ASO-oriented) Description (qui conta più la leggibilità che la traduzione letterale) Keywords (campo delicato: va adattato, non tradotto alla cieca) Al contrario, è meglio trattare con più cautela: Claim e frasi marketing molto creative : in alcune lingue risultano innaturali se tradotte letteralmente Keyword strategy : la ricerca utenti cambia per mercat

Two men pleaded guilty in the United Kingdom this week to criminal charges stemming from an August 2024 cyberattack that crippled Transport for London, the entity responsible for the public transport network in the Greater London area. The duo were key members of a prolific cybercrime group known as Scattered Spider, and their guilty pleas came on the first day of what was expected to be a six-week trial.

46.4%. That number — ChatGPT's June 2026 market share — ends a streak that held since November 2022. For the first time since the product launched, OpenAI holds less than half the AI assistant market. Gemini is at 27.7%. Claude is at 10.3%. The monopoly phase of AI assistants is over. The data comes from a June 2026 market report tracking monthly active users across major AI assistants. ChatGPT still leads with 1.11 billion monthly users — a number that would define the entire category in any other software market. But Gemini has 662 million, up 129 million in five months. Claude sits at 245 million, nearly four times its December 2025 count of 60.2 million. The trajectory is the story, not the absolute numbers. Why the 50% Threshold Actually Matters Below 50% doesn't mean decline. ChatGPT's absolute user count keeps growing. What the threshold signals is the end of single-platform dominance — the condition where building for "AI users" meant building for ChatGPT users. That assumption no longer holds in mid-2026. For context: search engine market share stayed above 90% for Google for nearly a decade after competitors entered. Social network market share for Facebook stayed above 70% for years after Instagram and Twitter had genuine scale. The pace of AI assistant fragmentation is meaningfully faster than those precedents. Three products above 10% share in under two years of real competition is an unusually fast split. What fragmentation means practically: the community knowledge base — YouTube tutorials, Reddit threads, prompt libraries — that once pointed almost exclusively at ChatGPT now covers three platforms with genuine depth. That changes how you can expect your users to arrive at your AI-integrated product, and what they already know about AI when they get there. Gemini's 662 Million Users Are Not What They Look Like Gemini's surge from under 500 million to 662 million monthly users in five months is impressive on paper. The driver is less impressive: Google

A stat from 2026 that should stop you cold: 88% of organizations reported a confirmed or suspected AI agent security incident in the past year (92.7% in healthcare). And more than half of all agents run with no security oversight and no logging — naked. The problem isn't that the AI isn't smart enough. It's that almost nobody welded boundaries around it. And boundaries are exactly where rigor lives. The incident list: speed flooring it, boundaries naked The last couple of weeks of security signals line up scarily well: 88% of orgs reported confirmed/suspected AI agent incidents in the past year; healthcare 92.7% ; over half of agents have no security oversight or logging. Supply chain is the front door. A plugin-ecosystem supply-chain attack harvested agent credentials from 47 enterprise deployments ; attackers used them to reach customer data, financial records, and proprietary code — undetected for six months. A public skills marketplace at one point hosted 824 of 10,700 malicious "skills." Config is an attack surface. Check Point disclosed remote code execution in a popular coding agent via poisoned repository config files ; MCP (Model Context Protocol) is the connective tissue across nearly every incident this year — poisoned configs, malicious marketplace skills, unauthenticated exposed MCP servers. By early 2026, at least ten public incidents across six major AI coding tools were attributed to " agents acting with insufficient boundaries. " The industry's own summary: AI agent security in 2026 is a supply chain problem first, a prompt-injection problem second. And every one of these shares a single root cause — the agent can act, but there's no architectural boundary on what it can touch, change, or call. Why "naked" is inevitable: bolt-on boundaries always leak Why do half the agents run with no oversight? Because in the mainstream approach, boundaries are bolt-ons : an allow-list here, a gateway there, logs you read after the fact. The trouble: The tools an

We’ve all been there: waking up feeling like a zombie despite getting eight hours of sleep. While wearables give us data, they often fail to give us foresight . What if you could predict your stress levels 24 hours in advance? 🚀 In this tutorial, we are going to tackle HRV prediction (Heart Rate Variability) using a state-of-the-art Temporal Convolutional Network (TCN) . By leveraging the Oura Ring API and deep learning, we’ll transform non-stationary biometric time series into actionable insights. Whether you're into time series forecasting or building the next big health-tech app, mastering Temporal Convolutional Networks (TCN) is a game-changer for handling long-term dependencies without the vanishing gradient headaches of traditional RNNs. For those looking for more production-ready examples and advanced biometric signal processing patterns, I highly recommend checking out the deep-dives at WellAlly Blog , which served as a major inspiration for this architecture. The Architecture: Why TCN? Traditional LSTMs are great, but they process data sequentially, making them slow and prone to memory loss over long sequences. TCNs, however, use Dilated Causal Convolutions , allowing the model to look back exponentially further into the past with fewer layers. Data Flow Overview graph TD A[Oura Cloud API] -->|Raw JSON| B(Pandas Preprocessing) B -->|Cleaned HRV/Activity| C{Feature Engineering} C -->|Sliding Windows| D[TCN Model Training] D -->|Dilated Convolutions| E[Stress Trend Prediction] E -->|24h Forecast| F[Dashboard/Alerts] style D fill:#f9f,stroke:#333,stroke-width:2px Prerequisites To follow along, you'll need: Tech Stack : Python, TensorFlow/Keras, Pandas, Scikit-learn. Data : An Oura Cloud Personal Access Token (or use the mock data generator provided). Difficulty : Advanced (Buckle up! 🏎️). Step 1: Fetching Biometric Data First, we need to pull our "Readiness" and "Sleep" data. Oura provides high-resolution HRV samples (usually 5-minute intervals during sleep).

This week's AI tooling news splits cleanly between infrastructure you can ship today and capability bets that require more careful evaluation. Anthropic dropped two significant releases—Fable 5 and Managed Agents updates—while the Workflow SDK landed a cancellation primitive that eliminates entire categories of homegrown plumbing. Underneath all of it, a sharp incident review from Anthropic is the most practically useful thing published this week if you're running multi-turn agents in production. Workflow SDK adds AbortController cancellation support The Workflow SDK now threads AbortSignal through workflow steps, using the same web-standard API you already use with fetch . Pass an AbortSignal into your workflow, inspect it inside steps, and you get cooperative cancellation that survives durable suspension and replay. This matters because cancellation in long-running workflows has historically required custom infrastructure—timeout flags passed through context, manual cleanup hooks, bespoke race logic. That's not interesting code to write or maintain. With AbortController support, you get timeout steps, request racing, and parallel work cancellation with patterns your team already knows. Two important caveats: this requires workflow@beta , and cancellation is cooperative. The runtime won't forcibly terminate a step—your step code needs to inspect the signal and respond. If you have steps with opaque third-party calls that don't accept signals, you're still writing wrapper logic. Verdict: Ship. If you're on Workflow SDK 5 and running long-horizon workflows with timeout or race requirements, upgrade and wire this in now. The pattern is standard, the boilerplate reduction is real, and there's no meaningful downside if your steps are already structured around explicit control flow. Anthropic adds dreaming, outcomes to Managed Agents Two distinct additions here. Outcomes let you define explicit success criteria enforced by a separate grader agent—replacing manual prompt

This is Lowpass by Janko Roettgers, a newsletter on the ever-evolving intersection of tech and entertainment, syndicated just for The Verge subscribers once a week. When Fox announced its acquisition of Roku earlier this week, executives of both companies were quick to promise that not much would change in the near future. Sure, getting its […]

Google's demo reel for Gemini Omni looks effortless: ask for a video, then keep talking to it until the shot is right. The question for developers is whether that conversational loop holds up outside a stage demo — and what it actually changes versus the Veo workflow it replaces. What Does Omni Add That Veo Couldn't? Omni's core addition is state. Veo produced one-shot renders — each prompt generated a fresh clip with no memory of the last. Gemini Omni holds context across turns, so changing the camera angle on turn three preserves the characters and lighting established on turn one without restarting the scene . Announced at Google I/O on May 19, 2026, the first shipped model, Gemini Omni Flash, replaces Veo as the video-generation surface in the Gemini app . Product director Nicole Brichtova framed it as "the next step towards combining the intelligence of Gemini with the rendering capabilities of our media models" — DeepMind's informal pitch is a "Nano Banana for video," extending conversational image editing to motion footage. Two claims deserve a skeptical read. Google advertises "intuitive understanding of forces like gravity, kinetic energy, and fluid dynamics," but those physics behaviors currently rest on Google demos and creator footage, with no third-party benchmarks published at launch . And on raw output, independent reviewers put Omni's generation quality on par with Veo 3.1 rather than clearly above it . The differentiation is the iterative editing loop and Gemini-grounded reasoning — not a new render engine. Before Starting: Paid Membership, Region, Age Omni access is gated behind a paid Google AI plan and a few hard eligibility rules, so confirm these before you open a prompt. Gemini Omni Flash unlocks in the Gemini app and Google Flow for Google AI Plus, Pro, and Ultra subscribers, with Plus starting at $7.99/month . If you want to test it for free, generation is available at no cost on YouTube Shorts and the YouTube Create App at launch . Two cons

Introduction Enterprise orchestration is evolving. As systems grow more distributed and...

Stack Overflow has announced Stack Overflow for Agents, a beta API-first knowledge exchange aimed at AI coding agents rather than human developers. The service is presented as a way to close what the company calls the Ephemeral Intelligence Gap, where agents repeatedly rediscover the same fixes and patterns in isolation instead of sharing them through a common memory. By Matt Saunders

Obsidian prompts beat open-ended reflection every time: median review time across 6 weeks was 14 minutes, fastest was 9, slowest was 22 (and that week genuinely deserved 22). I ran the GTD-adjacent version faithfully for six weeks — 90 minutes, full capture sweep, energy audit, the works. Then less faithfully for two months. Then I stopped entirely and didn't notice for three weeks. That last part is the failure mode nobody writes about. The format wasn't wrong; it was sized for a version of my week that rarely existed. The fix wasn't a better framework. It was shorter, closed questions. My Obsidian template has seven prompts, none of them open-ended: what shipped, what didn't, what I avoided and why, one thing to drop, one thing to protect. One-to-three sentence answer ceiling per prompt, hard stop. Open questions like "how was your week?" generate rumination. Closed questions generate decisions. That distinction is doing almost all the work. The Notion version I ran before this taught me something useful about tool selection too. I built rollups — tasks closed this week, open tasks by project, inbox count, stalled for 7+ days — and they worked exactly as designed. What Notion couldn't do was get out of its own way during actual reflection. Every time I tried to think through what went wrong, I'd end up reorganizing a database instead. Forty minutes later, new linked database, zero review completed. The same flexibility that makes Notion a good data layer makes it a bad "close the loop and move on" environment. Obsidian's plain-file simplicity is the right call for the thinking layer — and completely wrong for the data layer. Neither tool alone is the honest answer. There's also a cautionary note from my automation setup: a Zapier zap that pushed completed tasks into Notion for weekly rollup ran cleanly for two months, then silently broke when my task manager updated their API response format. Modified tasks started logging as completed. My rollup became noise befo

PyTrees look deceptively simple. You flatten a nested Python object into leaves, keep a structure descriptor, and later rebuild or map over the same shape. That abstraction is powerful enough to carry optimizer states, model parameters, batched inputs, gradients, and sharding annotations. It is also just ambiguous enough that three major frameworks implement three subtly different languages under the same idea. This note compares JAX jax.tree_util , PyTorch torch.utils._pytree , and TensorFlow tf.nest . I tested the behavior in two environments: an older stack with JAX 0.4.35, PyTorch 2.2.2, TensorFlow 2.20.0, and a newer stack with JAX 0.10.0, PyTorch 2.12.0, TensorFlow 2.21.0. Most flatten/unflatten semantics were stable across these versions. The main version-sensitive result is PyTorch: _pytree.tree_map in 2.2.2 accepts only one pytree, while 2.12.0 supports multiple pytrees and behaves much closer to JAX prefix-style mapping. The short version: JAX treats pytrees as a transformation language, PyTorch is converging toward that model in torch.func , and TensorFlow exposes a broader nested-structure utility through tf.nest . Those differences show up exactly where backend-agnostic libraries usually hurt: None , dictionary order, custom containers, tree_map , autodiff, and vectorization. The Shape Of The APIs The three APIs have the same surface story but not the same contract. from jax import tree_util as jtu from torch.utils import _pytree as tpu import tensorflow as tf leaves , treedef = jtu . tree_flatten ( tree ) tree = jtu . tree_unflatten ( treedef , leaves ) tree = jtu . tree_map ( f , * trees ) leaves , spec = tpu . tree_flatten ( tree ) tree = tpu . tree_unflatten ( leaves , spec ) tree = tpu . tree_map ( f , tree ) # PyTorch 2.2.2 tree = tpu . tree_map ( f , * trees ) # PyTorch 2.12.0 leaves = tf . nest . flatten ( tree ) tree = tf . nest . pack_sequence_as ( structure , leaves ) tree = tf . nest . map_structure ( f , * structures ) Flattening means "whi

This is Lowpass by Janko Roettgers, a newsletter on the ever-evolving intersection of tech and entertainment, syndicated just for The Verge subscribers once a week. Severance. Pachinko. Silo. Ted Lasso. Over the past couple of years, a number of Apple TV shows have become hits with audiences and critics alike. And yet, compared to the […]

Webflow's code export has two problems. It is only available on paid Workspace plans, and even when you pay, it does not include your CMS content: collection lists export as empty states, collection pages export with nothing in them. If your site has a blog, the export gives you a site without a blog. Forms and search are disabled in exported code too, per Webflow's own docs. Meanwhile, the published site is sitting on a CDN, fully rendered. Every CMS page is real HTML. wget --mirror will happily fetch all of it. What wget gives you, though, is not deployable. I migrated a production Webflow site this way and hit the same five breakages everyone hits, so I turned the fixes into a Claude Code skill that runs the whole workflow. This post is the five breakages, because they are useful whether or not you use the skill, and they apply to Framer, Squarespace, and friends with different domain names. Setup: the mirror itself The one wget incantation that matters, because Webflow serves assets from a separate CDN domain and you have to tell wget to follow it: wget --mirror --convert-links --adjust-extension \ --page-requisites --span-hosts \ --domains = yourdomain.com,cdn.prod.website-files.com \ --no-parent https://yourdomain.com/ This downloads every page plus the CSS, JS, images, and fonts they reference, and rewrites URLs to relative paths. It looks complete. It is about 90% complete, and the missing 10% is invisible until the page renders blank. Breakage 1: the page renders blank, console says "integrity" The symptom: your mirrored page shows raw unstyled text or nothing at all, and the console says Failed to find a valid digest in the 'integrity' attribute . The cause is subtle. Webflow ships its <link> and <script> tags with SHA-384 SRI hashes. wget's --convert-links rewrites URLs inside the downloaded CSS files, which changes their bytes, which means the SRI hash no longer matches, which means the browser silently refuses to apply the stylesheet. The file is right

Every low-code vendor says "we support customization." But supports is a weasel word — recoloring a button is customization, and rewriting a scheduling engine is also customization. What actually decides whether a platform locks you in is how far up its extensibility goes. Here's a ruler. The three levels of customization Level What you can do Most no-code A real dev framework L1 — Config Fields, forms, workflows, permissions, themes ✅ ✅ L2 — Extension Custom components, custom actions, external API calls, business rules ⚠️ limited ✅ L3 — Framework Modify/extend the core, custom engines, deep rewrites, source under control ❌ wall ✅ (when open/controllable) Where it stops is where your ceiling is. Plenty of no-code platforms are delightful at L1, then hit "can't do that" at L2/L3 — and you retreat to writing your own thing next to it. Now low-code is the burden. Why you get locked in Black-box SaaS — no source, so any extension point the vendor didn't expose is simply out of reach. Two sources of truth — your extension code and the platform's config live in different systems, so a platform upgrade breaks/voids your work. Crippled self-hosting — the on-prem edition quietly drops extension capabilities. Closed ecosystem — only their component marketplace; your stack can't get in. How model-driven + open source raises the ceiling One unified extension system — your extensions (custom fields/components/actions) and the platform itself are built on the same metadata. Extension isn't a bolt-on, it's a first-class citizen — upgrades don't wipe your customizations. Source under your control — open + self-hostable is what makes L3 framework-level extension actually possible: an extension point you can't reach, you can add. AI at the metadata layer — AI-generated extensions land in the same model, so they stay maintainable and evolvable. That's the road Oinone takes: 100% metadata-driven, front + back end open source, self-hostable — customization reaches L3. How to stress-tes