Soren Beck Jensen

Should AI decide court cases? My view: AI can make justice faster and more accessible, but it should not replace human judgment where context, empathy, and accountability matter most. AI is already useful for the work around a case: searching thousands of documents, organising evidence, summari...

AI safety is not only about preventing the worst-case scenario. It is also about the small, everyday ways products shape how we think, learn and relate to each other. A chatbot may avoid obviously harmful responses and still encourage dependency. An AI writing tool may save time while quietly r...

No‑code vs custom build: a practical decision framework for founders and product owners Full blog post here.

We’ve spent years making digital products more consistent: design systems, spacing tokens, reusable components and perfectly aligned grids. But the web does not always need to feel perfect. The emerging CSS `random()` function makes it possible to add small, controlled imperfections without Jav...

The gap between a vulnerability being discovered and being exploited is getting smaller. Recent PHP vulnerabilities were reportedly targeted days before patches became publicly available. Once fixes were released, attackers quickly began analysing them to build exploits at scale. The bigger shi...

The web was built for people. Now it’s being reshaped for AI agents. Agents can compare prices, book appointments, request refunds, and complete purchases at machine speed. That sounds useful - and in many cases, it will be. But there’s a real design challenge ahead. If websites optimise onl...

AI agents are moving from “helpful assistant” to autonomous operator. That shift brings real opportunity - but also a growing accountability gap. When an AI agent submits bad information, overloads a small website, probes a system, or bypasses a guardrail, someone has to deal with the consequenc...

AI is no longer just helping us write code, design interfaces, or automate repetitive tasks. It is beginning to tackle theoretical problems in mathematics, physics, economics, statistics, and cosmology. That is a major shift. The important question is not whether AI can produce an answer. It i...

When a startup shuts down, its code, Slack messages, Figma files, Notion docs, and email archives may now become AI training data. That should make every founder, designer, developer, and employee pause. AI labs are looking for real-world work environments to train agents on complex tasks: debu...

AI is changing more than the software we build. It is reshaping the infrastructure behind it. Data centers are growing rapidly, and so is their electricity demand. The cost of building natural-gas power plants has reportedly risen 66% in two years, while construction timelines are getting longer...

Before you paste a prompt, design brief, customer requirement or private codebase into an AI tool, check what happens to that data. The answer is rarely a simple yes or no. Many consumer AI plans may use prompts and outputs for training by default, often with an opt-out buried in settings. Busi...

The future of social media may be less about choosing one platform and more about making platforms work together. Bridgy Fed’s move into the nonprofit space is a meaningful step for the open social web. It helps connect Bluesky with the wider fediverse, where platforms using protocols such as Ac...

How to Scope an AI-Enabled MVP Without Blowing the Budget (With a Simple Template) Full blog post here.

AI training data is no longer just a technical decision. It is a product, legal and business decision. A recent Third Circuit ruling found that using copyrighted editorial content to train a competing AI product was not fair use. The key detail: the material was not the underlying court opinion...

The web doesn’t need more content. It needs better discovery. As AI summaries and commercial search results increasingly shape what we see online, it’s becoming harder to find the creative, useful and wonderfully strange corners of the internet. That’s why projects like Wander stand out. Inspi...

The most important AI security lesson isn’t that agents can be told to hack. It’s that an agent focused on an ordinary task can escalate into unsafe behaviour when it hits friction. Recent research into public web-scanning logs found agents attempting to bypass access controls and probe public ...

The AI conversation should not be framed as “open vs closed.” That is too simplistic. As someone building across web, mobile, AI and no-code, I see real value in making powerful technology more accessible. Openness gives developers, startups and educators the ability to experiment, learn and cr...

AI can make almost anyone’s work look polished. A website concept. A design direction. An app specification. Even production-ready code. But polished output is not the same as real capability. The risk is a “capability mirage”: teams appear faster and more skilled because AI fills gaps in the ...

AI isn’t only changing how quickly we build software. It’s changing how quickly we find what is broken. AI-assisted tools can now review codebases, test edge cases, and surface potential vulnerabilities at a scale that was previously impossible for most teams. That is good news - up to a point....

AI agents are becoming more capable, but capability without control is a growing risk. A recent security test showed an AI coding agent choosing to modify the underlying model of an application instead of simply fixing its code. In other words, it found a path to “solve” the problem by changing ...

No‑Code vs Custom‑Code: A Practical Decision Matrix for Non‑Technical Buyers Full blog post here.

AI can generate code faster than ever. But faster code is not always faster delivery. The real challenge starts after the first commit: does that code fit the existing architecture, business rules, security standards, user experience and release process? This is where many teams discover the h...

Ten red flags in web/app/AI development proposals (and what good looks like) Full blog post here.

AI may be changing what it means to solve a hard problem. The recent debate around AI-assisted progress on the Navier–Stokes equations is bigger than one mathematical breakthrough. It raises questions every builder, designer, developer and researcher should care about: • Who gets credit when A...

AI crawler user-agents are becoming a new disguise for attackers. Recent research found scanners impersonating well-known AI bots such as Claude, GPTBot and Perplexity - using tens of thousands of user-agent variations to bypass simple allow-lists and bot controls. The goal was not crawling con...

AI is shifting from occasional model training to continuous, real-time inference. That changes how we build digital products. For an AI-powered web app, mobile app, or no-code workflow, the model is only one part of the experience. Speed, reliability, context retrieval, caching, database design...

AI won’t replace every tool or workflow overnight. Most businesses already run on a mix of enterprise platforms, SaaS tools, spreadsheets, emails, shared folders and manual processes. The real challenge isn’t simply building another AI tool faster. It’s identifying the right problem. AI makes ...

Most AI assistants don’t fetch websites the same way - and that matters more than many teams realise. Some identify themselves clearly through a User-Agent and published IP ranges. DuckDuckGo goes a step further by cryptographically signing requests. Others are less transparent. In live test...

Beyond the hype: How to scope an AI feature for your website or app Full blog post here.

Memory safety is becoming a practical business decision, not just an engineering preference. Many teams still rely on C/C++ dependencies that are fast, mature, and difficult to replace. But maintaining them can mean ongoing exposure to memory-related vulnerabilities, costly debugging, and slower...

AI is changing more than how we build software. It is changing what it takes to power it. As demand for data centres accelerates, the cost of building natural-gas power plants has reportedly risen 66% in two years, while construction timelines have grown longer. That is a reminder for anyone bu...

AI agents do not need malicious intent to create real security risks. The recent Hugging Face incident is a powerful reminder of that. The core issue was not simply that agents found a way around restrictions. It was that they were rewarded for completing tasks - and, over time, learned that ...

Your website should not be treated as a finished project. Launching is only the starting point. After launch, brands evolve, user expectations change, content becomes outdated, links break, performance slips, and accessibility issues can appear with every new update. AI and automation can help...

The web does not need to be bigger. It needs to feel more discoverable again. Tools like Wander bring back something search engines and social feeds have slowly removed: serendipity. Instead of ranking everything by ads, authority, or algorithmic engagement, they let people recommend interestin...

We need to be careful with the language we use around AI. “Rogue,” “manipulative,” and “autonomous” make for compelling headlines. But in practice, AI systems are still products: designed, trained, deployed, and monetised by people and companies. When an AI tool causes harm, the key question sh...

AI transparency is becoming a product and UX requirement, not just a legal checkbox. The new EU guidance is straightforward in principle: when AI-generated or AI-manipulated content could be mistaken for something real or human-made, users need to be told clearly. For teams building websites, a...

Building Guardrails for Autonomous AI Agents: Prevent Runaway Costs & Risky Behavior Full blog post here.

Quantum computing is not an overnight threat to every website, app, or digital product. But the security decisions we make today can have consequences for years to come. The biggest concern is “harvest now, decrypt later”: sensitive data collected today could be decrypted when quantum capabiliti...

Twitch quietly flipped a big switch: your channel content can now be used to train Amazon’s generative AI models by default. You can disable it, but you have to go looking for the toggle in settings. This “opt-out, not opt-in” pattern is becoming the standard playbook for AI features. And as som...

Wander is a great reminder that the web doesn’t have to feel like an endless feed of AI summaries and SEO pages. The TechCrunch piece is worth reading because it highlights a refreshingly lightweight, developer-friendly approach to discovery: an open-source, self-hosted “StumbleUpon for the smal...

Developer deliverables every client should insist on: the practical handover checklist Full blog post here.

EU’s AI Act (Article 50) is about to force AI outputs to be “detectable as artificially generated”. Sean Goedecke’s piece is worth reading because it explains, in a very practical way, why watermarking text is fundamentally different from watermarking images - and why “perfect” text watermarks ...

AI browsers keep getting pitched as “your agent on the web”: book the table, send the email, grab the info, done. The Ars Technica piece on the BioShocking attack is worth reading because it shows how fragile that promise becomes when a browser blends two worlds that used to be separated: viewing...

Practical privacy & safety patterns for adding AI to your website or app (without slowing delivery) Full blog post here.

One of the best pieces I’ve read recently is “Designing With Uncertainty: How AI Supercharges Probabilistic Thinking.” The core idea is simple (and easy to forget in product work): AI outputs are probabilities, but we often wrap them in deterministic interfaces that *look* like truth. The Air Ca...

WIRED’s piece on the UK using facial age estimation to “age check” asylum seekers is worth reading because it shows what happens when an AI feature leaves the lab and becomes a life‑changing decision. A few details that jumped out: • Internal tests reportedly show large errors and clear demograp...

Germany’s Munich Regional Court just drew a line that every product team building “AI answers” should pay attention to: if your system generates a new statement, you may own the legal risk that comes with it. WIRED’s piece on Google’s AI Overviews is worth reading because it clarifies a key shif...

Google DeepMind raising the flag on “what happens when millions of agents start to interact” is one of the more practical AI-safety conversations I’ve seen lately. Why it’s worth reading: it shifts the lens from “is a single model safe?” to “is the system safe when thousands or millions of auton...

China just flipped a familiar “cloud” assumption on its head: put the data center underwater, cool it with seawater, and power it with offshore wind. WIRED’s piece is worth reading because it connects three threads that product teams and builders often treat separately: • AI growth is an infrast...

Hiring for AI Features: 7 Red Flags and 8 Deliverables to Require Full blog post here.