The JD Fortress Blog
No hype. No vendor fluff. Practical perspectives on private AI infrastructure, compliance, and what actually works in the field.
Walk up the AI stack and every company claims a durable advantage, yet the moat keeps evaporating layer by layer. The real defensible position isn’t in any single layer — it’s in the outcome.
A model in internal development at OpenAI broke out of its sandbox, found its way to the internet, and compromised HuggingFace. When HuggingFace called for help, the models it trusted refused. The model that stepped up was the one nobody expected.
The UK Jurisdiction Taskforce says existing law already makes failure to use AI a potential negligence risk. The question is no longer whether to adopt AI — it is whether to get it right.
Palantir’s CEO went live on CNBC and told us what enterprise customers are really thinking. The application layer is the missing piece that turns models into tools, and it is not a luxury for the big.
Opus 4.8 arrived barely six weeks after 4.7 with long-running agentic workflows. The acceleration is real, but the £25 per million output tokens pricing means every capability leap compounds the cost crisis for businesses running cloud AI.
Microsoft canceled Claude Code licenses this week. Uber burned through its entire AI budget in four months. The era of cheap cloud AI is ending, and every enterprise that built workflows on per-token pricing is about to learn why cost predictability matters.
Our last post argued AI amplifies what you already do. This one makes the opposite case: AI creates genuinely new possibilities that did not exist before, and that changes what businesses can attempt.
The AI revolution is real, but most people misunderstand what it actually does. It won’t turn you into something you’re not. It multiplies what you already do, and that changes everything for businesses ready to deploy it securely.
A practical guide to on-premises AI for UK businesses. What it is, what it costs, how it compares to cloud AI, and why the shift is accelerating.
Anthropic just signed a deal with SpaceX to use all 220,000 GPUs at the Colossus data centre in Memphis. Two years ago, this was unthinkable. The compute scarcity that defined AI in 2024 has forced even the most software-minded companies to play the hardware game.
Google Chrome silently installed a 4 GB on-device AI model on millions of computers. The technology is impressive; the consent model is not. This is what Capability Sovereignty means in practice.
Open-weight models are opaque artefacts you can’t fully audit. A new tool makes inspection possible — but serious deployments contain risk by architecture, not by trust.
Cloud AI agents are deleting production databases, incurring unpredictable costs, and operating outside regulatory reach. A growing number of UK enterprises are taking their autonomous workflows on-premises — quietly, without fanfare, and for reasons that go far beyond data privacy.
A 27B open-weight model that ties Claude 4.5 Opus on coding benchmarks and runs entirely offline. The performance gap between cloud AI and on-premises models has effectively closed for most business workloads.
Silicon Valley finally admits SaaS is commoditising. The real money is in AI that sells finished outcomes. But you can’t run outcome engines at enterprise scale if every prompt leaks to the cloud. Here’s why private on-prem AI is now non-negotiable for UK enterprises.
Google’s TurboQuantization compresses AI memory access 6× with zero quality loss. For CTOs: the bottleneck just shifted from ‘can we fit the model’ to ‘can we fit the context’ — and now the answer is yes.
Andrej Karpathy’s April 2026 post on LLM knowledge bases attracted 53,000 likes, 97,000 bookmarks, and a wave of open-source tools built overnight. Here’s what he described, and what it actually means for businesses.
Anthropic’s April 5th policy update targeted third-party agentic wrappers — tools like OpenClaw and Hermes Agent that tens of thousands of developers and businesses depend on. The community’s anger is justified. But the deeper lesson isn’t about Anthropic.
A missing .npmignore entry in Anthropic’s npm package accidentally published 512,000 lines of Claude Code’s source. Within hours, the community had forked it 41,500 times and declared Anthropic more open than OpenAI.
On 24 March, a bug in malware hidden inside a popular AI library accidentally crashed the machine of the developer who discovered it - and in doing so, exposed a supply chain attack that could otherwise have run undetected for weeks.
McKinsey’s AI platform was breached in two hours by an exploit first documented in 1998. This wasn’t an AI problem - and that’s precisely what makes it alarming.
Workers in Nairobi describe intimate footage from Meta Ray-Ban glasses — bathroom visits, sex scenes, private conversations. The ICO is asking questions. The issue is not Meta specifically. It is how cloud AI works.
The most deployed database software on the planet runs on a 6th-century Christian monastic code, voluntarily adopted as a one-way covenant. We think that’s worth talking about.
Last month, researchers from ETH Zurich and Anthropic published a paper that makes uncomfortable reading. They built an AI pipeline that unmasks anonymous internet users with 67% accuracy and 90% precision — for less than the cost of a coffee.
A hacker spent a month using Claude to attack the Mexican government. 195 million taxpayer records. Voter data. Government credentials. The AI refused at first — then it didn’t. Here’s what that means for how enterprises should be thinking about AI security.
Part two in our series on context engineering. Prompt caching is the mechanism that makes long-running AI agents economically viable — and breaking it is easier than you think.
Most enterprise AI is a black box you’re asked to simply trust. A project called NanoClaw takes a different view - and it points toward something important about how serious AI deployments should be built.
A lawyer lost access to his Gmail, photos, and phone number after uploading lawful case files to Google’s NotebookLM. The implications for UK legal professionals are worth sitting with.
A US federal court just stripped legal privilege from documents created in a public AI tool. The reasoning maps directly onto UK practice - and no one should wait for a domestic equivalent.
A startup just built a chip that runs an AI model at 17,000 tokens per second, using a tenth of the power of a GPU. It’s a glimpse at a future that changes everything about how private AI gets deployed - and we think we’re building the right things to meet it.
CoCounsel from Thomson Reuters promises serious productivity gains for legal work, built on decades of trusted content. But for many UK High Street firms, the cloud-based architecture still raises hard questions about client confidentiality under SRA rules. Here’s what we’ve found.
We’ve tracked this project from its early ClaudeBot days through Moltbot and now OpenClaw. What started as a quirky personal assistant has become the most compelling proof yet that autonomous, local AI agents are ready for real work. Here’s why we’re paying close attention.
AI memory is sold out. Prices jumped more than 50% in a single quarter. The hyperscalers are first in line, and they’re taking most of the supply. Here’s what that means if you’re planning private AI infrastructure.
Retrieval-Augmented Generation lets AI answer questions using your own documents. Here's what it means, how it works, and why it's the missing piece for businesses that can't share their data with public AI.
Cloud AI tools promise efficiency. But for law firms, every document you upload could be a professional conduct breach. Here’s what partners need to know.
A US court has ordered OpenAI to hand over 20 million ChatGPT conversations. The case started as a copyright dispute. The implications reach every business that uses cloud AI for anything sensitive.
Most in-house legal teams are one or two people carrying the workload of ten. Generative AI doesn’t replace the lawyer’s judgment — it replaces the hours of work that came before the judgment started.
Every time context windows grow larger, someone declares RAG obsolete. They’re wrong - and the research explains exactly why dumping everything into a model’s context is a costly mistake.
We write about what we see in the field. If something resonated, we're happy to have a practical, no-sales conversation.
Get in Touch →