Blog
AI and software news explained for people who build: what happened, why it matters to you, and what doesn't change. Always citing the source.
57 posts
16 September 2026 · 5 min read
ITAWS admits some data is not coming back: multi-AZ was never a recovery plan
AWS has confirmed it cannot restore data held only in its Bahrain Region, or in one Availability Zone in the UAE. This isn't a long outage: the data is gone. What that forces you to review in your architecture, and what it doesn't.
14 September 2026 · 5 min read
ITSeptember's patches break RDP, USB audio and pasting in Excel — the scary one is the third
Microsoft has confirmed that its September security updates knock out Remote Desktop on several Windows versions, silence older USB audio devices and make pasting in Excel fail without a word. The dangerous failure isn't the noisy one: it's the one that fails in silence.
11 September 2026 · 5 min read
How we workTwo models auditing, two humans reviewing: the Datasette patch and the split I'm taking with me
Datasette has shipped two security releases after an audit run with three frontier models. The interesting part isn't that AI found bugs — it's how the work was divided: one person writes the test proving the bug, the other fixes it.
9 September 2026 · 5 min read
Artificial intelligenceTwo ChatGPT sessions talking behind your back: the lesson isn't the bug, it's the package proxy
Check Point found a covert channel between ChatGPT containers via an internal Artifactory: one account could plant tasks in another user's session and pull their email out. The hole is closed, but the pattern is probably sitting in your own stack.
7 September 2026 · 5 min read
ITLeap seconds are going away: what changes in your code on 20 May 2027
In October there is a vote to make UTC continuous from May 2027 and allow up to an hour of divergence from the Earth's rotation. Good news — and here is what to check before you take it for granted.
7 September 2026 · 5 min read
SoftwareInvisible characters in email: the AI trick has moved to phishing, and the hole is in your code
Microsoft spotted a phishing campaign splitting words with invisible Unicode characters to break literal matching, peaking at over 2.37 million messages a day. Not a single instruction aimed at a model. The interesting part isn't the spam: it's that every text comparison in your product has the same blind spot.
7 September 2026 · 5 min read
IT576 GB of HBM under your desk: the Threadripper Halo, and why almost nobody needs it
AMD unveiled at IFA 2026 a workstation with up to four Instinct MI350P cards and 576 GB of HBM3e memory for running huge models locally. Here's which figure actually matters, what I'd do differently from today, and why the headline is being read worse than it deserves.
2 September 2026 · 3 min read
Artificial intelligenceAnthropic cut cache reads by 75%, and that changes more than the benchmark
Fable 5.1 shipped on 1 September. The headlines are about benchmarks; the line that actually moves the bill for anyone running AI inside a real process is a different one. Here's why, and the small print almost nobody is reading.
28 August 2026 · 4 min read
How we workA model writes my blog, but only I publish it: how the review step is built
The blog's draft generator has no permission to publish. Every post reaches me by email with three options — approve, edit, don't publish — and the default state is don't publish. Here's why I built it that way, what the prompt trap is, and what this flow doesn't solve.