Ugh, sorry for the rant, but if anyone suggests that Linux on the desktop is "ready" -- it is not. Tried to spin up Ubuntu 26.04.1 and 26.10 on my Lenovo Thinkpad X1 Carbon Gen 13 and see how the "new" (in the works for how many years now?) TPM hardware support works. It is horribly broken: 1) The iwlwifi driver missing completey on the TPM kernel, so no wifi after install 2) there is some issue with Ubuntu's hardware root-of-trust evaluation and how it probes the Intel ME, which means you still have to use a PIN or passphrase to unlock the disk and 3) when you use the PIN, the Absolute Persistence Module has to be turned off, otherwise the PIN doesn't work and you have to manually enter the recovery code every time. github.com/canonical/secboot… github.com/canonical/secboot… discourse.ubuntu.com/t/the-p… Yes, I can live without the TPM kernel and hardware-backed encryption, but I shouldn't have to. These are features that have shipped and available in the installer that are just plain broken for a subset of commonly available, modern hardware. The blame is likely the intersection of Lenovo, Ubuntu, and Intel, but this stuff just works on Windows and I can get a working Linux environment with WSL2. I've been running Linux (along with FreeBSD, NetBSD, OpenBSD, Solaris, etc.) in some form on the desktop since around 2000, but there are still a lot of sharp edges. I moved past hacking XFree86 configs and recompiling kernels years ago -- it is just easier to use a Mac or run Windows. Again, sorry for the rant.

Was never really a Tailscale fan after trying it briefly. It is super convenient and easy to get going, but almost too easy. Lots of reinventing the wheel including its own SSH daemon that turns out to be roughly equivalent to distributed trust with ssh keys between nodes. Cloudflare's Zero Trust does a lot less of this and just focuses on being a good mesh with some added (but needed) complexity.

Opus 5.5 is surprisingly good at de-sloppifying code that its predecessors wrote.

I've had a PHP 8.4.26 honeypot up and running since Monday morning, and there still are no signs of RCEs. Just OOMs and worker crashes, but no RCEs, yet.

OK, the Claude Code Desktop WSL integration is pretty cool.

Might become a Claude Code Desktop convert. Going to be hard to break away from the CLI completely, though.

Gemini 4 is better than I thought it would be. We have some internal AIOps evals and Gemini 4 is the first Google model to pass one of the more difficult scenarios. Gemini 3.7 and 3.8 failed for various reasons and 3.8 specifically exhibited some misaligned behaviors.

This is definitely the future. Small teams delivering huge change. Anthropic wrote a whole new SCM and migrated its very large Git monorepo to it over the course of 5 months for the entire organization.

There's no way Muse Wonder is Opus 5.5 or Astra level. Looks like Meta benchmaxed another model.

Big IF true -- >Fable 5.1 & Opus 5.5 performance. Meta has a history of benchmaxing, though.

Had an O'Reilly learning subscription for close to 20 years (back when it was called Safari) and decided to cancel recently. Part of me thinks it is unfortunate, but there's the more pragmatic side that recognizes most of the new content I would read on it is likely AI-generated and that it would be more efficient just to converse with Claude or Codex anything I want to learn about. For select books where I respect the author, I'll continue to buy an ebook or a printed copy.

Wild perf numbers from VR