Oct 26^th 2026 · 3 min read · #ai #anthropic #finops #llm #openai #piclaw #pricing

The Price of a Token

OpenAI launched GPT-6.1 Terra today at $1 per million input tokens, $5 for output and $0.05 for cached input. That is pretty similar to what happened with 6.1-Sol, since it is half of what GPT-5.6 Terra costs now, and the cached input rate is a fifth of what my piclaw pricing table recorded for it back in July.

This is kind of interesting to me for a number of reasons, not the least of which is that I have been tracking these numbers for months both at work and as I use my OpenAI subscription–my agents burn through a fair chunk of tokens, so anything that makes them much cheaper is noteworthy.

Following the numbers

In case you’re new here, piclaw keeps a dated pricing reference for its token chart, plus a set of finops snapshots with the provider pages they were taken from. Looking at the git history for a few model lines (USD per million tokens, input/output, then cache reads), the trend is pretty interesting:

Line	Earlier	Now
Claude Opus	4/4.1: $15/$75, $1.50	5.5: $4/$20, $0.20
Claude Sonnet	4.5/4.6: $3/$15, $0.30	5.5: $2/$10, $0.10
OpenAI “Sol”	5.6 (July table): $5/$30, $0.50	6.1: $2/$10, $0.10
OpenAI “Terra”	5.6 (July table): $2.50/$15, $0.25	6.1: $1/$5, $0.05

In short, headline prices are down by anywhere from a third to three quarters depending on which line you pick, which is nice but not the interesting part–the interesting part is the last column, and, if you’re doing any sort of agentic work at scale, how caching prices factor into the equation.

Cache Impact

When I audited a week of usage on my main instance in September, it came to 2.17 billion tokens. 94.3% of those were cache reads, which is entirely by design since I have been aggressively optimising piclaw’s context structure (and compaction) to maximise caching to lower costs, but still left a bit of fake money on the table.

Since 90% of my workloads are development/agent loops, fresh input on each turn is typically 1.5% and output a measly 0.1% (i.e. one or more tool calls). Agents just ship the same long context over and over to providers (to a point where I have to wonder why the heck we still use the same payload structure…), and 93.7% of the volume was in requests over 272K tokens, which OpenAI bills at twice the input and cache rates.

What this means, in volume, is that the price that matters for agents is actually cache reads, not raw tokens. If you’re burning money to light up your cigarettes (and have a disgusting smoking habit), you probably don’t care that Claude Fable 5.1 and GPT-6 Astra have the same $10/$50 list price, but over time Fable 5.1 cut cache reads from $1 to $0.25, which made it 70% cheaper than Astra on the same replay ($2,064 versus $6,858).

I don’t use Fable (I don’t have access to it anymore, even at work, and Anthropic doesn’t give me the time of day as an individual open source developer), but running my old benchmark’s token counts against today’s prices is… kind of interesting, and makes me really happy I am not paying per token:

Model	Input/output	Cache read	One week, same workload
GPT-6 Astra	$10/$50	$1.00	$6,858
Claude Fable 5.1	$10/$50	$0.25	$2,064
Claude Opus 5.5	$4/$20	$0.20	~$1,030
GPT-6.1 Sol	$2/$10	$0.10	~$990
Claude Sonnet 5.5	$2/$10	$0.10	~$515
GPT-6.1 Terra	$1/$5	$0.05	~$490

Of course, the above assumes a full replay: identical tokens as over that week, identical cache behaviour, etc. I’m glossing over retries, reasoning effort and whether the model actually gets the job done, so it’s just a data point.

I do find it interesting that Sonnet and Sol come out differently despite identical list prices because of OpenAI’s long-context surcharge, which Anthropic doesn’t charge–but, again, subscriptions mask all of this under their flat rate.

What OpenAI is claiming

OpenAI’s announcement says Terra matches Astra on Agents’ Last Exam (a test of agents using office software, which I’ve always found hilarious given there’s so much more useful stuff to test them on) “at one-tenth the price per token”, and that it finishes tasks over three times faster than both Astra and Sol.

There are also claims that Terra is over three times as token-efficient as Opus 5.5 at maximum effort, which would make it about nine times cheaper per task–likely more because Terra would also be using fewer tokens.

Now, I’m much more focused on actual results, and even if I don’t care anything about synthetic benchmarks like BenchLM, they are less flattering on stuff I care about:

Terminal-Bench 4.0: Terra scores 50.3%, against 70.6% for Sonnet 5.5.
OSWorld 2.1 (offline subset): Terra 74.2%, Opus 5.5 87.2%.
OpenAI itself still points people at Sol or Astra for “in-depth research or complex coding”.

Still, if you consider that the smart approach is to have Luna and Terra carry out pre-defined instructions drafted by Sol or Astra and use extensive standard automation, Terra is starting to look really good for churning through things like ticket triage, running test harnesses and even managing CI.

But why does Anthropic keep charging premium prices?

This is the bit that puzzles me. If Sol and Terra now get you most of the way to… well, anywhere, really, for a fraction of the price, why does Anthropic still sell Fable (and Mythos) at $10/$50?

I have a few theories, ranging from good marketing, lack of competitive pressure at the premium level, and the current shape of the market:

They are not alone. Astra costs exactly what Fable does, so OpenAI also keeps a premium tier. It’s just not the one it advertises, which is an interesting inversion of Anthropic’s marketing spiel.
The mid-range has already converged. Sneakily (well, quietly, at least) Sonnet 5.5 now has the same list price as GPT-6.1 Sol, down to the $0.10 cache reads, and Opus 5.5 is at $4/$20, so the price war is happening one tier down (maybe at the level Claude users actually set most of their agents to).
They nominally still win where it matters for coding agents. My results with Opus have been consistently lousy across the board where it regards code quality (and the models outright lying to me regarding implementation), but market perception seems to be completely opposite to mine.
Most people never see a token price. Claude Max, Codex and Copilot sell allowances, not tokens, and a flat $200 a month both hides most of the difference unless you hit the limits and is psychologically justifiable as being “best of breed” (even if it isn’t).

What I do know is that, on my own workload, the cheapest Anthropic model I’d trust with real work (Sonnet 5.5) now costs about the same as Terra, and Opus 5.5 costs half of what Fable does. So Anthropic is, surprisingly, cutting prices too.

Still, I’ll be running Terra against my usual piclaw tasks over the next few days, mostly to see whether the token-efficiency claims hold up. My guess is that it’ll replace a lot of my bounded, well-tested work and none of my hard debugging, planning or code audits, but we’ll see.

Oct 26^th 2026 · 1 min read · #apple #hardware #ipad #tablets

The New iPad mini ➹

I miss using my iPad Mini 5, but ever since iOS 27 came out it’s been massively slower and I’ve mostly switched to my M1 iPad Pro (which also makes a few other things simpler, like travel, note taking, using external displays, etc.).

I very dearly want to go back to the Mini since the Pro is too large as a permanent companion device, but the inflated pricing is not helping. and neither is my feeling that Apple has likely nerfed the hardware in ways that matter to me.

As an example, there’s a sidenote in the announcement where Apple promises extended display mode with iPadOS 27.2. But I am going to wait until detailed reviews show how well that actually works, since the Pro can drive my Kuycon P20 at full resolution, and if the new Mini doesn’t do at least as well, then I will likely start hunting for ~~unicorns~~ good Android tablets.

Oct 26^th 2026 · 1 min read · #ai #art #music #science-fiction #video

Claudia ➹

Here’s a fun rabbit hole I fell into over the weekend, and that should make you wonder if William Gibson is a visitor from a future timeline–because Claudia not only echoes Idoru to a great degree but also might be somewhat of an inflection point. Just not necessarily because of the base content on the site: this pretty amazing French “slop core” video has been making the rounds in socials, and even considering it’s AI-generated, I don’t think you can just ignore it without appreciating the creativity its human director put into it.

Yep, buckle up!

Oct 25^th 2026 · 1 min read · #agents #ai #wellbeing #work

The Quiet

I was watching Gergely Orosz the other day for some reason, and he was going on about the impact of AI on the industry (what I like to call the Cambrian Slop Explosion). He mentioned that one of the major changes–and one of the personally taxing effects of using AI–was constant context switching and how fatiguing it was.

And I completely agree. Nearly two months after returning from vacation and after a particularly nasty overlap of work and personal projects, the thing I missed the most was the quiet. To just sit in a room reading or writing with no pressure to check on my agents (work or otherwise, which also feels like work now), letting ideas come to me unbidden, in a Douglas Adams-esque parody, wafting in the wide open space somewhere between my ears and bumping into each other.

It’s kind of nice, the quiet.

Oct 25^th 2026 · 4 min read · #reviews #teams #ugreen #videoconferencing #webcams

The UGREEN FineCam Pro

It’s been a while since I wrote about my desk setups or how much I rely on having decent audio and a good camera with face tracking, so my return to work in early September seemed like a good time to do that. I’d meant to post this then, but said work was enough of a pain to delay it until late October… But that’s beside the point right now.

Apple users are used to face tracking and auto-framing by now, but I have been using professional videoconferencing systems for many decades. In fact, back in the mid-90s, I did quite a bit of work on videoconferencing over ISDN and ATM, including hacking NetMeeting into submission… But it took the pandemic and my OBSBOT camera to finally feel personal tech had caught up with the times.

After another bout of back pain this year and a return to my makeshift standing desk, I decided I’d take a look at the webcam market again to see if I could get another OBSBOT (or at least a true gimbal camera) to sit atop my battered Surface Pro 3, which was (until recently) the centrepiece of said “desk”. It has a perfectly usable Windows Hello camera, but I tend to get restless and something that keeps me in frame is nicer to my callers.

I considered the software route first, but I couldn’t find anything that would work with the internal camera and wouldn’t kill the PC (Teams is enough of a challenge for it these days). Trying one of the (excellent) Logitech Brio 4K cameras I still have made me realise I needed to move the camera itself, too. And I didn’t want anything that would require extra software in Linux or on a Mac later.

There’s enough junk software out there already, after all.

But OBSBOT-style cameras were around €200 a pop when I looked, so getting a new one for a secondary desk was a no-go, and I decided to try something cheaper, gimbal-less, but with enough sensor resolution for decent cropping and zooming.

The UGREEN FineCam Pro stood out to me for two reasons–what seemed a decent price for a nominally 4K-capable sensor and its compactness, since it took up far less volume than most other options.

Disclaimer: I bought this one myself; however, my usual review policy applies, because I consider this a bit of a “low effort” review.

The reason I bought it in April was that the built-in firmware handles gesture-controlled framing and zoom itself–with driverless support for Windows, macOS and Linux, and thus complete software independence (which I think will be useful for anyone in alternative platforms).

UGREEN FineCam Pro with its mode dial, privacy cover and monitor mount — The FineCam Pro, sporting the perennial dust layer that pervades my office

And why finally write about this now? Well, because I’m testing a brand new Qualcomm-based laptop that does head tracking completely in software, and thought this was worth noting while comparing setups–more on that in a future post about the laptop.

Hardware

The camera itself is… OK, I guess. It’s an 8-megapixel sensor capable of 4K at 30fps or 1080p at 60fps, has “phase-detection autofocus” and pretty decent lighting sensitivity, as well as two microphones with noise reduction, positioned for stereo capture if you use it for vertical video, which… I didn’t. Audio from those was reported as being good, though.

The rest is mostly creator frills: the mode dial switches between seven dubiously useful built-in filters, and there is a privacy cover.

The camera comes with a magnetic mount and a small tripod, which is another concession to the “creator” audience, and the magnetic holder fits neatly atop my Surface display and lets me detach and reposition the camera without disturbing the mount (it also made it easy to move the camera off-centre when using a sidecar monitor, which is a very “me” thing to do, but which might be of interest to other people as well).

Annoyances

For some stupid reason (yes, I know it’s probably the creator audience, but it’s still stupid) the camera defaults to vertical orientation, which I suppose comes hand in hand with its mode dial and effects.

That dial is clearly geared at people doing social media, but gleefully ignores the fact that even on TikTok (which I don’t use but am regularly kept apprised of) nobody seems to use those effects, and, as you’d expect, I didn’t use them either (except once by mistake).

Additionally, the USB-C port seems to have been purposefully positioned at the most inconvenient position possible on the back–the weight of the cable tore off the camera from the magnetic mount and the cable protrudes from the back too much, adding to the imbalance.

Teams Usage

It took me a little while to figure out the right distance at which to hold my hand to enable tracking, but after a while that became second nature (I’m used to having to do that with my OBSBOT as well, because for some reason there is no way to set that default even after years of firmware upgrades).

Picture quality in Teams calls was good, with smooth tracking–but of a different flavour than the OBSBOT. The differences in quality and motion are subtle, and in practice I’m mostly the one who notices.

But here’s a pro tip: perform the magical tracking gesture incantations before you enter the call, if you can, otherwise people will feel compelled to wave back, even in very formal official settings…

Verdict

I don’t regret my purchase, but I do think it would be pretty safe for UGREEN to simply remove the mode dial and move the USB port to the side. I’d much rather have a smaller camera than a selection of filters I never use.

I haven’t tried the plainer FineCam, but its shape is pretty much the form factor I’d want. Put the Pro’s camera hardware and tracking into that enclosure, and I’d be perfectly happy not having to see the dial…

Oct 24^th 2026 · 1 min read · #agents #ios #notes #weekly #wifi

Notes for October 18-24

I’ve been more interested in exploring how to simultaneously enjoy Greek yoghurt, orange marmalade and the glorious golden-hour light that pervades my living room than, well, most other things, so this week was a bit of a bust as far as personal projects go.

Part of that is because I have finally swapped the Surface loaner I was using for a brand new machine (more on that later) and am also returning the Lenovo Chromebook I recently reviewed, so I had to spend a little while making sure nothing of relevance was lost.

The rest was… mostly housekeeping, really.

Wi-Fi Tweaks

A sharp-eyed reader (hi!) noticed that I mentioned the Radxa Linkr had some WPA3 issues, and told me an OpenWrt update had some WPA3 interoperability fixes for the Cudy M3000, so I upgraded one of my APs–that didn’t fix the Linkr, but after some soaking in I upgraded the rest with no ill effects.

Roaming around the house still works fine, but there are a couple of reported issues with fast-transition (FT) associations in this build, so I’m keeping an eye on OpenWrt PR #23181 and mt76 issue #1098 in case I come across any issues in the future.

Local Agents

I’ve found an interesting workaround for my “no local agents on my Mac” rule–I’ve been using Open Minis on my iPad for a while now because of its iOS integrations (it can handle Reminders, Calendar, HomeKit, etc.), and I realised that it can run as a fully sandboxed iOS app on my MacBook Pro and sync via iCloud with my iPad and iPhone.

Since I had previously gotten my iPad to sync skills with memento, it took me 5 minutes to get everything working on my Mac, including mapping a couple of “external” folders to it.

Since it’s using the same engine as ios-linuxkit (even if they didn’t take any of my PRs and fixes, which is a bit sad), there’s additional ARM-on-ARM emulation going on, but on a MacBook that barely registers–and having a sandboxed agent that can run its own native Safari WebKit browser completely isolated from my Mac is pretty nice.

The Grind

I’ve been patiently nudging piclaw to port itself to pi-durable. Most of that has already happened–the new version is self-healing and is gradually removing the redundant abstractions I had built atop the original SDK.

No hurry there.

Oct 22^nd 2026 · 1 min read · #apple #foldables #iphone #reviews

The iPhone Duo ➹

I’ve been bemusedly reading and watching several iPhone Duo reviews over the past day or so (besides John’s piece, The Verge has a series of excellent pieces and a thorough video review), and the consensus seems to be that it’s a polished, chunky prototype that hasn’t really figured out what the user experience should be. Being a Mr. Mobile fan and watching his videos on Android foldables for years now, it comes as no surprise. But it also will come as no surprise that two years down the road this thing will have solved most of its teething issues–except the price, of course.

And I do find it amusing that we’ve managed to get to a point where a premium phone actually makes you want to use it less as a phone. I, personally, count that as a good thing.

Oct 21^st 2026 · 1 min read · #hardware #kits #mac #macclock #raspberry-pi #retrocomputing

Macintosh Mini ➹

The Maclock hack I linked to many months ago is now a finished product (with various flavours of kit), which is kind of fun to see. I have managed to procrastinate assembling mine for a stupefying amount of time now, and have some hope that posting this will shame me into finishing it (I actually have two Maclocks, plus my own 3D printed SE/30)…

Oct 18^th 2026 · 7 min read · #agentic #ai #hardware #kvm #networking #reviews

The Radxa Linkr

I’ve done quite a few KVM reviews in the past, and there’s a reason for that: All of the homelab and industrial IoT stuff I test needs some sort of physical access for setup, and even if I do have monitors with multiple inputs, USB video capture dongles, etc., it gets really old to have to constantly re-wire stuff, switch between keyboards, etc.

Additionally, most of the newfound range of cheap/compact KVMs that have popped up over the past two years still need an Ethernet connection, and it’s become ridiculously hard to have even semi-permanent test setups–I keep running out of both power and Ethernet ports, and that’s been the reason why I’ve been limping along on Wi-Fi based Pi Zero KVMs as well as one or two KVMs that, sadly, require proprietary apps.

So when I spotted the Radxa Linkr, I just had to try it. And that was even before I knew I could get it to work with an AI agent…

Disclaimer: Radxa supplied me with a Linkr free of charge, and this review follows my usual review policy.

Hardware

Radxa Linkr and its supplied cables — The Linkr and its supplied cables.

In the box, you get the Radxa Linkr and a set of cables (pictured above in real life, and below in the Quick Start guide).

Radxa Linkr Quick Start guide showing its ports, accessories and connection options — The supplied Quick Start guide (image: Radxa).

The hardware itself is pretty comprehensive, and packs quite a lot into the small 70 x 22 x 12 mm enclosure:

Rockchip RV1106G3 SoC, with a 1.2GHz Arm Cortex-A7 CPU, a RISC-V co-processor, 1 TOPS NPU and a hardware video encoder.
256MB of integrated DDR3L memory.
16GB eMMC storage or a microSD card slot (I got the variant with internal storage) for firmware and virtual media.
HDMI input up to 2560x1440 at 30Hz, with lower-resolution modes including 1920x1080 at 60Hz and custom EDID support.
Dual-band 2.4GHz/5GHz Wi-Fi 6 using an AIC8800 module, with client and hotspot modes.
A USB-C device port for keyboard and mouse emulation, power (5V/2A) and firmware flashing.
A separate USB-C network port for a direct host connection or 10/100Mbps Ethernet through the supplied USB-C to RJ45 cable.
One FN user button, a system status LED and an HDMI/video status LED.

The device itself is tiny, and the single FN button is used to identify it on the discovery page or (with newer firmware) to toggle the Wi-Fi hotspot with a double press. You can see the USB-C connectors on the back and side through the semi-transparent casing in the photo above.

A nice touch that you will appreciate if you manage many machines is that there are indicator lights that show when someone is remotely connected (the blue power LED will turn green) or there is a video input problem (signalled by a blinking video status LED).

Connectivity and Video Quality

Besides direct USB connectivity to your machine (which I honestly didn’t spend any time on), the Linkr supports both 2.4GHz and 5GHz Wi-Fi. However, it doesn’t seem to support my particular flavour of WPA3 (at least in current firmware versions), so I stuck to 2.4GHz/WPA2 for all of my testing, which is plenty fast enough.

That’s a decent tradeoff for me (I had zero latency issues, although I spend so much time in VDI and remote console setups that I don’t notice delays anymore) and, at least in one of my tests, was the only way I could connect to a machine in a remote corner of the house where 5GHz has very little reach–but it managed 60fps just fine.

The Linkr is quite flexible in terms of video capture. I’ve needed to hack in support for custom EDIDs in my old PiKVMs, so it’s nice to see that is baked in, and the built-in HDMI bridge goes from 1280x720@30/60Hz to 2560x1440@30Hz, with standard 1080p in the middle.

I also found it a nice touch that the Linkr includes a virtual sound card that it exposes via USB to the target system, although I have had zero use for audio in any of my tests…

Thermals

The Linkr, like all devices, does warm up while in use, and its metal underside can be surprisingly warm to the touch–and yet, my IR thermometer didn’t see it get past 45°C even after hours of continuous use (disconnecting, of course, lets it idle cooler).

Setup and UX

As you’ve probably divined from the Quick Start shot above, this is a pretty trivial web-based setup: go to linkr.now (with Chrome or Edge, because Apple still hasn’t figured out most of the modern device-related browser APIs), and you get an HTML app that will find the Linkr and help you connect to it – via USB or, later, via Wi-Fi.

The entire thing is very slick, and the user interface is both responsive and easy to navigate–here’s a little slideshow of the original firmware version:

Radxa Linkr web interface slideshow — A tour of the web interface captured with firmware 1.4.2.

Software

Firmware updates and release notes are available on GitHub, and since I got mine (v1.4.2), Radxa has added quite a few new features, as well as a bunch of fixes, in what is today v1.5.2:

A Plugin Center, which is a sensible way to keep adding more features without changing the core firmware–I haven’t yet had time to investigate how I could build my own, since I only noticed it as I finished this draft.
ZeroTier and EasyTier support (I use Tailscale myself, but it’s nice to see options).
Screenshot OCR to recognise text in screenshots and edit or copy the results (see below); recognition runs locally, and an optional plugin supplies the OCR resources without Internet access.
Experimental IPv6 support (which works, at least on my LAN).
FN-button hotspot toggle (double-press it to turn the Wi-Fi hotspot on or off).

One fix in particular was very useful already (they fixed resolution switching with the Radxa Dragon Q8B as video input, and I’m still testing that this weekend), and the pace at which they are currently shipping updates is reassuring, even though I couldn’t find published source code for the firmware in the public repository.

Other than that, you get all the typical remote KVM features (as well as Tailscale support, which I use extensively):

Video resolution settings (and custom EDID, as pointed out above).
Wake-on-LAN (for any MAC address, although you can pin the ones you need).
Keyboard and mouse control (as well as the virtual USB audio capture, all exposed via USB ID 1d6b:0104 - Multifunction Composite Gadget).

Mouse control can be set to both absolute and relative, as well as reversing the scroll wheel (which I thought was a very nice touch).

However, Mac and iOS users can’t really use some of the more advanced web UI features like screenshots, recordings or screen OCR, because they’re not available in Safari. Some of this is Apple’s problem, but I’ve hacked Safari-compatible screenshot support into other KVMs, so Radxa can probably improve that.

I did find the screenshot OCR feature interesting, but it requires pretty crisp text, and some SBCs just aren’t willing to play ball on their boot consoles–however, it can be immediately useful in a pinch:

Linkr screenshot OCR showing boot-console text and the recognition results — Screenshot OCR on a boot console; the recognition results can be edited and copied.

A nice touch about this is that recognition happens fully in-browser:

It uses tesseract.js to recognise English and Simplified Chinese.
By default, it downloads OCR resources over the Internet, including the worker script from cdn.jsdelivr.net.
The screenshot is processed locally by the browser worker, which returns the recognised text and bounding boxes to the UI.

The offline plugin just caches a local copy of the worker, engine and language data inside the Linkr itself, removing the external-download requirement.

APIs and AI Agent Control

But the real reason I am very much enjoying using the Linkr is that, unlike most KVMs, it actually exposes an API that you can either script against or just point your AI agent to.

I have been using that for over a month now to great effect (including doing a couple of Fedora and Windows installs, which is a veritable quagmire of edge cases) and built a piclaw add-on you can use as a reference.

The release notes also describe new API controls for resolution and video quality, which my add-on doesn’t cover yet. But the basic screenshot and input controls work amazingly well if you pair them with a vision-enabled model, and are really simple:

The documented public API uses an access token in the Authorization: token <token> header, and boils down to two endpoints:

GET /api/public/snapshot returns the current screen as a JPEG.
POST /api/public/control accepts an ordered events array containing keyboard, mouse, text and delay operations.

The control primitives are small enough to wrap directly in agent tools:

keyboard presses or releases keys using browser KeyboardEvent.code names, so combinations are sequences of key-down and key-up events.
mouse_abs positions the pointer using normalised coordinates between 0 and 1 (this can be a little tricky for some models, but they get it eventually).
mouse_rel moves it by pixel offsets (both support mouse buttons and scrolling).
text types up to 1024 characters, limited to printable ASCII, tabs and newlines.
delay pauses between events for a specified number of milliseconds (useful when the target needs time to react).

That gives an agent the usual screenshot/action loop without installing anything on the target machine.

I had to burn through quite a few tokens in initial testing (and the skills my add-on bundles are the result of that), but you should be able to use a relatively cheap model to automate typical system administration tasks, as long as it supports vision.

Can it do fully unattended installations? Well… It’s tricky. Can you figure out when to enter the BIOS on a random machine and what key to press? That is so variable across machines (and sometimes even on the same machine) that the only way I got it to work was to get into the BIOS (after a few tries), pick the boot volume myself and then hand over to piclaw.

The results and reliability are highly dependent on the AI model you use. I’ve had good experiences with astra, opus or sol on a variety of desktop automation tasks, and they were all able to use the Linkr to go through a standard OS installer. I also got them to check Windows updates (clicking through all the steps and choosing when to reboot) and even do console rpm-ostree and flatpak updates (and reboots) without any issues except the high token count that comes with handling screenshots (and model-side OCR).

This is what currently makes the Linkr worth looking at if you want to automate hardware setup, with the caveat that it can, indeed, do away with the odd bit of drudgery, but you need to be smart about how to do it (only retrieving screenshots at intervals/decision points to save tokens, for instance) and, preferably, script things so that the model is only really phoning it in.

For me it was almost a straight conversion of some of the stuff I’ve actually been doing in Proxmox for a while, so it’s nice to have a way to set up a blank box.

Minor Hassles

Hardware-wise, I should note that you can’t hit a reset button with the Linkr, so if your hardware doesn’t honour Ctrl+Alt+Del, you’re stuck. Other KVMs have dedicated interfaces to do that (and you can always use a smart plug, like I do with my test SBCs), but it’s something to keep in mind.

Additionally, I’ve had a couple of instances where the Linkr had a little trouble fitting directly into PCs or SBCs (including Radxa’s own) because of lack of clearance around the HDMI port or having the USB-C port on the side (which in a few cases meant it would point straight into the desk).

Nothing that you can’t fix with an HDMI extension cable (I used a Chromecast male-to-female extension, which is around 10cm), but worth mentioning since it doesn’t come in the box.

Another minor hassle is that if you power the Linkr directly from your target device, many devices (even standard PCs) won’t power up USB until they’ve gone past display initialisation, so the Linkr may skip UEFI/BIOS screens while it boots.

When testing SBCs, I had to use that power/data splitter cable to keep the Linkr powered up independently–otherwise it would boot up after the servers had tried (and failed) to initialise HDMI, so I’d get no video out (and, depending on device, they wouldn’t reactivate HDMI after USB input or Wake-on-LAN, either).

Verdict

I’m obviously biased towards anything that lets me put a machine on Wi-Fi and get at it through a browser without installing yet another proprietary app, so the Linkr already had quite a bit going for it.

But having an API that I can script against, or hand over to an agent once I’ve persuaded the machine to boot from the right device, makes it considerably more useful for the sort of testing I do, and I am quite likely to replace some of my other KVMs with Linkrs solely to free up Ethernet ports.

I’ll still need the odd HDMI extension and a separate power supply for a few machines (especially ARM ones), but at least the wiring will be much easier, and who knows, I might even start testing more than one SBC at once…

Oct 17^th 2026 · 1 min read · #notes #weekly

Notes for October 12-17

Had a fairly slow start to the week, then a progressively intense ramp-up that I tempered by way of a couple of strategically timed outings and a little time on my own projects.

I also spent quite a bit of time playing with the FM-1, although I’m having trouble finding the right kind of inspiration. And all sorts of stuff keeps breaking…

I’m Done With Time Machine on Synology

Time Machine broke again, mysteriously, on Golden Gate, and for some reason only my MacBook Pro was affected. Completely fed up with this, I just went and did what I had previously threatened: brought up a Samba Time Machine container (using mbentley/timemachine) in an unprivileged Debian LXC on tnas, backed by a bind mount from ZFS.

That entailed a bit of UID/GID finagling, and I must say I’m happy I don’t have to do it by hand myself again–took all of two minutes for piclaw to set everything up after I pointed it to my earlier post, and the LXC setup looks like this:

# lxc.conf
arch: amd64
cores: 2
features: nesting=1
hostname: timemachine
memory: 512
mp0: /local-zfs/folders/shares,mp=/mnt/shares
net0: name=eth0,bridge=vmbr0,firewall=1,ip=dhcp,type=veth
onboot: 1
ostype: debian
rootfs: local:102/vm-102-disk-0.raw,size=8G
swap: 512
unprivileged: 1

…whereas the container itself is managed via this snippet:

# docker-compose.yaml
services:
  timemachine:
    image: mbentley/timemachine:latest
    container_name: timemachine
    restart: always
    network_mode: host
    environment:
      # share is smb://hostname/TimeMachine
      - TM_USERNAME=timemachine
      - TM_GROUPNAME=timemachine
      - PASSWORD=<redacted>
      - TM_UID=1000
      - TM_GID=1000
      - SET_PERMISSIONS=false
      - VOLUME_SIZE_LIMIT=0
    volumes:
      # this is a pass-through mountpoint to the ZFS volume in Proxmox
      - /mnt/shares/timemachine:/opt/timemachine
    tmpfs:
      - /run/samba

Agent Stuff

I’m quite happy with how both kintsugi and my Obsidian server/agent harness are turning out (I’m drafting this and all my upcoming articles in it). kintsugi needed a bit more work than I expected, though, because:

iOS self-signed certificates and general Safari weirdness required a bunch of fixes across the board
despite applying container isolation to agents, it bafflingly lacked any sort of non-network resource limits and I had one workspace stall the entire machine it was running in
I needed better remote management and telemetry

Most of those are now done, and I have it running on my Orange Pi 6 Plus (yes, I’ve pretty much sorted out ARM support) to add USB device passthrough and other features I need to do embedded/ESP32 development over the next few weeks.

Media

Finally watched the end of Ted Lasso, which completely did not disappoint. Our current TV time is mostly handpicked YouTube and Apple TV fare (there’s just no real competition in overall quality) and I feel like there’s not enough time to watch all of the good stuff (like Dark Matter, which we’re behind on).

I also spent a while fixing various Calibre issues (my custom weekly news downloads were not working after an upgrade) and fiddling with my Xteink X4 a bit.

Oct 17^th 2026 · 1 min read · #food #photo #photography #restaurant

Miyazaki

Oct 15^th 2026 · 1 min read · #ai #anniversary #career #microsoft #work

11 Years at Microsoft

Unlike last year, I feel little need to mark the occasion. I could repeat most of last year’s reflections, with a few nuances.

Actually, let me do just that and save you the trouble:

AI has finally become a core consideration in business (pretty much all the customers I’ve worked with now understand the need for governance, foundational designs, standardised patterns, etc.), so we’re well past the experimentation phase.
Unsurprisingly, corporations have very different expectations from startups, so the Cambrian explosion I see outside work is much more muted and nuanced among customers, who are much more interested in getting actual value than in doing science experiments. The entire situation reminds me of the Big Data & ML hype cycle, only faster.

I am constantly reminded of this line from William Gibson’s Neuromancer:

“Night City was like a deranged experiment in social Darwinism, designed by a bored researcher who kept one thumb permanently on the fast-forward button.”

Geopolitical implications and polarisation are very much in people’s minds (inside and outside work). There is no avoiding discussions about the risks of relying on US inference providers, and the insularity of Europe regarding ownership of both core infrastructure and skilled resources is excruciatingly obvious to everyone even as they try to keep the business running.

Work is… worse than last year, I think, and yet on average the same:

My current organisation has been seesawing between regional and continental blueprints, which has again made being in Portugal a limiting factor in multiple respects (I guess this is cyclical, but it is extremely discouraging every time it happens).
I am definitely feeling stuck under the “Architect” label–I mentioned that last year, but the combination of constant reorgs and doing more and more things “outside the box” has become oppressive–if not, in fact, somewhat depressive.

Personally, things are OK in most respects, except:

My general physical condition: some back and neck trouble, and other things I’m figuring out
The biggest issue I have right now is clutter. And by that I mean largely mental clutter, given the sheer number of things I keep juggling.
Thanks to AI and my never-ending curiosity, I now also have trouble deciding what to do in my free time (there is just so much I can achieve), but I just can’t seem to relax–or stop having ideas.

And, of course, I am still letting my work define who I am, which I probably need to revisit at this point–I still want to build products, but that is just not happening here, at least right now.

Oct 14^th 2026 · 1 min read · #macos #productivity #ui

TrellisWM ➹

Tab grouping and auto-tiling aren’t new by any means, but zooming to a subset of windows is very, very neat.

It isn’t worth paying a subscription fee for (and if it proves to be really productive it might get Sherlocked within a couple of years, depending on whether Apple manages to shrug off the mess that is Stage Manager), and is likely to have three open-source clones tomorrow, so I’ll let the hype settle and give the concept a try once I find an implementation I can hack.

Oct 13^th 2026 · 1 min read · #ai #apple #hardware #homekit #siri

Apple HomeView ➹

This feels like Pixar’s Luxo or the classic half-dome iMac crossed over into the WALL-E universe, and looks like it is going to be cloned to oblivion. The $299 price is… interesting considering the tech inside it (parts of what it can do are highly non-trivial, and very much the kind of AI Apple likes doing).

I feel somewhat vindicated that Apple seems to have built a Siri for Families of sorts, but at the same time I have to wonder what flavour of it (and when) will be available in Europe, and exactly how much more bang for the buck it affords than, say, hacking a cheap Android picture frame–which is what I’ve been doing for a while now with the Lenovo ThinkSmart View.

I do think I want one, but there are a few things to settle besides price and value. For starters, having more screens in the house doesn’t seem like a winning proposition (the HomePods we’ve got work OK for alarms, reminders, etc.), and the American fascination with the kitchen island (which, by the way, needs power sockets in the middle…) doesn’t translate to European flats.

I find the Apple Home Collection a lot more interesting, but I’m going to reserve judgement until I understand whether it’s worth considering switching any of my home automation hardware… and all the LG gear is extremely expensive as such things go these days.

Oct 12^th 2026 · 2 min read · #apple #audio #automation #privacy #shortcuts #siri #watch

My Watch Should Listen To Me

I have been pondering the Audio Intelligence Live Rewind feature that Apple announced in the last keynote, and even though I know I am very late to this entire discussion, I still want to go on the record about it.

Oct 11^th 2026 · 9 min read · #fm #ios #midi #music #mvave #synthesizers

The M-VAVE FM-1

Regular readers will know I’ve had a long fascination for the Yamaha DX7–after all, besides getting a Reface DX quite some time ago, I also built a MiniDexed box and accumulated more software instruments than I have time to use, as well as an embarrassing amount of other physical gear that brands me as a victim of Gear Acquisition Syndrome.

Oct 11^th 2026 · 3 min read · #ai #go #hardware #life #notes #rust #virtualization #weekly

Notes for October 4-11

Sometimes I feel like I am only really alive on weekends. Sadly, that only seems to happen part-way through Sunday afternoons, which speaks to my utter lack of timing of late.

Oct 10^th 2026 · 5 min read · #agents #ai #coding #obsidian #tools #writing

We Need To Start Seeing Other Agents

I’m feeling the edges of burnout creep in, and having gone through that before, I’m pulling the brakes on anything that is neither immediately useful nor satisfactory.

Oct 8^th 2026 · 1 min read · #music #op-1 #synths #teenage-engineering

Teenage Engineering's CEO Says It'll Stop Making Synths ➹

This was not on my bingo card for… well, this decade, given the relative success of their more recent gear like the KO II.

Sure, there are some really strange things in the line-up, like the medieval flavour, and their UX can stray well beyond quirky into cuckoo land. But they’ve always been a source of whimsy and musical creativity–my OP-1 being a case in point.

Then again, things like the M-VAVE FM-1 are vastly cheaper these days (and arguably have better hardware features). They’re hackable, too, with an active ecosystem–which I suspect will dictate the success of future music hardware.

Oct 6^th 2026 · 1 min read · #ai #europe #llm #mistral #open-weight

Introducing Mistral Large 4 ➹

At last, Mistral has shipped a model that, judging by the model card, seems to compete with Chinese open-weight models–on paper, at least. It’s available via their API now, with the weights promised for the end of the month.

Now they need to demonstrate that the EU can do this consistently. There are scattered labs in Prague and Germany (which never seem to make it to Hacker News, for some reason), but the perception of AI work in Europe is still pretty poor–something I’ve ranted about before, including Mistral’s sovereignty pitch, which is at least backed up by their own European datacentres (the 3,800 Grace Blackwell GPUs used to train this model are relatively modest compared with hyperscaler clusters, but they did build out the infrastructure, and that deserves recognition).

Oct 6^th 2026 · 1 min read · #android #arm #google #googlebooks #intel

Google Admits Not Every Android App Runs Great on Intel Googlebooks ➹

Google acknowledging that some Android apps don’t run well on Intel Googlebooks is about what I’d expected. One of Android’s hidden catches is that Google has been neglecting it on Intel for a long time, and there isn’t a polished ARM-to-x86 translation layer to make up for it.

The main reason the Lenovo Chromebook I tested had such a seamless Android app experience was its MediaTek processor, and I really think the best Googlebooks are going to be the ARM ones.

Update: There is a fly in the ointment here that has me worried a bit, though. Googlebooks seem to absolutely require a Google login for activation (like Chromebooks), which makes it a little harder for me to consider them–you can use an Apple device (and a regular Android device) without an account, but if Google enforces this for Googlebooks and then they decide to arbitrarily nuke your account (as they’ve been known to do on occasion to people doing off-label things, even normal development), you’re completely hosed.

Oct 4^th 2026 · 17 min read · #android #arm #chromebook #chromeos #lenovo #review

The Lenovo Chromebook Plus

A few months ago, around the time Googlebooks were announced, I became really interested in Chromebooks again, partly because I’d completely lost touch and partly because I heard good things about the new ARM chipsets, namely the MediaTek Kompanio Ultra. But Chromebooks are not a common thing in my neck of the woods.

Oct 3^rd 2026 · 2 min read · #ai #dev #gherkin #life #tdd

Notes For September 27-October 3

This week was a bit different. For starters, it rained. And I visited the local office. And I consciously set aside hours to catch up on my writing instead of herding AI or doomscrolling, and again, failed at a bunch of that because–well, I lied.

Oct 2^nd 2026 · 1 min read · #apple #macos #photography #raw

Redlamp ➹

Redlamp was developed by one of the most talented Mac developers I’ve worked with in the past, and I’m very happy to see it–especially considering the way Apple completely neglected Aperture, and the shambles Photos still is if you have an inkling about real photography workflows.

It’s a native RAW editor for Apple Silicon, written in Swift and Metal, that opens folders directly and keeps edits in sidecar files. RAW development, masks and local adjustments, reusable Recipes and film looks are already there–although it’s still pre-alpha, with crop, healing and lens corrections on the roadmap.

Library management is further down the road too, so it isn’t a full Aperture replacement yet. I’ve added an app page with a screenshot and a few notes, and I look forward to using it from now on.

Oct 2^nd 2026 · 1 min read · #apple #exchange #macos #mail #microsoft #outlook

Apple Mail on Mac May Stop Syncing Microsoft 365 Accounts ➹

The writing has been on the wall for EWS in Exchange Online for ages, and Apple’s promised Graph support still hasn’t shipped; M365 can temporarily keep EWS enabled until April 2027, but right now some tenants may effectively kick out Apple Mail/Calendar, which is just stupid because it is definitely not Microsoft’s fault that EWS is being phased out in favor of a better, much more flexible approach.

I have zero idea about why it’s taking Apple this long to get their act together, since EWS has been feature-frozen since 2018 and Graph has been meat and potatoes stuff for ages. Even though I do run Outlook locally for my (very rare now) consulting/advisory work, I am not looking forward to being locked out of my accounts just because Apple effectively stopped maintaining its cloud account integrations and has been phoning it in for the last 3 or 4 macOS releases–not to mention that, as I’ve found over the past couple of years while trying to automate stuff, it is effectively impossible to have a unified local API for contacts, calendars and to-dos that actually works, let alone e-mail…

So I guess it’s a moderately good thing I’ve been using Graph wrappers for years now to get around Apple’s inability to maintain things, no?

Sep 30^th 2026 · 2 min read · #agents #ai #privacy #software

The Agents Are Among Us

I’ve been watching the launch of Meta’s Muse and OpenAI’s Dots somewhat… bemusedly, if you will. Being a regular Verge reader/listener, I am partial to Nilay Patel’s take that Silicon Valley’s “software brain” mentality has obscured both a lot of the actually practical uses of AI, and the unvarnished truth that the vast majority of people simply don’t care—or that they do care about the privacy and productivity impacts.

Sep 30^th 2026 · 1 min read · #apple #bugs #ipados #macos #ux

macOS Golden Gate Is a Buggy Mess ➹

I’ve come across most of these, but I’m still stuck on the papercuts I wrote about a few months ago. I have other priorities these days, and repeated exposure to Apple’s lapses in UX and QA judgement has somewhat dulled my senses.

iPadOS, meanwhile, still has weird bugs involving Spotlight and external keyboards, and those have been driving me nuts constantly.

Sep 26^th 2026 · 1 min read · #development #hardware #notes #weekly

Notes for September 21–26

Work was gruelling this week, with the rather unwelcome result that I didn’t exercise at all. I also fell asleep on the couch most afternoons/evenings just after closing shop–which ought to give you some idea of how much energy I had left for anything else.

Sep 22^nd 2026 · 7 min read · #ai #apple #automation #management #privacy #software

Some Old Expectations for the New Apple

In Nilay Patel’s Decoder conversation with Mark Gurman about John Ternus and Apple’s next big thing, Gurman describes something camera-equipped AirPods might eventually do: look at an email on your computer screen and help add an event to your calendar. And that was when he lost me.

Sep 20^th 2026 · 4 min read · #agents #ai #books #codex #gitea #hardware #notes #social media

Notes for September 13-20

It’s a bit insane that we’re past mid-September and my living room is still sitting at 29^oC at 7AM, but such is life in the late Holocene, I guess. Regardless, the shift towards Autumn is starting to show, and I’ve had a couple of days where I suddenly realized I was working “late” into the evening and the lights started coming on automatically (never a good sign unless you’re having fun).

Archives • 3D Site Map