Notes for November 14–21

After a relatively blissful weekend doing absolutely none of what I’d planned, nor clearing my personal backlog (I ended up doing some work admin stuff that had been ), I decided to be a little bit more systematic about things.

I am now back to Pomodoro timer land because that’s the only way to carve out 20-25 minutes of uninterrupted focus time for anything, including writing, and I think this week it served me well, for besides taking a couple of fairly momentous decisions at work, I also managed to get a surprising amount of stuff done in focused chunks.

Bye, Parallels, It Was Nice Using You

I cancelled my subscription last weekend, for three reasons:

  • It was a €69.99 subscription, and they charged me without any noticeable warning this time around.
  • There is no way running Windows VMs on a Mac is worth paying that much every year anymore.
  • Absolutely nothing they were offering as distinguishing extras and “features” was worth it, even for professional use.

In short, I might have kept it if they had sensible pricing, because the only thing I will grant does better than most options is hardware support. And yet the moat there has pretty much dried up now that is starting to implement very usable 3D acceleration via viogpu3d, which gives me in excess of 60fps in WebGL Aquarium and the couple of CAD programs I tried so far.

The trick to get it working well, if you’re setting up a new VM today, is to use the uninstaller script for the default video driver in the UTM utilities CD, then reboot to the Windows fallback video driver before running the installer. This could be better documented (or done right in the installer), but it’s easy enough to sort out.

Given what I’m seeing in emulation and graphics throughput, I expect that a few months of AI-accelerated development will close that gap (and others).

But on top of that, is very much automatable these days. was too, but ’s approach seems nicer, and you can use tooling to handle its disk images instead of yet another proprietary format.

All I think it’s missing from a management perspective is a snapshot tree (i.e., not just a linear stack of OS snapshots), but I suspect that too will come in time.

Slimming Down piclaw

The long, hard slog to slim down piclaw and remove the redundancies I had built atop the original SDK that are now embedded in pi-durable continues, and like I’ve written , in the process I tore out entire chunks of machinery and features I (at least) have no use for:

Chart of source and test code added to and removed from piclaw between 20 October and 20 November 2026
A month of piclaw churn, with more code removed than added.

I suspect I will eventually write a bit more about this at length as a follow-up to , but for the moment I’m just enjoying polishing it down and systematically closing issues without much of a hurry.

In parallel, kintsugi has become my long-haul/heavy-CPU project environment, and I’ve been using it to do whatever and work I need to get done. I’ve yet to figure out a decent way to manage build caches consistently, but at least now I don’t need to rebuild entire sandboxes every weekend.

Measuring Agent Calls

I’m back building building blocks, and since I needed a stable, reusable web wrapper for the GitHub Copilot SDK (which is now Microsoft’s core agent harness), I decided to create one and model it after pi-durable–but with built-in tool observability, because I will be using this in places where I will want to drill through the inter-agent and tool call chain without all the nice Azure observability tooling.

The result is pretty decent, and here’s what it looks like when I ask the base agent to create four sub-agents to render a 3D animation of a MacBook Pro, complete with interactive approvals:

ms-durable interface showing the MacBook Pro animation request, model and tool calls, and interactive approvals
The ms-durable trace for the MacBook Pro animation, including tool calls and approvals.

…and here’s the output render, for posterity:

Animated 3D rendering of a MacBook Pro
The resulting MacBook Pro animation.

Sorting Out Mail.app

I’ve been falling behind on my personal e-mail to a point where it’s just ridiculous, and since Apple killed Mail plug-ins years ago (including , which was the best way I ever had to file messages) and has utterly failed to deliver any effective way to manage inboxes (no, the “smart move” thing they injected into the toolbar does not really work for me), I decided to have another go at it.

helps a lot, but since I didn’t want to do it semi-manually and have become used to having Copilot do that for me at work, I decided to throw AI at the problem–however, I had a few constraints:

  • No login access to any e-mail account (local automation only)
  • No way to do anything more than list folders and messages or move them between folders (i.e., no destructive actions)
  • Minimal metadata access (although I am allowing message text search as well now)
  • Minimal (or no) local agent access

So I created a restricted server to do exactly that, and set up (which runs completely sandboxed on my Macs, seeing as I’m running the iPad application and it uses ARM emulation inside to run the agent) to use a private Azure OpenAI endpoint to drive that (I don’t have Macs powerful enough to run decent local LLMs, and Strata still crashes my RTX 3060, but Qwen does work for a few minutes and is good enough at this).

And boy, what a shitty mess Apple Events automation still is. I’ve been ranting about how badly Apple has let automation deteriorate , and there have been rumours of Apple doing servers of their own to automate macOS, but the foundations are so bad that it took me hours to get something as simple as moving messages between accounts to work reliably.

The code is up on GitHub, and I will likely be adding a few things to it or incorporating it into somehow–but I won’t be investing a lot of effort into it until I learn more about what Apple intends to do–if anything.

Ocarina of Time

A video game title screen with a sword-and-shield emblem, a rider on horseback by a river, and Continue, New Game and Options menu choices.
In my defense, I was left unsupervised.

Notes for November 9–13

Boy, what a week.

To begin with, thanks to the inauspicious alignment of a couple of depressions (weather ones, for a change), the weather was miserable enough that I barely left the house, which I am only rating positively because it prevented me from attending Web Summit and its sycophantic, utterly divorced from reality hype fest.

WetSummit Was The Best Summit

As it turns out, the Portuguese government decided to shutter schools and non-essential public services and we all stayed home on Wednesday. Luckily, depression Charles took a more southerly track than forecast (making for a day of intense but otherwise very much non-tropical rainstorms here), but by that time “WetSummit” had shut down for the day, and it was probably its best yet.

This may sound harsh given it was indeed a great event when it came to Lisbon and that (if I had made it there) it would be a rare opportunity to actually meet some people I haven’t seen in years, but I just can’t sit through another talk on how X is going to transform the world (for any value of X, especially AI). It has become genuinely impossible to sort out the wheat from the chaff in startup land over the past couple of years unless a) you know their team or b) you know, from your personal experience, that they are trying to solve an actual problem.

Civic Responsibility Is Cool

Either way, I don’t blame the government for doing the responsible thing–the forecasts on Tuesday looked really bad (to the point where we battened down the hatches and I planned for both Internet and power outages):

Windy weather forecast the day before, showing a depression off Lisbon and strong winds along the coast I don’t care about the politics, but watching commentators complain that the economy was shut down for a day is both predictable and proof that we have the worst kind of intelligentsia in the political sphere–the exact same people would be gleefully harping on about floods and disaster response failures if the worst had come to pass.

The trend over the past decade (as I perceive it) is that the weather here is becoming sub-tropical in the unstable sense (i.e., bigger swings in rainfall, temperature, etc.), so I’m happy to see prevention being enacted.

As an aside, a part of this trend I am increasingly concerned with is that my back and neck certainly feel it when this kind of weather swings by, and it’s only likely to get worse.

Work and Industry

Work-wise, like atmospheric pressure, things were also highly unstable: I got involved in a pretty amazing new project even as other things reached an astonishing new low (with barely a glimmer of light at the end of any of those tunnels), and that was just, well, Monday.

Fortunately, the Jensen Huang soap opera has provided ample entertainment of the most fascinating kind. What other industry might even dream of attaining such peak intensity of scandal and drama week upon week, I do not know, but this timeline is even better than a William Gibson novel right now: (checks notes) wars, mad dictators, almighty industrialists, critical technology developments, minor cataclysms, we have it all. Sometimes two of each!

And, of course, the M365 outage was a somewhat apt way to both start Friday 13th and make it obvious that in between that and a touch of fever/migraines, should just pack it in early and throw these notes together.

Older Toys Are The Best

With all this, my brain felt an unusually high compulsion to do a bit of exploration, so I tinkered late into a couple of evenings and built rexweb, which combines ideas from Rex and webterm and gives you a web front-end to multiple persistent terminal sessions using a vaguely compliant approximation of the rex protocol, specced out to my agents by means of what I can only classify as interpretive dance.

In short, I felt compelled to flip some of the assumptions of both on their heads and figure out if terminal multiplexing was a thing I actually needed done in that particular way, but at least I had a good design to mimic.

The jury is still out on how useful it all is, but back in the early 90s, I would dial into my college VAX using a 2400 bps modem and multiplex LAT sessions using some long-forgotten abomination that allowed me to do Kermit file transfers and have independent terminal windows (not too many, obviously) on Windows/286–so I highly appreciate how far we’ve come, but at the same time it’s fascinating to see how we keep reinventing some wheels…

Notes for November 1–8

It’s dark and rainy and annoying enough outside that I am both enjoying sitting on the couch with a blanket and pining for my regular walks to the nearest park. And as usual for the season, work is steadily stacking up until the holidays, so this was an even more intense week than the last.

But besides , I’ve had a couple of free hours here and there to keep track of some stuff…

Nature Is Healing

Yes, US elections are inescapable, and no, I don’t care about the politics (just their outcomes), but I think this quote from Nilay Patel on the Vergecast (which I watch almost religiously), via Simon Willison’s note, is the best summation of the zeitgeist yet:

The products aren’t good enough to sell people on the costs.

You cannot tell people that you have to listen to the whine of a data center in their backyard for AI overviews and some of the most hilarious deep fake slop in history.

It’s not worth it. It just isn’t worth it to anyone. And agents aren’t real and they might not be being used the way that anyone wants them to yet. All these companies are spending more than they’re making on these products. Refrigerators holding blood storage in Europe are being hacked.

The costs are too high. The benefits are not real. And if you connect that to corruption, you’re like, in order to impose these costs on me, these CEOs are in the White House shaking Trump’s hand, renaming it to the Gulf of America, paying for this ballroom that we don’t want. It’s obvious crypto-corruption, and it’s still slop.

The AI industry will not get out of it until they make better products. Until everyone can see without needing to be convinced, without needing an ad campaign, without any of this stuff that just using these products makes your life better.

Because, you see, outside our little technology bubble, hasn’t really improved anything overall–yes, there are indirect benefits in many industries, there are cost savings, even the occasional new non-tech product, but all of it has to be offset against the stupefying increases in complexity.

Just like during the Big Data hype cycle, companies now have to have new processes and tools to manage , and the elephant in the room everywhere is the , which affect businesses and consumers alike in more fundamental ways, like how much it now costs to have any sort of computing device to work on even if all you need to do is e-mail and spreadsheets.

And that, I think, together with reality creeping in around the edges of the Anthropic IPO, might finally break the hype. But until then (and even for a good while after), prices for everything will just keep creeping up.

Penny Pinching

As an indirect offshoot of that, I’ve set my agents to go through my cloud services and nodes to relentlessly optimise RAM usage, storage allocations and (in the case of my Azure subscription) disk writes and backup cycles, which are a significant part of my monthly Azure bill.

I’ve been neglecting power consumption for a while, but as we start getting into the cold season I’m again looking at wattage, starting with optimising my servers a bit:

Homelab

Free from the tyranny of upgrading piclaw to the latest upstream (it’s happening, but first I have a couple of agents slowly whittling it down to the core to remove bloat), I’ve gone back to tidying up my home services, which need some foundational work (I am completely fed up with self-signed certificates), a few usability fixes and a bit of resource optimisation.

Fortunately I have been able to sort of run Strata on my RTX 3060 for a few hours at a time after a few more kernel and driver updates, so quite a bit of the work has been done with a pastiche of local models, but with just one discrete GPU across seven nodes, things have been slow.

Naming Things Is Hard, But Finding Them Is Harder

The first part entails finishing dance to the point where it’s usable as a personal CA (no, I am not going to let it use 4 as a private key).

I am still mildly annoyed at the fact that using .local for, well, local, in-house LAN services in 2026 still seems to be a corner case, but I am determined to have proper TLS working without having to use for stuff that is literally on the same wiring. If there isn’t a simpler way than installing my own CA on my own devices, I’ll at least own the issuer and have my machines use ACME to self-provision.

So far that has meant minor tweaks to mdnsbridge as well, and I am still debugging the entire CA/issuer lifecycle, but fortunately I can just yell at a pair of LLMs to finish the work and deploy it–the number of SKILL.md and note lookups involved in the process is pretty wild at this point, but let’s just say that piclaw’s core memory hacks and having them add notes to are still paying off handsomely.

Near-frictionless Writing

The second point is mostly fixes to my server, which has completely replaced and my publishing workflow. On a good day, I can just pop open the URL in a browser, edit my drafts directly in including pasting in and cropping images, and then have a couple of rounds with the built-in agent to review the post, fix any typos, resolve any internal links (which was the most time-consuming post-processing phase), pick tags and just issue direct GitHub API calls to push the exact changeset (which then starts a delta-optimised site rebuild).

As an authoring environment, it is pretty awesome–and if it is broken I can just SSH in or use piclaw’s built-in terminal to edit drafts (which is what I’m using now as it’s fixing an edge case with TLS in the server).

I think it will be good enough to use 100% of the time in a few weeks, since we’re past the 80%/20% mark between feature completeness and polish.

Measure Thrice, Cut Twice

Besides power, I’m currently looking at the way I gather data across my little menagerie of machines. I already gather data into Graphite for bare-metal hardware and some internal VM stats (using telegraf), and services (using both their native telemetry and a bit of telegraf trickery), token usage from all my agents (via Graphite exporters) and all my home automation telemetry.

Graphite is still pretty great and I have zero interest in changing the metrics schemas or the protocol itself, but I have been looking at replacing it with VictoriaMetrics on the server side solely because I would like a leaner, meaner server daemon to remove some load from the .

Entertainment

We finally finished Dark Matter’s second season last evening, which turned out to be a great example of how to extend an original plot and promises an interesting third season. We’ve skipped Silo and some of the other stuff for the moment, but it’s time to revisit our watchlist for these longer seasonal evenings…

Reviving a Touch Bar MacBook with Fedora Silverblue

Following up on my , I decided to install onto my old MacBookPro13,2 (the 2016, T1 Touch Bar model), which I once (somewhat naively) termed .

With the benefit of hindsight, I can now finally come to terms with why this particular hardware revision is much maligned in general, and revisiting it was an interesting experience because:

  • I had forgotten how bad the keyboard is
  • The Touch Bar is worse than useless in general, and causes a number of issues in Linux
  • It can’t use modern 5GHz Wi-Fi (well, it can, but this setup doesn’t support WPA3, and probably never will)
  • The battery is somewhat shot
  • It warms up considerably at the slightest provocation

It does, however, have an absolutely beautiful 2560x1600 panel that reminds me of the I had to send back (still sad about that), and four USB-C ports so it’s trivial to keep topped up enough to use as a thin client:

MacBook Pro running Fedora Silverblue, with the GNOME About panel and Touch Bar visible
MacBook Pro running Fedora Silverblue, with the GNOME About panel and Touch Bar visible

Installation

The first step is, as you’d expect, to just plug in a USB drive and boot from it. ’s hardware support is, as usual, stellar (power management, backlight, touchpad, etc., all worked out of the box) and I was immediately productive on it by dint of installing pi, asking it to map Esc to Caps Lock, and get the rest of the hardware to work while I just spun up , logged onto my server and kept writing.

The result is a sort of overpowered Chromebook (it has 16GB RAM and entirely too much storage, and even if it has a relatively paltry two cores and four threads, it’s still an i5), and were it not for the pretty hefty weight when compared to more modern machines, I would call it a very satisfactory conversion.

Kernel Fixes

The entire process took something like eight hours altogether (loosely spread over a couple of days), and I’m pretty sure you can get it down to under 4 hours if you slave away at it as a meat proxy to the AI–I just kept doing other stuff and occasionally rebooted for it to test the fixes.

What pi ended up doing was:

  • Building akmod packages for both the Touch Bar and audio initialization
  • Debugging microphone and camera inputs
  • Tracking down and fixing incorrect NVRAM settings that prevented the Wi-Fi from working in 5GHz
  • Taking several stabs at getting full suspend/resume to work while ensuring all the weird hardware recovered correctly (this is what took the longest, by far)

As you’d expect, I got it to publish all the notes, steps and packages on GitHub, so this should be easily reproducible if you have an identical machine (which, mind you, you might not–there are several subtle variations of this).

Next Time…

Since I forgot to turn on LUKS disk encryption in the first install and I want to move to Fedora 45 (which is launching this week), I will eventually have to do the whole thing again–this time likely going straight to my Niri desktop setup rather than staying in GNOME. However, given the number of changes expected in Fedora 45, I will let that settle first.

All in all, reasonably smooth sailing, although power management and sleep/resume have been a major pain, since it also had to patch and tweak the Touch Bar and audio drivers to come back on when picking up the machine.

There are a couple of things that still don’t really work, but they’re fairly minor considering this is probably going to live out its days as a couch machine:

  • External displays (might have to do with the old Intel chipset–some monitors are detected but nothing renders)
  • There’s an annoying phantom tap thing in the trackpad driver, so for the moment I am back to press to click, which is slightly maddening

I like our new AI overlords, especially considering I can use them to breathe new life into this kind of otherwise useless hardware. I might even look into replacing the battery on this one as well…

Notes for October 25-31

Have I ever mentioned how much I hate daylight savings and the utter stupidity of disrupting my (already poorly attuned) circadian rhythms for a completely arbitrary change that makes mornings only slightly more bearable and evenings a dark pit of mild despair?

Well, I guess I just have. It might have also been the tipping point for my feeling slightly feverish and achy towards the end of the working week, which certainly did not help my overall mood or productivity.

Hacking The Norns

The week did have an auspicious start, since spurred by tinkering with the I decided to get some music/audio hacking in. I’ve been toying with the idea of turning my Norns into a Microcosm effects pedal for , and not only did things finally start to gel with SuperCollider after deciding to reorganize the code but I also thought it would be nice to throw AI at the problem of getting a usable UX on that and iterating through all the visible options:

Mosaic A Norns interface views showing Activity, Shape, Mix and Time controls, settings, presets and error states
Mosaic A interface views and secondary controls, with settings, presets and status screens.

Economic Madness

I’ve looked at Anthropic’s S-1 and am torn between it being one of the best works of investor fiction ever written (especially regarding profitability vs compute and training costs) and a potential proof point for industry “transformation”.

If it goes through (and they don’t get their lunch eaten by OpenAI, open weights or anything else that will sink per-token pricing, which will certainly have more interesting implications) they might have something interesting there, but I am not betting my money on anything that is pure AI at this point.

3D Printing

I finally managed to make a small dent into my 3D printing backlog–I had been meaning to design 3D printed cases for the , the and another couple of SBCs I have, but a fortuitous search found serviceable ready-to-print ones, so my office is now looking slightly more tidy, and I have fresh new surfaces for dust to settle on, with two caveats:

  • The ’s fan noise was much more noticeable with the case on
  • I am temporarily lacking the required M3/14mm screws to button down the Sigma properly

Still, it was nice to go back to popping stuff off build plates, and the cases make for a much neater desk. Plus I realized I had an entire roll of white TPU stored away that has come in pretty handy to print feet for most of the cases.

Reverse Engineering Renaissance

One of the AI-related things I am actually quite happy about is the blooming reverse engineering/hardware repurposing scene, which I’ve been tracking since I would indeed like to make better use of some of my older hardware.

There are now credible, largely working efforts to do stuff like get Linux running on the Lenovo ThinkSmart View and iPad and iPhone hardware (sadly, triggered by Omarchy, but I guess the outcomes are a net positive), re-targeting or recompiling games for the Mac and Windows, and various other retrocomputing initiatives.

Of course only time will tell how good and actually useful any of those things are, but the barrier to entry is now so low that taking over Android photo frames and debloating Smart TVs is almost commonplace, and I’m totally here for it.

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 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 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.

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.

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.

Read More...

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.

Read More...

The Radxa Linkr

I’ve done quite a few KVM reviews , 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.

Read More...

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.

Read More...

Miyazaki

Miyazaki
An excellent place to have lunch

11 Years at Microsoft

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

Read More...

My Watch Should Listen To Me

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

Read More...

The M-VAVE FM-1

Regular readers will know I’ve had a long fascination for the –after all, besides getting a quite some time ago, I also built a 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.

Read More...

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.

Read More...

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.

Read More...

Archives • 3D Site Map