iiinova — where your digital self keeps evolving

Andrew Cowie @istathar.space · 24/09/2026

not* Clearly but written with AI assistance.

Andrew Cowie @istathar.space · 24/09/2026

I switched to using Opus 5.x medium a while back for daily work because frankly emitting code is not the bottleneck, it's s consideration of impact and discussing nuance of implementation options. Occasionally I think to "launch a Sonnet subagent in the background" but often but worth the bother.

120

Reposted by Andrew Cowie

Ryan North @ryannorth.ca · 23/09/2026

Pompeii is a UNESCO world heritage site not because it was remarkable in its prime, but because it survived to our era untouched: one moment, preserved for all time. But future generations won't know what OUR world was like unless we do the same. This is why we MUST bury Toronto with a volcano

Picture of the streets of Pompeii, from https://commons.wikimedia.org/wiki/File:Pompei_-_panoramio_(26).jpg. Taken by Tanya Dedyukhina, used under CC

11867601625

Andrew Cowie @istathar.space · 23/09/2026

This is awesome: a terminal font in the style used by computers in WarGames (1983). mw.rat.bz/wgterm/ Mainframes deserve mainframe fonts!

The clipboard lady walking around the War Operation Plan Response mainframe computer in the movie WarGames (1983).

000

Andrew Cowie @istathar.space · 23/09/2026

Today I learned that YouTube has an internal category for useless content called `brainrot`. They also noted that most AI generated slop ends up being tagged this way. They still serve it to people, but aren't you reassured that they know they're doing it?

000

Andrew Cowie @istathar.space · 21/09/2026

Terrestrial ISPs are awful. LEO satellite constellations are great but kinda crowded (or will be until Saint Kessler has the last laugh). You want to think bigger this time: ISP via Moon Bounce. No pesky NIMBY objections to deal with.

110

Andrew Cowie @istathar.space · 21/09/2026

The best ISPs are those started by frustrated techies. I'd be a customer! Until you get burned out, smothered by dealing with regulations, and then get acquired by an incumbent telco anyway. And then *I* would have to start an ISP. That would be bad.

110

Andrew Cowie @istathar.space · 19/09/2026

This is a candidate for best sign of the year. In an era of demagogues, false profits, and mafia dons running countries, clear messaging and simple truths are priceless.

A prominent sign on the side of a Transport for NSW bus in Sydney, written in a clear, large font saying "I am a new bus"

121

Andrew Cowie @istathar.space · 19/09/2026

PM 🇨🇦 Mark Carney's speeches in Strasbourg and Toronto have gained considerable attention, but the real juice are his "Forward Guidance" videos. youtu.be/eYAwioM3RQI Worth watching. Reminds me that the quip "Canadian identity is merely the assertion that you're not American" goes back a long time.

youtu.be

Forward Guidance: A Stronger Canada

YouTube video by Mark Carney

021

Andrew Cowie @istathar.space · 14/09/2026

(subtweeting www.warhammer-community.com/en-gb/articl... and their riveting announcement of nothing in particular except another Ultramarine character nobody asked for)

warhammer-community.com

Sunday Preview – Space Marines drop in from orbit

Check out what’s coming up next week.

000

Andrew Cowie @istathar.space · 14/09/2026

How are the Space Marines "dropping in from orbit" if there are no Drop Pods and the Thunderhawk crashed? #warhammer40k

120

Andrew Cowie @istathar.space · 07/09/2026

(h/t youtube.com/watch?v=58x8... in a comment about the use of Space T-Rexes as a unit of measure)

youtube.com

Why Nancy Grace Roman Telescope Is About to Change Everything

YouTube video by Anton Petrov

000

Andrew Cowie @istathar.space · 07/09/2026

"They're called Imperial units because Metric Star Destroyer doesn't sound as intimidating"

100

Andrew Cowie @istathar.space · 22/07/2026

The ability of a cutting-edge model to move laterally is getting extraordinary.

010

Andrew Cowie @istathar.space · 15/07/2026

You know there's something special about the Universe because yesterday it let you make a release! only to find mere hours later that your release has a bug in it.

000

Reposted by Andrew Cowie

Darth Putin @darthputinkgb.bsky.social · 09/07/2026

Russian admiral talking to Gerasimov. "So, you get them all to line up, in full view of the enemy, and have them march to a certain death by drone?" "Da. Exactly" "Get my ship captains on the phone. We have a new strategy!"

2259389

Reposted by Andrew Cowie

Jonathan McDowell @planet4589.bsky.social · 09/07/2026

I think this is pretty serious: for several months the UNOOSA (UN Office of Outer Space Affairs) has not been making public new sat registrations apparently because of an IT policy issue. A month or two I could accept, but at this point it's a failure of the international space governance regime

39731

Andrew Cowie @istathar.space · 03/07/2026

We know the Devil exists because He created Continuous Integration and fixing CI is Hell. GitHub Actions are like the Fifth Circle of Hell, "wrath". Or maybe Seventh, because oh yeah baby, "violence". The sinner is doomed to re-run the same Workflow again, and again. And again.

020

Reposted by Andrew Cowie

mr. TIM @timkellogg.me · 25/06/2026

Ornith 1.0: full spectrum of open weights coding models from 9B to 397B The largest appears to rival Opus 4.7-4.8 All fine tuned from Gemma 4 and Qwen 3.5 but no, the cool part is how they made it… 👇 deep-reinforce.com/ornith_1_0.h...

This bar chart, titled **"LLM Performance Evaluation,"** compares the performance of various large language models across eight benchmarks, with a primary focus on **Ornith-1.0-397B** (represented by the prominent orange bar) against other well-known models like Claude Opus, DeepSeek, and Qwen.
* **Terminal Bench 2.1 (Terminus-2):** Ornith-1.0-397B leads this benchmark with a score of **77.5**, outperforming Qwen3.7-Max (**73.5**) and Claude Opus 4.7 (**70.3**).
* **SWE-bench Verified:** Ornith scores **82.4**, placing it ahead of DeepSeek-V4-Pro-1.6T (**80.6**) and Claude Opus 4.7 (**80.8**), though it trails behind Claude Opus 4.8, which leads at **87.6**.
* **SWE-bench Pro:** Ornith achieves the top position with a score of **62.2**, narrowly edging out GLM-5.2-744B (**62.1**) and Qwen3.7-Max (**60.6**), while significantly outperforming DeepSeek-V4-Pro-1.6T (**55.4**).
* **SWE-bench Multilingual:** Ornith secures the highest mark at **78.9**, closely followed by Qwen3.7-Max at **78.3** and DeepSeek-V4-Pro-1.6T at **76.2**.
* **NL2Repo:** Ornith scores **48.2**, slightly behind GLM-5.2-744B (**48.9**) and well below the benchmark leader, Claude Opus 4.7, which scores **69.7**. However, it remains ahead of Qwen3.7-Max (**47.2**).
* **Claw-eval Avg:** Ornith posts a strong **77.1**, outperforming DeepSeek-V4-Pro-1.6T (**75.8**) and Qwen3.5-397B (**70.7**), while falling just short of Claude Opus 4.7 (**78.2**).
* **SWE Atlas - QnA:** Ornith reaches **41.2**, beating out Claude Opus 4.7 (**40.3**) and DeepSeek-V4-Pro-1.6T (**27.2**), but falling behind Claude Opus 4.8 (**48.8**).
* **SWE Atlas - TW:** Ornith claims the top spot with a score of **39.1**, maintaining a narrow lead over Claude Opus 4.7, which scores **38.5**.

4819

Reposted by Andrew Cowie

🦋Special Kherson Cat🐈🇺🇦 @specialkhersoncat.bsky.social · 18/06/2026

A Russian air defense missile blows up a fuel tank at the Moscow Oil Refinery following a failed attempt to shoot down a Ukrainian drone.

1312750504

Reposted by Andrew Cowie

David Burbach 🇺🇸 @dburbach.bsky.social · 19/06/2026

Analysis of the Russian Burevestnik nuclear cruise missile. The crazy bastards appear to have done it. Air exhausted straight and spicy out of the core. 🥳

1217132

Andrew Cowie @istathar.space · 09/06/2026

@shashj.bsky.social my compliments for the choice to re-air, if that's the term, your interview with Sir Alex Younger. Watching again I was struck by his clarity of language and commitment over a lifetime of service. People such as he spend their lives in the shadows and are rarely recognised.

010

Andrew Cowie @istathar.space · 30/05/2026

Tomorrow is not Q day. It's K day. The day after is Y day. Then C day. Then C day again. It continues forever. If only you could decipher the enigmatic but beautiful code you would at last unlock the secret of reality.

000

Andrew Cowie @istathar.space · 29/05/2026

Good preliminary analysis from Scott Manley about the Blue Origin test failure and loss of launch site youtu.be/aaR6yEE-Myo

youtu.be

Blue Origin's Rocket Explosion - How Bad Is It?

YouTube video by Scott Manley

011

Reposted by Andrew Cowie

Senator David Shoebridge @davidshoebridge.bsky.social · 25/05/2026

Asking the Attorney General's office this morning about this outrageous information - they intentionally delayed publishing FOI documents until 1 minute before our Estimates hearing in February commenced so we couldn't ask them about the documents released.

49033

Andrew Cowie @istathar.space · 25/05/2026

That's actually not a bad idea at all. Might even be able to do it in pure systemd, and if not, that would be a pretty simple wrapper.

010

Andrew Cowie @istathar.space · 25/05/2026

We keep getting locked out because Tailscale is sending discovery and control plane traffic it shouldn't be.

000

Andrew Cowie @istathar.space · 25/05/2026

Mostly it's a defense in depth request, being able to reason about what traffic is on what interfaces but also, our plant is on corporate networks that we multi-home on have IDS that flags and shuts stuff down if there are VPN packets going through, so that's an additional reason;

100

Andrew Cowie @istathar.space · 25/05/2026

(this is a new problem, different from the request we've been making for a year to be able to force Tailscale Mac client to only use a specific interface so as to get around incorrect default network choices)

100

Andrew Cowie @istathar.space · 25/05/2026

Internal network is not trusted, got tens of thousands of foreign devices on it, any of which could actively be attempting to subvert Tailscale. Any vulnerability and we're hosed, so we'd really prefer if Tailscale wasn't trying to use it to communicate via it or consider any requests coming from it

210

Andrew Cowie @istathar.space · 25/05/2026

I'm sorry to bother you @apenwarr.ca but please have someone consider this? We keep asking but no one @tailscale.com is reacting. Our CTO is at the "we should just do Wiregard ourselves" point and I'd be very disappointed to lose the benefit of all our hard work implementing Tailscale over this.

100

Andrew Cowie @istathar.space · 25/05/2026

it shouldn't be bloody listening on interfaces it wasn't told to. Even having Tailscale report the existence of those internal networks up to the control plane is unacceptable. Given the prevalence of zero-days it's no longer something we can hand-wave away.

100

Andrew Cowie @istathar.space · 25/05/2026

We're not using Tailscale for LAN hosts to talk to each other, so having direct local connections isn't an optimisation we need. Yes, sure, we can put firewall rules on the machines to only allow Tailscale egress on a specific net, but

100

Andrew Cowie @istathar.space · 25/05/2026

Doing discovery on all interfaces is a nice default but we have legitimate security reasons to not want Tailscale on gateway machines having anything to do with internal interfaces.

100

Andrew Cowie @istathar.space · 25/05/2026

Is there a way to stop @tailscale.com from listening on anything other than a specific single nominated interface?

100

Andrew Cowie @istathar.space · 18/05/2026

In today's episode of cool-things-I-have-hallucinated-that-should-exist-but don't, we have "cryptobiological" links used to secure transfer of information. I'm trying to figure out what the biology part is. DNA coding? Pheromone sampling? Or maybe it's just "human enters passphrase". #cryptography

Gemini hallucinating how to enable one AWS account to read an encrypted AMI created in another account. Not only was the AI completely wrong, but it invented the term "cryptobiological" to explain the transfer of secret key material between the two sides.

000

Reposted by Andrew Cowie

Weedle @weedle.bsky.social · 12/05/2026

Three months from now the average voter is lined up at a gas station waiting for their ration and still getting headline alerts on their phone about YUGE FREE MONEY CHECK HITTING BANK ACCOUNTS SOON where if you actually click the link it's an AI summary about a town in Iceland enacting a UBI program

212120

Andrew Cowie @istathar.space · 07/05/2026

Well they did say this was going to keep happening. #linux security embargo broken so meet Ditry Frag github.com/V4bel/dirtyf... You had an entire week since the last one, don't complain.

github.com

GitHub - V4bel/dirtyfrag

Contribute to V4bel/dirtyfrag development by creating an account on GitHub.

020

Reposted by Andrew Cowie

Shipwreck @shipwreck75.bsky.social · 06/05/2026

Trump surprised Gulf allies by announcing “Project Freedom”, the officials said, angering leadership in Saudi Arabia. In response, the Kingdom informed the U.S. it would not allow the U.S. military to fly aircraft from Prince Sultan Airbase or fly through Saudi airspace to support the effort.

637165

Andrew Cowie @istathar.space · 06/05/2026

ROU A Lightly Militarized Misunderstanding Involving Explosives. @cultureshipnames.bsky.social

Screenshot from source article with text "Until then, it’s just a lightly militarized misunderstanding involving explosives."

185

Andrew Cowie @istathar.space · 06/05/2026

I always get worried when there are widespread delays in logging.

000

Andrew Cowie @istathar.space · 05/05/2026

Don't worry. I'm sure VMware will buy them before anything bad happens.

010

Andrew Cowie @istathar.space · 04/05/2026

Sadly it's still such a long way to go. Can you imagine 40 years from now? React programmers will be the Cobol grayhairs of the next generation, still working at banks on codebases that even the AI robot overlords are afraid of.

100

Andrew Cowie @istathar.space · 04/05/2026

I was more going for why there appears to be an asymmetry of the experience between front-end devs doing server side work (and being judged harshly for it) and back-end devs pulling off something that works in a web browser (and people being basically ok with it).

100

Andrew Cowie @istathar.space · 04/05/2026

Testing (for all that we go on about it) is hard, but AI sees implicit spec and ... implements it. So front-end devs can interate to "works" + "looks right". Back-end engineers work on nuance and "correct" is either defined or ... #yolo. Small wonder we think it's substandard.

100

Andrew Cowie @istathar.space · 04/05/2026

Front-end can *see* the outcome, both in terms of API responses but most of all in the page in the damn browser. Back-end server-side work is complex, multiple independent asynchronous parts, but more than anything is *opaque*.

110

Andrew Cowie @istathar.space · 04/05/2026

I have been converging on the idea that the only thing that matters is your ability (in the sense of opportunity and being enabled) to be clear in your requirements. Your prompt is almost irrelevant. Whatever the tests or examples specify is exactly what you're going to get.

110