iiinova — where your digital self keeps evolving

hah precisely why I didn't bother renewing my Anthropic sub despite all the glowing praise for Opus 5.5

000

Alienities sounds like something I would have liked to study at university.

It's not just Adobe's dark patterns or the hidden cancellation fees in your gym membership. I don't think it's functionally possible for a company to get "big" without extracting money from customers that they don't actually want to spend.

Nick Fisher @hydroxide.dev · 26/09/2026

Never underestimate how much of the economy relies on exploiting consumer inertia or information asymmetry. www.apollo.com/wealth/insig...

131

Dart + WASM is starting to gather speed, the great Simon Binder has started putting together a package to implement the missing host components needed to run a Dart-compiled app via non-JS runtimes like wasmtime. github.com/simolus3/was...

github.com

GitHub - simolus3/wasm.dart: Tools to run Dart in any WebAssembly runtime

Tools to run Dart in any WebAssembly runtime. Contribute to simolus3/wasm.dart development by creating an account on GitHub.

051

Nick Fisher @hydroxide.dev · 26/09/2026

Guy who just discovered TLA*: "I think TLA could be really useful here" * me

000

Nick Fisher @hydroxide.dev · 25/09/2026

Trance music, best listened to at 2am driving on empty roads with the windows down with the cold winter air blowing in.

040

Nick Fisher @hydroxide.dev · 25/09/2026

Some more lip rigging work, using the ribbon approach from Rik Schutte's course on Blender Studio.

040

Of course, otherwise your program won't run on a 1979 8-bit desktop calculator from a now defunct Eastern European manufacturer. And if it won't, then why are you even programming bro?

100

That was my reaction too, but I've seen some incredible output from Opus 5.5 (wasn't released until after I ran this test) which makes me think it might be closer than that. If anyone has an Anthropic sub and some quota to burn, a comparison would be great.

Can LLMs retopologize a facial mesh?

Can LLMs retopologize a facial mesh? (I'll save you a click if you don't want to watch the video, the answer is no). youtu.be/oIB9J_NCZaw

youtu.be

YouTube video by Nick Fisher

1130

This is what I’m hearing. It’s tempting to resubscribe but I know I’ll just get bitten by their guardrails again.

110

Also: - Sol 6 is terrible, zero point in using it. - Astra is good at high or above, but the 20x plan now lasts me about 3 days with *modest* usage. No longer seems worth it.

100

I cancelled my Anthropic sub over this, so looks like it's time to reconsider my OpenAI sub too.

230

If you had the awareness to ask them to hold their hand in front of their face in the first place, you're probably savvy enough to know something's off to begin with, so I guess it's not really going to derail things significantly.

000

Nick Fisher @hydroxide.dev · 23/09/2026

I've seen a lot of people sniffing out deepfake scammers on video calls (fake job interviews etc) by asking them to hold their hand or their fingers in front of their face. That hack won't last long though, so I don't know what people are going to do next.

100

This doesn't look like an official release, just a random jev implementation atop diffusiongemma via cloud run?

112

Nick Fisher @hydroxide.dev · 22/09/2026

My recent contributions to the Dart SDK have landed, adding support for importing multiple and/or shared WebAssembly memories. Getting closer to "native" interop between Dart and external WASM libraries (rather than hacky JS bridges). github.com/dart-lang/sd... github.com/dart-lang/sd...

github.com

[dart2wasm] Fix multi-memory operands and optimizer support · dart-lang/sdk@1dce1b4

Closes https://github.com/dart-lang/sdk/pull/64247 GitOrigin-RevId: 6b9d1a809f0d6eca72f0bd6b5dde3341a0bf9368 Change-Id: I70e4fdcae69cd2f18568b05706d88bbadbaaf0de Reviewed-on: https://dart-review.g...

161

Nick Fisher @hydroxide.dev · 22/09/2026

Your toddler has good taste

This was my jam back in the day.

160

I watched Two Prosecutors (2025) last night. Don't recommend to be honest, it was bleak and depressing, but in all the wrong ways. www.imdb.com/title/tt3552...

imdb.com

000

I'm just trying to avoid participating in The Discourse (TM).

I've been around long enough to know I shouldn't google that

100

When life gives you lemons, have a lemon party.

220

Have you tried hooking it up to Prolog (or maybe even Lean) to create ontologies so you can gradually offload queries into a more deterministic (and less compute intensive) system?

121

This is about airgapped networks so the assumption is that there’s no ethernet interface at all (and often the USB ports are superglued shut). However, power modulation is a known attack vector too, eg www.scitepress.org/Papers/2026/...

scitepress.org

But it's so low bandwidth that you would only use it for things like small snippets of text (passwords, botnet commands, etc). It is a legitimate cybersecurity concern in ultra-secure networks but I don't think we have anything to fear when it comes to a self-replicating AI escaping containment.

100

It's a legitimate data exfiltration method, in the BitWhisper paper they report success up with the receiver up to 40cm away (so still pretty close). It's not just temps, you any other channel works as long as you have a receiver (fan noise, power draw, etc). ieeexplore.ieee.org/document/724...

ieeexplore.ieee.org

200

Someone from OpenAI was using this as an example of how AI escapes containment from an airgapped network. Theoretically possible and fun to muse about when you're high, but totally irrelevant. The mundane risks are much more worrying, but much less fun to talk about.

220

Increase CPU load, CPU gets warmer, nearby machine with temperature sensor detects a 1 Decrease load, CPU gets colder, sensor detects a 0 We now have a (noisy) binary communication channel!

200

(if I got my numbers wrong please correct me, I'm still half asleep on a Saturday morning)

000

Kimi K3 has ~2.8 trillion params, at FP4 that is 2.8e12 * 4 = 1.12e+13 bits 1.12e+13 bits / 800 bits / hour = 1.4e10 hours to exfil weights 1.4e+10 / 24 / 365 = ~1.6 million *years* to copy itself ...I think we're probably fine

111

Back-of-the-envelope calculation for "super-intelligent AGI can self-replicate over airgapped machines via CPU thermal signalling". BitWhisper did this, maxing out at 8 bits per hour, but let's be generous and assume 100x improvement in future hardware/sensors, so 800 bits / hour.

330

When I was living in Melbourne I worked in a few coworking spaces which was a decent balance too, the last one was about 10 minutes walk from my apartment. That was the best of both worlds.

110

I do miss being shoulder-to-shoulder with people but it's the daily commute that gets me (and even worse now that I'd have to work "drop the kids off at school" into that routine). On balance I'd prefer WFH with the occasional trip to the office (maybe once every week or two).

120

(the caveat for Gumroad is that they unilaterally decide to pay themselves an affiliate fee if someone buys after reaching your product via their own discovery feature, which is fairly questionable). Also Stripe pays out within 7 days, vs 14 days for Gumroad (or up to 30 for Superhive).

000

Sales from the Holotype website are starting to trickle in, which is great to see. Advantage is I only pay Stripe fees, no commission Gumroad takes 10% (with a caveat), Superhive is 30% BEFORE merchant fee & Blender fund contribution which is...a lot.

100

My German professor at university accused me of plagiarism because he thought my (English language) writing was too sophisticated to have been written by an undergrad. Any student who grew up actually using the phrase "load-bearing seam" is definitely in for a bad time nowadays though.

070

Overnight I set DeepSeek Harness + V4.1 Flash head-to-head with Codex + Astra to try and retopologize an awfully messy face mesh. Results and video coming soon!