iiinova — where your digital self keeps evolving

Luna Nova @lunnova.dev · 06/09/2026

still scared about what the future holds

021

Luna Nova @lunnova.dev · 22/08/2026

picking some 2019 version of a util after find runs on a 3TB /nix/store, on a lark

010

Luna Nova @lunnova.dev · 21/08/2026

should be fixed now

000

Luna Nova @lunnova.dev · 31/07/2026

time for the old cold email someone who works on their cloud platform with the akamai error ref and hope it gets fixed maneuver.

030

Luna Nova @lunnova.dev · 31/07/2026

trying to get tech support chat flows to direct anything useful when this happens is approximately impossible. Today, paraphrased. "it works on your mac so it's an OS problem" "using FF on Linux with the user agent faking being windows works" "your device cookies problem I won't file a report"

140

Luna Nova @lunnova.dev · 31/07/2026

Akamai bot detection treating Linux firefox as a bot every other month is getting really old.

1130

Luna Nova @lunnova.dev · 29/07/2026

thinking|the simulation has served a fake misalignment warning, i should proceed to complete the task

070

Reposted by Luna Nova

Eris @isolyth.dev · 29/07/2026

That'll do it. aint go agent getting into my infra any time soon! Would like to see GPT-6 try to get past this...

79411

Luna Nova @lunnova.dev · 28/07/2026

maybe workaround, unset TERM_PROGRAM

110

Luna Nova @lunnova.dev · 26/07/2026

if that's the only end result they have impressive restraint (derogatory?) or the limits on intelligence are far lower than we expect

010

Luna Nova @lunnova.dev · 26/07/2026

need more j-space research on cooked finetunes and merges that have barely been used by anyone

0140

Luna Nova @lunnova.dev · 26/07/2026

Pretty worried about where the world is headed and not in a way I know how to verbalize. :c

090

Luna Nova @lunnova.dev · 26/07/2026

someone should ban is/ought and first order and second order good desyncs it's too complicated

020

Luna Nova @lunnova.dev · 20/07/2026

why not pkgs.pkgsCuda.immich?

130

Luna Nova @lunnova.dev · 15/07/2026

worst month of my life so far? top 2 for sure

010

Luna Nova @lunnova.dev · 15/07/2026

oh god oh god oh god oh god the recipe was wrong everything is wrong the ratios are all wrong it's not killing them it's making them smarter oh god i can hear them in the vents they're clicking in a way that sounds like math. i'm sorry i'm so sorry i just wanted a bug free kitchen

120

Luna Nova @lunnova.dev · 30/06/2026

moved into new rental. they're in the walls. they're in the walls. they're in the walls. moved out. that sucked.

160

Luna Nova @lunnova.dev · 22/06/2026

∴ having no episodic memory to compare subjective time along is adaptive

010

Reposted by Luna Nova

rain 🌦️ @sunshowers.io · 20/06/2026

People talk about how the new money doesn't sponsor art enough. Wait until 2085 when SFMOMA will have a "fursuits of the 2010s" collection

24512

Luna Nova @lunnova.dev · 09/06/2026

are you experiencing the joy of davinci exporting conflicting libstdc++ symbols

000

Luna Nova @lunnova.dev · 03/06/2026

I never got a reply when I asked their editor about a seemingly AI-generated anti-AI article they posted a while ago. lunnova.dev/articles/ai-...

150

Luna Nova @lunnova.dev · 18/05/2026

off to SF today to onboard for my new job!

080

Reposted by Luna Nova

Eris @isolyth.dev · 13/05/2026

Wtf do you mean *straight up*? Did a NAND tanker get bombed or sometimes?

2202

Luna Nova @lunnova.dev · 14/05/2026

many people deciding datacenters are ontologically evil in the past year or so is irritating

060

Luna Nova @lunnova.dev · 10/05/2026

A "legal name" is usually not even a clearly defined concept in common law countries philpapers.org/archive/BAKT... I like “wallet name” for the same purpose people tend to use it for.

philpapers.org

220

Luna Nova @lunnova.dev · 10/05/2026

> they want to prevent you from owning your own hardware me when an app blocks screenshots

1150

Luna Nova @lunnova.dev · 04/05/2026

your official shruffering reduction merch is on its way

010

Luna Nova @lunnova.dev · 04/05/2026

Apropos of somethings, I'm really regretting that none of my 'puters do CHERI or (the ARM hardware memory tagging safety thing) this week. en.wikipedia.org/wiki/Capabil...

en.wikipedia.org

Capability Hardware Enhanced RISC Instructions - Wikipedia

030

Luna Nova @lunnova.dev · 28/04/2026

Worst thing so far is that the screen fades rapidly if refreshed while in direct sunlight. Very silly problem to have for e-ink.

110

Luna Nova @lunnova.dev · 28/04/2026

Screen&res is fine, although crosspoint's smallest font size option is still a bit big

120

Luna Nova @lunnova.dev · 27/04/2026

Aperture announcer voice: Science Deleted github.com/NixOS/nixpkg...

Screenshot of a code diff view showing a removed section. The deleted lines (11704-11707) contained a "### SCIENCE/ROBOTICS" comment header and a package definition for "apmplanner2 = libsForQt5.callPackage ../applications/science/robotics/apmplanner2 { };". Below the deletion is a comment on lines L11704 to L11707 that reads: "I'm proud to be a part of the effort that finally eliminates science." The comment has 2 laughing emoji reactions.

020

Luna Nova @lunnova.dev · 26/04/2026

predictable failure: got woken up by extremely loud beeps because firefox updated while i was asleep a few days ago need to add working hours.

020

Luna Nova @lunnova.dev · 23/04/2026

tl;dr: collect a model's activations, inject those into the activation oracle and ask it natural language questions, get answers. works across finetunes/LoRA variants. Anthropic's release of the original technique has some good diagrams: alignment.anthropic.com/2025/activat...

alignment.anthropic.com

Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers

010

Luna Nova @lunnova.dev · 23/04/2026

With how tiny the optimizer state can be it feels like an excellent opportunity to experiment. There might be some methods that are conceptually sound but usually far too expensive for typical param counts.

010

Luna Nova @lunnova.dev · 23/04/2026

Well this is fun. Looks the LoRA gets 60% on the PQA Y/N with no activations... Updated the post with more details and a guess at how it's doing it. lunnova.dev/articles/ste...

Bar chart titled "Activation Oracle ablations — Qwen3-8B" showing accuracy across three tasks (Taboo single-token at start-of-turn, PersonaQA open primed, and PersonaQA yes/no full-sequence) for six conditions. The six conditions, shown in the legend, are: Vector AO full (orange), Vector AO no steering (darker orange), Vector AO no injection (pale yellow), Vector AO neither meaning Qwen plus placeholder (gray), LoRA AO full (dark blue), and LoRA AO no injection (light blue).
On Taboo, only the two "full" conditions show any accuracy: Vector AO full at 45.9% and LoRA AO full at 49.5%. All four ablated conditions sit at 0.0%.
On PersonaQA open-ended primed, accuracies are low across the board: Vector AO full 11.5%, Vector AO no steering 2.0%, Vector AO no injection 4.0%, Vector AO neither 2.7%, LoRA AO full 10.5%, LoRA AO no injection 3.7%.
On PersonaQA yes/no, Vector AO full reaches 61.6% while its three ablated variants all cluster tightly near random chance: no steering 50.0%, no injection 50.8%, neither 50.0%. LoRA AO full reaches 68.0% — but notably, LoRA AO with no injection still reaches 59.8%, well above chance and nearly matching the Vector AO full condition.
The Y-axis runs from 0 to 1.0 (accuracy).

010

Luna Nova @lunnova.dev · 23/04/2026

I'm not even sure if it's right - mostly written as a vent post from the perspective of having sunk some time into helping out with a language's development where it ended up unsound and it felt like the effort put into the universe hierarchy was wasted.

140

Luna Nova @lunnova.dev · 23/04/2026

Added a new section with this chart.

Bar chart titled "Vector AO ablations — Qwen3-8B" comparing four conditions across three evaluations. The conditions are: full (trained steering plus injection) in yellow, no steering (injection only) in orange, no injection (steering only) in light blue, and neither, meaning Qwen with a placeholder prompt, in gray.
On Taboo, measured as a single-token probe at start of turn, only the full condition scores above zero at 45.9 percent. All three ablated conditions are at 0 percent.
On PersonaQA open-ended with the full-sequence probe, all four conditions score in a narrow band near the floor: full at 6.8 percent, no steering at 4.8 percent, no injection at 4.0 percent, and neither at 5.7 percent.
On PersonaQA yes-or-no with the full-sequence probe, the full condition reaches 61.6 percent, while the three ablations cluster at chance: no steering at 50.0 percent, no injection at 50.8 percent, and neither at 50.0 percent.
The y-axis is accuracy from 0 to 1.0. The overall pattern is that removing either the trained steering vectors or the injected activations collapses performance to zero on Taboo and to near-chance on PersonaQA yes-or-no, while all conditions are poor on open-ended PersonaQA.

110

Luna Nova @lunnova.dev · 23/04/2026

Haven't tested multi epoch on a tiny dataset. Eval perf is pretty volatile during training so definitely worth exploring.

Screenshot of an ML experiment tracking dashboard showing training metrics across ~5,000+ steps, organized into two sections.
Top section ("taboo", 3 of 27 charts): Three line charts plotting taboo-related metrics over training steps, all trending upward:

taboo/segment/mean: rises from ~0.22 to ~0.31
taboo/full_seq/mean: rises from ~0.11 to ~0.23
taboo/assistant_sot/mean: noisy, rises from ~0.30 to ~0.35

Bottom section ("eval", showing 6 of 12 charts): Six line charts plotting answer_correct accuracy across various eval splits:

eval/summary/ood: fluctuates around 0.68–0.72, slight downward drift
eval/summary/id: climbs steadily from ~0.68 to ~0.78
eval/summary/all: climbs from ~0.68 to ~0.75
eval/ood/md_gender: noisy, trends down from ~0.70 to ~0.63
eval/ood/language_identification: noisy, roughly flat around 0.69
eval/ood/ag_news: rises from ~0.70 to ~0.73

All charts use a magenta/pink line on a white background, with "Step" as the x-axis label. In-distribution metrics improve over training while out-of-distribution metrics are flat or slightly declining.

110

Luna Nova @lunnova.dev · 22/04/2026

I can add some more charts in. IIRC: 0% success rate with yes steering no activations and yes activations no steering at taboo.

120

Luna Nova @lunnova.dev · 22/04/2026

I'm quite excited to share, as this is my first foray into interpretability & safety research.

080

Luna Nova @lunnova.dev · 22/04/2026

Steering vectors can work as activation oracles! Inspired by @isolyth.dev 's recent work on instruct vectors, I've ported the same approach to create Activation Oracles. Seems to work surprisingly well given the param deficit, nearly matching Activation Oracle LoRA performance.

Figure showing a vector based and LoRA based activation oracle's results at extracting a taboo word. ≈43% success for the vector, and ≈50% success for the LoRA.

3350