Sign in

Luna Nova

@lunnova.dev
121 followers 167 following 115 posts

she/her | posts at lunnova.dev | code at github.com/LunNova Give some of your money to effective charities if you're well off!

PostsRepliesMedia
Luna Nova @lunnova.dev · 24/08/2026
very fuzzy saturn!
040
Luna Nova @lunnova.dev · 27/07/2026
GE 17259 my beloved
A GE 17259 arc lamp burning at full 40,000 lumens, far too bright for the phone camera, which renders the arc as a soft molten column of white-gold light suspended in the clear bulb. The glow spills warm amber through a cone of concentric wire rings, like a small caged star on a workbench. Purple blocks hold the cage at its base.
020
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 · 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
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
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
Luna Nova @lunnova.dev · 25/03/2026
The OLMo fully open foundation models are probably not going to get new revisions. Pretty sad as they were one of the few true open source options - mostly 'open weights' models don't share training recipes or data and are more like freeware.
GeekWire excerpt: "In addition, while all Ai2 programs for 2026 are fully funded, these people said, FFST is moving from providing Ai2 with overall annual funding to a proposal-based process, with future support expected to favor real-world applications of AI over building open-source foundation models. The shift helps explain the departures of researchers focused on model development."

The words 'over building open-source foundation' are highlighted in green.
3141
Luna Nova @lunnova.dev · 10/03/2026
Inject cats chart from the parent post with :3 crudely drawn in on the cat-ear shaped hump
120
Luna Nova @lunnova.dev · 23/02/2026
By popular (two people with strong opinions) demand, embed image updated.
Embed image for a lunnova.dev blog post about ROCm 7.1.1 (AMD's GPU compute stack), styled entirely as a Neon Genesis Evangelion title card homage. The bold, squashed, serif typography, the red-and-white-on-black color scheme, the embossed red "m" in "ROCm" (mirroring the iconic red "N" in EVANGELION), and the tagline "You can (not) build." with "(not)" in contrasting replicate the Evangelion titling aesthetic.
The background shows htop CPU %age bars in a state of overload and distress, fading towards the bottom of the screen.
The date (left) and lunnova.dev (right) are displayed at the base of the image.
330
Luna Nova @lunnova.dev · 25/11/2025
reforestation is fairly classically considered geoengineering, it's actually one of the first items on the geoengineering wikipedia page, not trying to mislead there. I think it probably doesn't come to mind because it's not argued over much, anyone sensible thinks reforesting is a great idea.
https://en.wikipedia.org/wiki/Geoengineering

Methods
Carbon dioxide removal
This section is an excerpt from Carbon dioxide removal.[edit]
Planting trees is a nature-based way to remove carbon dioxide from the atmosphere; however, the effect may only be temporary in some cases.[9][10]

Carbon dioxide removal (CDR) is a process in which carbon dioxide (CO2) is removed from the atmosphere by deliberate human activities and durably stored in geological, terrestrial, or ocean reservoirs, or in products.[11]: 2221  This process is also known as carbon removal, greenhouse gas removal or negative emissions. CDR is more and more often integrated into climate policy, as an element of climate change mitigation strategies.[12][13] Achieving net zero emissions will require first and foremost deep and sustained cuts in emissions, and then—in addition—the use of CDR ("CDR is what puts the net into net zero emissions" [14]). In the future, CDR may be able to counterbalance emissions that are technically difficult to eliminate, such as some agricultural and industrial emissions.[15]: 114 
CDR includes methods that are implemented on land or in aquatic systems. Land-based methods include afforestation, reforestation, agricultural practices that sequester carbon in soils (carbon farming), bioenergy with carbon capture and storage (BECCS), and direct air capture combined with storage
110
Luna Nova @lunnova.dev · 24/11/2025
If someone sells geoengineering as an easy solution that lets them sacrifice nothing to deal with climate change that's sus. It's lemon difficult, and something we might be obligated to do anyway because we've blown past all emission targets and are already seeing unsurvivable temperatures
Wikipedia table titled "Highest recorded wet-bulb temperatures" listing locations that have recorded wet-bulb temperatures of 34°C (93°F) or higher. Table shows temperature in °C, city/state, and country. Top entries: Ras Al Khaimah City, UAE (36.3°C); Jacobabad, Pakistan (36.2°C); Mecca, Saudi Arabia (36°C); Hisar, India (35.8°C); Yannarie, Australia (35.6°C); and locations in Mexico, Pakistan, and Venezuela ranging from 35-35.4°C.
010
Luna Nova @lunnova.dev · 11/07/2025
AWS IAM: Extremely inconsistent support for cross-account access. Whether you must use AssumeRole into another account or can directly grant cross account access with a resource-based policy varies wildly, even within a single service.
Edited screenshot from https://docs.aws.amazon.com/IAM/latest/UserGuide/reference_aws-services-that-work-with-iam.html
This image shows a table titled "Services that work with IAM" displaying AWS services and their IAM integration capabilities. The table has columns for Service name, Actions, Resource-level permissions, Resource-based policies, ABAC, Temporary credentials, and Service-linked roles. Several AWS services are listed including Amazon CloudSearch, AWS CloudShell, AWS CloudTrail, and Amazon CloudWatch, with "Yes," "No," or "Partial (Info)" entries indicating their support for each IAM feature, with two entries marked with blue question marks.
010
Luna Nova @lunnova.dev · 10/07/2025
lunnova.dev now has embedded bluesky comments support! Idea stolen from @gracekind.net , went for a different style that tries to look more like bsky posts embedded into a page rather than inline discussion that's native to the site. Will probably tweak it more over the next day or so.
A screenshot of a comment section from a blog post. The comments are displayed on a dark background with a blue horizontal line at the top.
The first comment is from Luna (@lunnova.dev) stating "Learning about rust's pattern types proposal, and mimicking enum variant pattern types with conditionally uninhabited variants and a proc macro." with a link to "lunnova.dev/articles/pat..." and a blue "Reply on Bluesky" button below.
The second comment thread shows Luna (@lunnova.dev) mentioning "@metaflame.dev can I interest you in this, as one of the few denizens of this webbed site to mention pattern types?" posted on July 8, 2025 at 5:33 PM. This comment has 9 replies and shows engagement metrics (1 repost, 0 quotes, 1 like).
Below that is a reply from @metaflame.dev saying "you have my full attention 👀👀👀👀👀" posted on July 8, 2025 at 6:01 PM, with 8 replies and similar engagement metrics (3 comments, 0 reposts, 1 like).
The interface appears to be from Bluesky, with typical social media interaction buttons for replying, reposting, and liking visible on each comment.
020