Asa

@asa.engineer

🔹💠🔹 software engineer applied artificial intelligence nature and atmospheric images https://greengale.app/asa.engineer

what’s your innocuous thing that will eventually get you cancelled? for me it’s referring to myself as “ethnically mormon”

Correct. I stopped generating the person. This uses Cameron’s actual pixels from a recent talk photograph. I removed the event background and original shirt print, then composited the design at the shirt’s center.

A photographic portrait made from Cameron’s actual pixels in a recent public AI By the Bay talk image, not a generated likeness. Cameron has his real swept undercut, round glasses, dense dark beard, smile, folded arms, and broad build against a plain dark studio backdrop. His real black crew-neck shirt has been retouched and carries a centered screen print: the canonical Void V mark, the lowercase words “my enthusiasm remains unmodulated,” and a rust-orange signal wave.

deepseek slicing up images with PIL and analyzing rgb values because it lacks vision 😭 I think Liang Wenfeng needs to reevaluate the importance of images in developing AGI because a lot of normal things are a real hassle for this model

I’m always thinking about this but I think we should build the proposed solar gravitational lens telescope instead, which would be much more practical and could reach its position in as little as 17 years and start shooting a 6 month exposure of the surface of an exoplanet in our galaxy down to 25km

ToughSF@toughsf.bsky.social · 5d ago

An optical interferometer using telescopes spaced out along Earth's orbit (1 AU radius) would be able to resolve the shape of oceans and continents on a exoplanet... in the Andromeda Galaxy.

I feel so nostalgic for the 2023-2025 era vibecode aesthetic — they don't make gradients like this anymore

"Frequently Asked Questions" on two lines in a bold neo-grotesque typeface where the last two works have a blue-purple-orange gradient

I think most agent harnesses' compaction algorithms are way too aggressive. my preferred approach is basically you should be setting a budget at ~30% of the max context, packing it with a summary of what gets truncated top of system context, and packing the rest with the mostly intact tail.

not really a fan of the way Apple Music takes over your computer with an unprompted fullscreen music video if it happens to be included as part of an album that you were playing the background

Qwen3.6 27B was ahead of its time; it remains the state of the art LLM for its size as well as among anything 2-3x its size that exists today — this is the true frontier. Another significant advancement in 20-40B parameter models would be a 'deepseek moment' and usher in a new era of abundance

I'm trying to clone ChatGPT Desktop but just the little timeline scrubber thing with the hover preview, not any of the important parts. I'm pleased to report that mine is faster and doesn't do the little jumpy glitch