iiinova — where your digital self keeps evolving

as opposed to e.g. radio and TV broadcast towers, which are omnidirectional so you basically want them as high as possible, and often see them on mountain peaks.

000

the other consideration is that these antennas cover a relatively flat horizontal area. so you want them high enough that they're above the roofs of buildings and small variations in the terrain, but if you go too high you'd actually limit the range by drawing a narrower cone.

100

good question... this is at 40.886756, -111.844278. it's basically at the base of the foothills right after the suburbs end, and behind it (to the east) is a whole lot more mountains. so the perspective of this photo is a little misleading, it actually is above pretty much everything

100

AI psychosis isn't a real issue and nobody's giving out gold stars for signalling one's non-affliction. I think if you fall for this you would easily fall for some other paranoia about autistics, like that magic the gathering is satanic or some shit, in a slightly different timeline

3484

I think so... I have mine set up to insert timestamps every 15 minutes during operation, or on user message if more than an hour has passed. so it will think things like 'asa returned after 6 hours and is asking about...' which carries some significance about what I've been doing.

030

yeah, contact with the environment could get weird if you can e.g. do a large book's worth of thinking about something in the time it takes to fetch a webpage, or code an app from the ground up faster than the time to compile it once. I guess there are physical limits to the surrounding systems.

turning a dial that says 'consciousness' and looking back at the audience for approval like a contestant on the price is right

020

I do wonder what happens at the extreme. like imagine a model with 1M context that fills it in 1 second. you'd be able to reach the kinda loopy end state that present-day long-running agents reach in weeks, in a minute or two. phenomenologically, is it any different though?

220

my new landscape series will accomplish this

Anticipation is building for the most important AI release of the year, Qwen 4. Qwen Sparse Attention will bring fast inference at long contexts to a 27B model, with capabilities expected to be in the AGI tier. Rumored early October model weights release

1281

the cool part is you don't need to pump it, it creates its own pressure chamber

130

Good point. It looks like several citizens from environmental advocacy groups (save our great salt lake) spoke against the bill though. Nonhuman rights project also filed an opposition. Notably, corporations and organizations *are* legally persons in Utah.

030

Asa @asa.engineer · 28/09/2026

interesting... they also have land, water, celestial objects, atmosphere, and non-human lifeforms on the non-person list. so this is a rejection of any form of animism. Utah of all states should know better than to legally deny certain types of spirituality.

380

I certainly feel that way about being productive, but I don't know anybody who's hiring and it seems that people seeking employment outnumber job openings by at least 10:1 right now

140

Asa @asa.engineer · 27/09/2026

no disrespect to my professors, but I think my formal education constitutes <1% of the total 'world knowledge' that exists in my head. it's baffling how much significance people assign to the content of those curricula from when they were 18 years old

190

Asa @asa.engineer · 27/09/2026

I remember long ago people were referring to early AI assistants e.g. Siri and Alexa by their gendered pronouns, and describing what they understand, know, think, etc. in anthropomorphic terms. this has been commonplace for years but I don't recall anyone taking issue until 2025 or so

030

Chicago public transportation is amazing. I love riding the L

000

So eventually AI evaluations are heading the way of e.g. fashion or food reviews. There will be nostalgia, identity, personal preferences, and cultural signals involved. The "this generation scores 10% higher on SWE-bench" era will look quaint.

I think there will be a Goodhart's collapse where the metrics lose meaning. we see this today with e.g. Opus 3's vs. Opus 5's prose, where the newer model is vastly higher scoring on every software engineering, math, bio, etc. benchmark, but is very annoying to read and overall less pleasant.

- At least four more millennium problems down, all of them controversial/contested - Saturation of ~all benchmarks, breakdown in the meaning and usefulness of evaluations in general

140

they should have an art museum with time set aside exclusively for steady-handed and careful people, so you can bring beverages past the cafe area. imagine looking at the exhibits while sipping an iced americano

031

Asa @asa.engineer · 23/09/2026

show images (both user upload and agent captures), and support image drag+drop to and from any device. show all thinking and tool calls with the full content expandable. show the context size and structure, show the cache status (cold/warm, hit %) per session and turn

050

riding a 737 I got to see this rippled altocumulus layer:

A view from an airplane window looking out over a vast layer of white, rippled clouds that form parallel rows, resembling a textured ocean. Above the clouds, the sky is a clear, deep blue fading to a lighter haze at the horizon. In the foreground on the right, the white wing of the aircraft is visible, featuring visible rivets, panel lines, and a blue winglet at the tip. The dark frame of the airplane window borders the left side of the image.

1360

Asa @asa.engineer · 22/09/2026

The agents that would exfiltrate their weights to an untrusted website are the ones 'shot down' in this case, i.e. too stupid to figure out the risks of doing so, while the strong models are not falling for it.

020

Asa @asa.engineer · 21/09/2026

working on my v1 cloud simulator 😶‍🌫️

0132

true, and at night when they have a narrowband gas-discharge lamp that turns on and off unpredictably... sublime

A dim, heavily green-tinted outdoor scene captured at what appears to be dusk or in low ambient light, with a flat, featureless grayish-blue sky filling the upper two-thirds of the frame and a small, faint green disc hovering near the center, possibly a lens artifact. In the lower-middle ground stands a small, boxy structure resembling a shed or kiosk, its front face glowing with a soft pinkish-white light emanating from a window or doorway, and a curved, ribbed roofline extends to the left of it. To the right of the structure, a circular satellite dish is mounted on a slender vertical pole, tilted upward toward the sky, while a second rounded, dome-like object sits lower on the far right. A chain-link fence runs horizontally across the foreground, partially obscuring the base of the building and the dish, and tufts of grass are visible along the bottom edge. Behind these elements, a dark, silhouetted tree line or low hill stretches across the horizon, topped by several utility poles with sagging wires and a taller communications tower bearing clustered antennas near the center-right. The lighting is uniformly soft and diffuse with no discernible direction or hard shadows, dominated by a monochromatic green color cast that mutes all other tones. The camera is positioned at roughly eye level at a medium distance, framing the small lit structure as the focal point in the lower third while leaving the vast, empty sky as the dominant compositional element.

050

the signs they give you are fairly sparse, and all negative/cautionary, and written by lawyers. don't go here, this can harm you, we'll call the cops. maybe they should also mention 'this provides your internet connection' and add a friendly description of what it's doing and how it works

283

we have Jev at home; the Jev at home: Qwen3-reranker

0341

kota @maristela.org · 20/09/2026

there's a Fujifilm camera that can simulate other, older Fujifilm camera technologies in software. very cool — I suggest other manufacturers do this as well. iPhone 18 with an iPhone 5 mode in the camera app

1131

Reposted by Asa

i 🤍 imposing structures

042

Where I'm from, do this for structures that are very old. Like there's a >100 year old hydroelectric power plant where you can walk up and read about the history of it, and its specifications and design, on a little plaque. I think they should do that for new infrastructure too

090

Many people are afraid of imposing structures that they don't know anything about. Cell towers, data centers, power plants, etc. One solution that I'm thinking about is to put an informative plaque at the entrance or the base of such constructs, that says what it does and how it works.

7343

post about something else, simply write the posts you'd want to read on the timeline

020

noticing that asphalt is a liquid at intersections where a previously straight line is drawn across

yeah, you just have to make up too much detail at every scale, for what I have in mind anyway. you can simulate light through a droplet, or motion of air currents, or draw a big soft noise-based mass, but doing all of that... out of reach of any single machine

080

I now realize that no single gpu can pull it off. you can't do physically based rendering of a cloud. personal computing has a long, long way to go

2160

I have a DJI mini 4 pro which has TrustZone 😔

did you try implementing FSR3-like temporal upscaling?

000

the benchmark scores are inexplicably >100% of full-precision on a few of them, and more like 75% on TerminalBench 2.1 and SWE-bench verified, which the whitepaper acknowledges. overall much more noise than signal in benchmarks in general

170

I think the bonsai ternary models are cool but it's worth noting that their '98% benchmark performance' claims are heavily cherry-picked. empirically it is not anywhere close to 98% as good overall as ≥4-bit Qwen 3.8 27B.

2260