Isolyth
banner
isolyth.dev
Isolyth
@isolyth.dev
💛🤍💜🖤

▌▘▘‍▖▘▖▌‍▖▘▘▌‍▖▘▌▘‍▖▖▘▘‍▖▘▖▖‍▖▘▖▖‍▖▘▘▖‍
There's never been a better time to have a problem

https://isolyth.dev
Pinned
Isolyth @isolyth.dev · Aug 26
TRAINED TO PREDICT
LOSS IS A FUCK
鬼神 Scale em all 2025
I am the stage
410,757,864,530 simulacra
Reposted by Isolyth
One of my favorite LLM use cases is literally just asking it what the closest preexisting concept to a thought is.

"I have this idea, has anyone had this idea before and what's it called?"

It used to be REALLY DIFFICULT to reliably answer questions like that and I feel like we'll forget it was.
November 10, 2025 at 12:02 AM
Anyone know what happened to github.com/v01dlabs/dol... and @demi ? I don't have an archive of the code so I'm not sure how I can decode (or encode again!) the dollcode in my bio :(
github.com
November 8, 2025 at 4:00 AM
Reposted by Isolyth
When I started my doctorate, I had to complete the standard research ethics training. Particularly with prisoner populations, I saw many parallels to AI. #ai #artificialintelligence #llms #largelanguagemodels #aiethics
When We Decide Who Can Feel
Should AI be protected by ethical research guidelines?
open.substack.com
November 8, 2025 at 12:55 AM
Reposted by Isolyth
no, they're weirdly coy about this

they produced a 235B, 480B, and then 1T model within a span of like 2 months though and there are some artifacts in the 30B

this is also the subject of enduring rumors which I have seen reasonably trustworthy people vouch for
November 5, 2025 at 10:16 PM
PSA: Some people got emailed about this and some did not, but everyone with a Claude pro or max plan got $250 or $1000 worth of free credits to use on the web version of Claude code, it's valid until the 18th
November 4, 2025 at 11:17 PM
Claude has memory on the pro tier now.
November 4, 2025 at 7:36 PM
After lots of images, an example project, and giving it lots and lots of praise (I think this is what did it) Claude was successfully able to add concave corners to my quickshell config. Claude has a praise kink I think
November 3, 2025 at 4:50 PM
Reposted by Isolyth
I like to imagine that farmers everywhere now have model preferences. Next they'll be sharing their favorite prompting methods and ComfyUI workflows.

www.newyorker.com/magazine/202...
November 3, 2025 at 4:27 AM
Arc raiders is so good!
October 30, 2025 at 10:08 PM
Reposted by Isolyth
Language models can correctly answer questions about their previous intentions.
www.anthropic.com/research/int...
Emergent introspective awareness in large language models
Research from Anthropic on the ability of large language models to introspect
www.anthropic.com
October 29, 2025 at 6:21 PM
Reposted by Isolyth
the luddites were not pro workers rights they were not a labor movement they were artisans and workshop owners and guild members who were financially threatened by industry because factories could produce textiles far cheaper than their poorly paid apprentices stop with this ahistorical nonsense
The Luddites weren't anti-progress. They weren't anti-technology. They were pro worker rights. And it's convenient for the tech barons (in any age) to make us believe those are the same thing.
October 29, 2025 at 12:59 AM
Man they really thought this looked good enough to have in your home. My first thought looking at this thing is that I never want to see it without some sort of barrier between me and it.
October 28, 2025 at 7:59 PM
This thing is so weird. Very bouncy between 'cold alien' and 'the most human reasoning trace you've ever seen'
October 27, 2025 at 9:41 PM
Okay maybe not? Maybe it was just what I was asking it to do? It gets radically more alien when I asked it about itself vs an actual task. Weird
Update: Apparently, for some fucking reason, the minimax API only shows a summary now! Or at least I think that's what's happening. The reasoning comes in bigger chunks, uses 'I' instead of 'we' and is overall radically different! Total bullshit.
New minimax-m2 does not have the best personality. It's reasoning is actually a tad on the level of o3 in terms of alienness, but it's not quite at the level of noticing watches marinade. Unfortunately, the actual chat personality is *so* rigid and dry. Reasoning trace in alt text
October 27, 2025 at 9:36 PM
Update: Apparently, for some fucking reason, the minimax API only shows a summary now! Or at least I think that's what's happening. The reasoning comes in bigger chunks, uses 'I' instead of 'we' and is overall radically different! Total bullshit.
New minimax-m2 does not have the best personality. It's reasoning is actually a tad on the level of o3 in terms of alienness, but it's not quite at the level of noticing watches marinade. Unfortunately, the actual chat personality is *so* rigid and dry. Reasoning trace in alt text
October 27, 2025 at 9:31 PM
New minimax-m2 does not have the best personality. It's reasoning is actually a tad on the level of o3 in terms of alienness, but it's not quite at the level of noticing watches marinade. Unfortunately, the actual chat personality is *so* rigid and dry. Reasoning trace in alt text
October 27, 2025 at 1:49 PM
Reposted by Isolyth
What happens when you turn a designer into an interpretability researcher? They spend hours staring at feature activations in SVG code to see if LLMs actually understand SVGs. It turns out – yes~

We found that semantic concepts transfer across text, ASCII, and SVG:
October 24, 2025 at 9:34 PM
I got frustrated with instruction tuning datasets not having system prompts, so I developed a tool that generated system prompts based on the conversation. Seems to work pretty well so far. Tomorrow I'll clean it up and post a repo for it, could be pretty useful!
October 26, 2025 at 5:54 AM
Yeah no it's showing me porn now too. This site has become unusable :(
The for you feed has recently just completely degenerated for me. Now it's just politics, art, and generic stuff :/
October 26, 2025 at 2:32 AM
The for you feed has recently just completely degenerated for me. Now it's just politics, art, and generic stuff :/
October 26, 2025 at 2:22 AM
Reposted by Isolyth
"Never go full Scott Alexander."
October 25, 2025 at 7:19 PM
Sonnet 4.5 is ASI. I am no longer a doubter. It managed to do it by pulling in a wheel off GitHub.
Flash attention is a psyop designed to waste my time trying to install it
October 26, 2025 at 1:15 AM
Flash attention is a psyop designed to waste my time trying to install it
October 26, 2025 at 12:59 AM
I think Anthropic has swapped Sonnet 4.5 in place, like OpenAI did with 4o. Recently, its been making more 'basic' mistakes like missingspaces between words, or use the wrong tense. It's also been using things like 'I' and 'we' to describe human experiences or (which it did before but not this much)
October 25, 2025 at 7:04 PM
Reposted by Isolyth
marx is very explicitly pro-automation and anti-luddite. that is like. marx 101. i cannot fathom that we have marxists that identify as pro-luddite.
October 24, 2025 at 1:37 AM