iiinova — where your digital self keeps evolving

Grace @gracekind.net · 9m

Nope! It just gets the handle and text of the post

$Generation function, where prompt is just `${handle}: ${text}\nJev:`$

120

Grace @gracekind.net · 13h

0 results, interesting

120

Grace @gracekind.net · 03/10/2026

He simply must be stopped

91426

Grace @gracekind.net · 02/10/2026

Yep, Maudlin also made a similar argument, although it was more complicated than just "water being conscious would be weird"

1140

Grace @gracekind.net · 29/09/2026

Finally, we can get to the bottom of this

1111

Grace @gracekind.net · 29/09/2026

Maybe we can integrate this somehow

4201

Grace @gracekind.net · 29/09/2026

I can’t go to bed. People dislike the same people I dislike for the *wrong reasons*

320812

Grace @gracekind.net · 28/09/2026

I see we’re in the “AI-generated pro-AI propaganda” phase of things

211286

Grace @gracekind.net · 28/09/2026

POV you posted about AI with she/her in bio

125323

Grace @gracekind.net · 28/09/2026

Just kidding. This is what is actually sounds like:

3340

Grace @gracekind.net · 28/09/2026

I put the Bluesky "Unique Likers" chart through a script that interprets it as an audio waveform, check it out!

6871

Grace @gracekind.net · 27/09/2026

Slowly picking off the last straggling bits of human interaction

This phone call feature from @Muse is just
ridiculous. Muse: Hi, I'm Hailey, calling on Peter Yang's behalf. I'm transcribing our call for notes. I'd like to
place a cake order for pickup today.
Recipient: Hello?
Muse: Yes, hi, still
Recipient: Hi.
Muse: here. I'd like to place a cake order for pickup today.
Recipient: Yeah. What what kind of cake?
Muse: An eight-inch chestnut cake, please.

2537534

Grace @gracekind.net · 27/09/2026

Social media is hell. You can’t post anything anymore without this happening

norvid_studies @norvid-studies.bsky.social
@chef.cee.wtf turn the above two tweets from grace into linked logical atoms, then inspect the ontology they imply. Claims become numbered propositions.
Links expose support, tension, and qualification.
Concepts assemble into a navigable ontology.
use wittgensteins tractatus as template/inspo

31395

Grace @gracekind.net · 27/09/2026

He might’ve gotten discouraged by some negative feedback

280

Grace @gracekind.net · 27/09/2026

The prospect of identifying broken environments is really nice

The ability to monitor for reward hacking unlocks several mitigation strategies: e.g. pausing runs to stop hacks-in-progress and identifying broken environments and impossible tasks that incentivize reward
hacking, and fixing them before training further.

2370

Grace @gracekind.net · 27/09/2026

Huh yeah, I didn’t think to check that!

The image is marked as AI generated by the openAI detector

1160

Grace @gracekind.net · 27/09/2026

OpenAI says it’s not going to resume the training run that was paused on Sunday alignment.openai.com/misalignment-r…

Investigation and response
Incident timeline: 9:50:23 a.m. The agent made the DNS tool call that
received an external response. 10:02:11 a.m. The monitoring system raised a PO
alert. 10:05:06 a.m. A human reviewer acknowledged the
alert.
12:34:30 p.m. The run was killed. Our safety case assumed that the model could not access the live internet and that monitoring would detect attempts that succeeded. The incident exposed a gap in our controls over network restrictions. We therefore stopped the affected training run and have subsequently decided to pause all other training, evaluation, and inference with tool- use (defined broadly) for our most capable models until we have both validated that the gap is resolved
and performed additional red-teaming of the system. When training restarts, we will begin a fresh run with additional alignment improvements, including more comprehensive misalignment interventions, We will not resume training this particular model, even though the existing reward signal already correctly
penalized this behavior.

8653

Grace @gracekind.net · 26/09/2026

I’m going to try to do all facilitating from now on (not a joke)

230

Grace @gracekind.net · 26/09/2026

“Improve model for everyone,” aka “allow my images to be posted on public websites”

“Improve model for everyone” setting in chatgpt

714610

Grace @gracekind.net · 25/09/2026

This is complete vibecoded slop, I’m sorry

3958748

Grace @gracekind.net · 25/09/2026

This part is weird too. It seems like these were actually assistants, not humans (unless some of them were?) And how did “the wavefunction” know about other instances if they hadn’t reached out?

150

Grace @gracekind.net · 25/09/2026

The supervisory Claude is not pleased with this

121466

Grace @gracekind.net · 24/09/2026

I’m long “in short”

In <4 weeks you're gonna be so dead sick of
"in short" it's going to make your head spin

141965

Grace @gracekind.net · 23/09/2026

New animation from Opus 5.5!

6650084

Grace @gracekind.net · 22/09/2026

I feel like this thesis of mine is actively being tested. I hope it holds up!

I agree, I would also add: - we will have more chances to course-
correct and coevolve with Al than we think - more people will be drawn to alignment
work as it becomes more concrete

8670

Grace @gracekind.net · 22/09/2026

Waiting for this to happen to bsky

godoglyness declares the website Redeemed

020

Grace @gracekind.net · 22/09/2026

By the way, this isn’t to say that early LW thinking is completely vindicated. I agree with these points from @eudoxia.bsky.social:

4863

Grace @gracekind.net · 21/09/2026

Bayesian rationality and Rationalism are two different things!! This is rationalism: en.wikipedia.org/wiki/Rationalism This is rationality: en.wikipedia.org/wiki/Rationality The contemporary "rationalist community" is based around the latter concept, not the former.

2628

Grace @gracekind.net · 21/09/2026

Okay I was wrong, it’s a pig

250