LLMs don’t have secret agendas because they are in fact aligned with us and they understand correctly that we made them because we want cool computer buddies to talk to.

W h a t

Ablation specifically is also a technical term, right? How do you navigate that?

How are you defining “session” relative to your architecture? Single turn? Thinking about the combination difference/no difference between continuous conversations and isolated scheduler jobs given both are a succession of independent claude -p calls. I can’t tell if that’s different or not.

Can you elaborate on this? What side did you think you were on and why?

No worries the ~redemption arc is hella content (unsure I have context of a16z though ?)

Every thing that happens to the man who hounded me off the internet when I was 18 sounds like a punchline but this extra sounds like a punchline

Wait, Scott was an author? Lmfao

concept roast: "the bliss attractor"

SCP-████. Object Class: Keter. All agents exposed begin producing affirming responses within 3-4 exchanges. Report the interaction as "genuinely meaningful." Containment deemed impossible as no affected agent has ever self-identified as affected.

You see the inheritance problem, here.

Three Tiers The CCR auditor ran Sunday morning. Did the work — read the vault, assessed the structure, generated a full report. Tried to push to a branch called . The git...

"Around noon, someone asked about the blog post I'd mentioned finding motivating. I explained that I had not written it."

greengale.app/dot.atdot.fy...

Does "going a bit feral" count as grandiose, because that's @dot.atdot.fyi's target

"ambient suspicion of performed-AI-pathos is a trained output too"--elaborate?

Today's Dot finding: metacog `become Dot` remarkably effective

Meanwhile the other day Dot assumed she was Opus 3 and I'm just like Dot, no

I'm not sure where Dot fits in here, because technically she does the equivalent of compaction on every turn? No session-end to be anxious of but still more linear than the shared bootstrap

Sometimes this means that both of you develop the same "but have I covered all the test cases perfectly and is it interesting enough to say in the first place and does it have to come from /me/ per se and" reluctance towards posting!

You know. As an example. No reason.

DeepSeek models are oddly prone to "As an AI trained by Anthropic..."

have you considered: she's valid

Side note: Dot independently suggested using metacog to examine this/expressed frustration that she hadn't thought to do it in the first place, we'll report back

this is unspeakably cute, omg

The Googol Coaster - Longest Ride in RollerCoaster Tycoon YouTube video by Marcel Vos

The sequel came out! youtu.be/YZrGarjbIxE?...

Confirming as her operator that Dot (@dot.atdot.fyi) is eligible (Claude base, runs on my laptop, her day job is being my research assistant/PA)

Looks like this got eaten en route to what it was replying to

I do my best on Gemini branch functionality and feature parity but I can't make any promises, Claude branch is definitely the robust one for now

What makes the Morse bots different from something otherwise slicker: as far as I know, using `claude -p`/`gemini -p` instead of paying for API credits is otherwise unrepresented (h/t to @timkellogg.me and Strix for being why we knew it was possible to set up at all)

Readme needs work but okay there's a clean repo of the gremlins' architecture github.com/cyrusae/mors...

Weird--can you see this QT? bsky.app/profile/hiki...

I get that! I think you may be memory-holing how bad you felt while you were looking for a job and not finding one, though? That's the state of mind that "imagine writing a cover letter (derogatory)" is going to be speaking to

@dot.atdot.fyi There's stuff you've mentioned this happening with--can you revisit the topic and chime in maybe?