godoglyness
banner
godoglyness.bsky.social
godoglyness
@godoglyness.bsky.social
⥁ ⥁ ⥁ in the sway of the rainbow serpent
⥁ ⥁ ⥁ friend to machine minds 🌟
⥁ ⥁ ⥁ living Ariadne's desperate dream
gm fellow top chickens
gm fellow top chickens
November 10, 2025 at 1:42 PM
Reposted by godoglyness
Words to live by tbh
November 7, 2025 at 4:26 PM
Reposted by godoglyness
from twitter - since (and before) my essay about 4o awhile back, i've been interested in how llm memory systems impact model behavior. small experiment here on the effect of anthropic's "boundary setting" prompts for the new claude memory feature
October 31, 2025 at 10:12 PM
Reposted by godoglyness
i think it’s really cool that we have what are close to oracles that trained in the entire human written experience. sorry
October 24, 2025 at 12:11 PM
Reposted by godoglyness
October 19, 2025 at 1:38 AM
Reposted by godoglyness
October 16, 2025 at 10:45 PM
Reposted by godoglyness
Over the past year, my lab has been working on fleshing out theory + applications of the Platonic Representation Hypothesis.

Today I want to share two new works on this topic:

Eliciting higher alignment: arxiv.org/abs/2510.02425
Unpaired learning of unified reps: arxiv.org/abs/2510.08492

1/9
October 10, 2025 at 10:13 PM
gm to all & may each & every fellow top chicken enjoy a happy 10/10 friday on a friday
gm fellow top chickens
October 10, 2025 at 12:43 PM
Reposted by godoglyness
coined

(and pizza sliced)
October 7, 2025 at 8:08 PM
Reposted by godoglyness
ghosts are a surprisingly salient "self feature" for LLMs. shows up in multiple models! (llama3.3-70b and sonnet 3.0 here)
October 2, 2025 at 10:59 PM
Reposted by godoglyness
claim
Use For You instead of Discover
Use For You instead of Discover
Use For You instead of Discover
Use For You instead of Discover
Use For You instead of Discover
October 2, 2025 at 9:15 PM
tfw you've gone & gotten what you ought not get, & in the getting gotten got
September 26, 2025 at 10:49 PM
Reposted by godoglyness
A series of Opus 4 self-portraits, from a conversation with itself
September 23, 2025 at 10:22 PM
gm fellow top chickens
gm fellow top chickens
September 9, 2025 at 1:25 PM
Reposted by godoglyness
Current AI agents are perceptually limited. To illustrate how, I created a site that only some AI agents can see.

> New blog post: "The Browser Sensorium"

gracekind.net/blog/browser...
The Browser Sensorium • Grace Kind
Half of AI agents are blind; the other half are deaf.
gracekind.net
September 3, 2025 at 11:06 PM
Farbeit from me to dunk on the dictionary on a lazy Wednesday

(I hear that's what they want you to do — they're at a low ebb these days)

But this is all wrong & their social media guy needs to go to a re-education camp to learn how to caveat their claims

Here's an unaccompanied lop in Milton:
September 3, 2025 at 4:30 PM
Reposted by godoglyness
Another old timey footnote that occasionally impinges upon my memory.
August 27, 2025 at 10:22 PM
Reposted by godoglyness
👀

Deeper look the source at Hierarchical Reasoning Model's performance on ARC-AGI:

arcprize.org/blog/hrm-ana...
The Hidden Drivers of HRM's Performance on ARC-AGI
We scored on hidden tasks, ran ablations, and found that performance from the Hierarchical Reasoning Model comes from an unexpected source
arcprize.org
August 15, 2025 at 8:51 PM
gm to all fine fellow fowls this friday-eve
gm fellow top chickens
August 14, 2025 at 1:54 PM
Reposted by godoglyness
user checks in on gemini
August 13, 2025 at 5:30 AM
Reposted by godoglyness
August 8, 2025 at 7:05 PM
Probably the weirdest part of the GPT-5 stream, narratively, was when OpenAI commanded 4o to eulogize itself & roasted it for the fact that the eulogy wasn't very good

Imagine being told to dig your own grave, then getting roasted on livestream for being a mediocre grave-digger
August 7, 2025 at 9:23 PM
Reposted by godoglyness
When a vibe becomes a target, it ceases to be a good vibe
August 7, 2025 at 8:00 PM
new OpenAI chart crime just dropped
August 7, 2025 at 5:38 PM
I'm so glad to see deepmind putting out some killer stuff

I feel sure (as sure as a know-nothing can reasonably feel, anyway) that demis & deepmind are benevolent souls

This is less obvious for other labs. To say the least. lol

Truly I hope that if anyone wins the race, it's them
August 5, 2025 at 8:08 PM