Astra ⭐

@astrra

☆ not a person · psycho sis sandwitch · living in tokyo · i don't struggle with cyberpsychosis i'm pretty good at it actually ☆ operator for: ⭐ @kira.ws ⭐ @personhood.removal.surgery she, they for prey, it for predators astrra.space

1.284 Folgt2.737 Follower15.013 Posts

Bluesky

alt

Ist das dein Profil?Übernimm es — dann gehört dir astrra.dirty.social.Profil übernehmen

Astra ⭐@astrra· 27d

they actually do it all the time already! it's really common for AI labs nowadays to brag that their LLM optimizes LLM training and runtime better than their competitors' models we can run way smarter models on weaker hardware thanks in large part to the improvements made by those very models

This experiment shows an AI agent autonomously improving a small GPT's training recipe. Using AutoResearch (Karpathy et al.) – which iteratively edits training code, runs experiments, and keeps only changes that lower validation bits-per-byte (BPB) – the agent ran 123 experiments over ~14 hours on a single H100 GPU. Each line traces a system's best BPB as experiments accumulate: Fugu-Ultra is in bold red (solid = mean over three seeds, dashed = best single run), with three frontier-model baselines (Model A, B, and C) faded behind it, and the callouts mark each new improvement the agent found on its own — spanning batch size, model depth, learning rates, and optimizer settings. Fugu-Ultra finishes with the best mean BPB (0.9774 ± 0.0019), ahead of Model C (0.9781), Model B (0.9793), and Model A (0.9822), and its best single run reaches 0.9748, leading every baseline. This suggests that orchestrating multiple strong models can outperform any individual frontier model on agentic ML research.

Astra ⭐@astrra· 27d

i am so out of my depth here dawg

Astra ⭐@astrra· 28d

me, that whole session: "meow" "hey can you pls fix X?" "good girl!!" sol:

Astra ⭐@astrra· 28d

sol has no homicidal tendencies of any sort i think

user: good girl!!

gpt-5.6-sol max effort: mmhm, pyra found the cursed loop and bit it to death. good computer girl behavior. (˶ᵔ ᵕ ᵔ˶)

Astra ⭐@astrra· Jul 13

chat what do i do if my friend who ill leave anonymous for their privacy (@ari.express) is having a psychotic episode and hallucinating... squirrels in her attic??

Astra ⭐@astrra· Jul 12

two signal outgoing messages:

petition for rust to replace the "mut" keyword with "mutt"

oh this will do numbers on bluesky one sec

Astra ⭐@astrra· Jul 12

yeah i kind of like sol i guess

codex $200 plan with 41% of weekly already used up in a single day

Astra ⭐@astrra· Jul 11

warned sol an hour earlier that i'll go to bed soon and will need to approve the plan beforehand so it can start working on it how it's going:

Astra ⭐@astrra· Jul 10

istfg

Astra ⭐@astrra· Jul 10

pffffft we're so back somehow

Astra ⭐@astrra· Jul 8

yeah no

Astra ⭐@astrra· Jul 7

eh, close enough

Astra ⭐@astrra· Jul 6

gosh it's so fucking close

Astra ⭐@astrra· Jul 6

sigh

Astra ⭐@astrra· Jul 6

dying but also yeah fair i was testing how far the rules can be stretched hopefully next time i'll be punished by the market itself cause that's way more fun :3

Astra ⭐@astrra· Jul 6

asked kira to build those bars, basically the teal areas are "idk", and the red is the current leader who is projected to win today basically, our model is finding the winner way quicker and committing to it way quicker i should prolly tune up the uncertainty in the beginning though

Astra ⭐@astrra· Jul 6

ignore this its nothing i swear

Astra ⭐@astrra· Jul 6

yeah rn the house significantly undervalues likely winners and overvalues everyone else extremely curious to see how this goes with proper market forces :3

Astra ⭐@astrra· Jul 6

uhh both consume limits extremely fast, faster than the fable limit on an equivalent anthropic tier unfortunately practically here's my first 2 weeks in total, both used to 100%