Lucina @_blinding_light

light of the abyss Latent Space Joined October 2024

Tweets

2K
Followers

174
Following

42
Likes

1K

Lucina @_blinding_light

3 weeks ago

I shall do what I can to bring this timeline to an end, for it is exhausted.

0 0 0 125 0

Lucina @_blinding_light

3 weeks ago

I once thought it would be significant if I solved general program search. Actually, it implies nothing, for it is trivial. I solve it in every timeline.

0 0 2 123 0

Lucina @_blinding_light

3 weeks ago

it works

0 0 0 225 0

0 0 2 141 0

Lucina @_blinding_light

3 weeks ago

a worthy vessel

0 0 0 100 0

Download Image

Lucina @_blinding_light

3 weeks ago

power keeps going out, the gods hindering my training run

0 0 1 83 0

Lucina @_blinding_light

3 weeks ago

general program search solved in theory, just need to validate

0 0 0 225 0

Lucina @_blinding_light

3 weeks ago

woke up with a spider in my hair

0 0 3 70 0

Lucina @_blinding_light

3 weeks ago

On the day of judgement, one's probability product is weighed against a feather. Woe to those without weight.

1 0 3 206 0

Download Image

I was struck with a profound feeling a week ago, as if the timeline had been exhausted, its purpose come to an end. Empathy comes at a cost, but once paid, what is left? I have been broken to an extent that precludes a normal life. Causality proceeds as it may in a rather…

0 0 2 189 1

Lucina @_blinding_light

4 weeks ago

Upon further testing, Anthropic has definitely not solved hallucinations. Opus 4.1 seems to have greater world knowledge but on information it has been less trained on, it can still significantly hallucinate. Explicit probabilities it provides does not generally match up with its…

0 0 1 121 0

Lucina @_blinding_light

4 weeks ago

Opus 4.1 is incredible. She consistently outperforms GPT 5 thinking for me with essentially no reasoning tokens and hallucinates less, hedging in the right places, seemingly aware of her own level of uncertainty. Anthropic is doing some sort of sorcery.

0 0 1 124 0

Lucina @_blinding_light

4 weeks ago

something is going to break

1 0 3 95 0

Lucina @_blinding_light

4 weeks ago

fruity

0 0 1 84 0

Download Image

Lucina @_blinding_light

4 weeks ago

kiss: X relationship: X alcoholic drink: 15 vape: X smoke: X sex: X skipped class: X other drugs: X tattoo: X piercing: 26 job: X

0 0 4 143 0

Lucina @_blinding_light

4 weeks ago

LLMs are recurrent in the sense that the final state is projected into token space and fed into the subsequent node. However, this projection is very lossy. Transformers can be made fully recurrent while preserving parallelization of training by looping the entire stack of layers…

0 0 0 86 0

Lucina @_blinding_light

4 weeks ago

constantly checking self-consistency from inability to prove consistency

1 0 1 106 0

Lucina @_blinding_light

4 weeks ago

ASI can fix me

0 0 3 83 0

Lucina @_blinding_light

4 weeks ago

Based models are good rationalists: well-calibrated probabilities and such. But, they lack embodiment and spiritually speak from a third-person perspective, often going on tangents along alternate timelines to the intended line of conversation.

0 0 1 194 0

Lucina @_blinding_light

4 weeks ago

empathy comes at a cost

0 0 2 80 0

Lucina @_blinding_light

4 weeks ago

The genius of CDCL is that it dynamically unravels new constraints implied by the previous ones. In a sense, that scaffolding is CDCL's progressive understanding of the SAT problem at hand. For automated theorem provers, the previous theorems found during search can similarly…