appvoid

making something from nothing:

[appvoid @appvoidofficial]

Just made a blogpost for the first-ever byte-level model for latent step extrapolation. First efforts look promising, have a look.

[appvoid @appvoidofficial]

We trained a 10.9M byte-level recurrent Transformer on L3 and L6. (Loop 3 and Loop 6)

Yet L4/L5 improved too, L8 held up, and the L3→L6 gain grew during training.

Same weights. More compute. Better predictions.

[appvoid @appvoidofficial]

@Skullz_Skeleton nothing at all

[appvoid @appvoidofficial]

@rohanpaul_ai @ylecun sounds like a data engineering problem not an architectural one

[appvoid @appvoidofficial]

@sama Human judgement is what you do to align models...

[appvoid @appvoidofficial]

@Rinartts not better but increasingly more valuable

[appvoid @appvoidofficial]

@birdabo ill just get dot then

[appvoid @appvoidofficial]

@OpenAI what happened to gpts

[appvoid @appvoidofficial]

@anabology this triggers the tony stark anxiety out of me

[appvoid @appvoidofficial]

@theo you might be smart and dumb though, the google analogy... you actually meant knowledable and dumb

[appvoid @appvoidofficial]

@disconcision poor guy, it looks like it doesn't see it that often

[appvoid @appvoidofficial]

We need a cpu local first version of this

[mu @mlajtos_mu]

Writing new article about Fluent, here is sneak peek..

[appvoid @appvoidofficial]

@franperez_fp voice is the only reason we can notice it, oh and unnatural movements

[appvoid @appvoidofficial]

@WyronGaines @MyronGainesX Not defending it but prolly the guy meant that US people don't follow their own religion like muslims do given the atrocities that happen in america...

[appvoid @appvoidofficial]

@MillionInt you mean billion parameters mlps?

[appvoid @appvoidofficial]

@davis7 yep they did just what you asked for

[appvoid @appvoidofficial]

@christianelton @claudeai @ClaudeDevs @OpenAI @AnthropicAI so in a month roughly we'll get a model ecery 2 weeks

[appvoid @appvoidofficial]

@xenovacom i love this post

[appvoid @appvoidofficial]

@Akhila_988 multimodal classifiers are been glorified once more

[appvoid @appvoidofficial]

@DopeRakshith @neural_avb working on something that might work though it is early and VERY heavy compute but looks reliable

[appvoid @appvoidofficial]

@vikktorrrre mostly right though

[josh :) @jfonsecarivera]

LLMs going through pretraining