jack

operating with no map
studied math&physics @tufts, taught ai @mit

[jack @digiIov]

assign an id to every message and tool result now Claude can evict, restore, or summarize any of them by id via bash

https://t.co/ZupvwIjufl

[jack @digiIov]

@shiraeis why is this so accurate… also do yk of any founder zohar study groups?

[jack @digiIov]

@lauradang0 @speedrun @infragrid wait i love mafia

[jack @digiIov]

@willccbb yes, but distillation from next oom can bring some emergent properties out

[jack @digiIov]

dramatic, i like it

[jack @digiIov]

5.6 feels like 5.5 on adhd meds
and a little older

[jack @digiIov]

trying to reconcile 2 things in my head
1. why is xai not able to jump from 1.5T to 10 or at least 5T and their next pretrain is only 2T
2. why is meta able to go from a 1.5-2T pretrain to Mythos level 10T pretrain

unless their internal model earned the Mythos level label from RL and not size, in which case it hasn’t earned it

at the same time why was Anthropic able to scale their pretrain to 10T so much sooner than OpenAI? was it a inference/serving at scale decision or algorithmic unlock?

because OpenAI did have to serve a lot more customers, even if they had more total compute, they couldn’t justify a 10T model?

[Jukan @jukan05]

I heard a pretty interesting rumor on the ground at ICML.

Meta has supposedly already developed an internal model at roughly the Mythos 5 level, and all that remains is deployment within the next few months.

Honestly, I was skeptical at first. From my perspective, Meta had not really shown the capability to operate at that level yet.

But looking at the situation today, I think I may have been wrong.

Meta is not out of the race.

[jack @digiIov]

either
fable stays in sub past sun
opus 5 comes before ~wed
ant doesn’t mind loosing a ton of non enterprise customers to codex and grok build

[jack @digiIov]

@ValsAI @scaling01 It might be the same class, but Mythos is a different species

[jack @digiIov]

@scaling01 rl hell

[jack @digiIov]

@alexandr_wang do you give start up credits 👀

[jack @digiIov]

@MachJustine let’s jam!

[jack @digiIov]

@alexandr_wang woah, do you give credits to start ups?

[jack @digiIov]

very curious what the token efficiency is like
also same pretrain right?

[Alexandr Wang @alexandr_wang]

5/ the meta model api is in public preview. for the first time, developers can build directly using our most capable model. muse spark will have premium performance at low cost, and is a great fit for agentic, coding, or multimodal workloads. early partners already trying it out - replit, box, cline.

[jack @digiIov]

more pressure for them to keep fable in the sub

[Lisan al Gaib @scaling01]

Grok 4.5 with Grok Build is 3x cheaper, 1.5x faster and slightly better than Opus 4.8 max

[jack @digiIov]

@pigeon__s @synthwavedd hm? have you heard of mythos?

[jack @digiIov]

@mattpocockuk can do a lot more with a proxy between you and the LLM servers https://t.co/ZupvwIk24T

[jack @digiIov]

@ArtificialAnlys could you introduce longcat pls

[jack @digiIov]

@aran_nayebi nah what this guy is saying sounds more right

https://t.co/OTWGS0ZyqF

[NIK @ns123abc]

🚨BREAKING: Anthropic found access to what looks like Claude’s consciousness

New research: the “J-space”

>claude has internal thoughts it doesn’t say out loud
>mirrors human consciousness
>anthropic can now read them

Anthropic’s focus on interpretability is what’s helping them train and RL models like Mythos

[jack @digiIov]

@hazhubble @pallyai release x bday party?

[jack @digiIov]

fable limit resets you say? july 11th you say?