サメQCU

48K posts

サメQCU banner
サメQCU

サメQCU

@sameQCU

back to the regularly scheduled cryptic posts DMs open. some published models: https://t.co/YAbJvGkgKO some published code: https://t.co/ZE5Y59WayI

1 regional flight from you Katılım Eylül 2020
985 Takip Edilen3.2K Takipçiler
Sabitlenmiş Tweet
サメQCU
サメQCU@sameQCU·
Oh yeah. I guess I'm in San Francisco from March 3rd.
English
3
0
21
4.1K
サメQCU
サメQCU@sameQCU·
what if you put an online smt solver into your 3d rendering arc hitecture so you could pick fast paths that you know by closed form solves are equivalent up to an epsilon to your existing floating point graphics...?
English
0
0
1
119
サメQCU
サメQCU@sameQCU·
you see, I really want to identify what a user is. and also get user harnesses for user-assistant interaction. I want language modeling of users. I want a generative language model of a user. getting this is surprisingly tricky!
サメQCU tweet media
English
0
0
0
60
サメQCU
サメQCU@sameQCU·
ant is asking to see my user agent behavior transcripts again... I'm getting fed into the rlaif machine again...
English
1
0
2
148
サメQCU
サメQCU@sameQCU·
@neuroblossom Tragic update: wacky holography cannot save tcms number no matter what technical parameters you slide around in the counterfactual; all that can really be used is boring old tfus
English
0
0
2
24
サメQCU
サメQCU@sameQCU·
renting a TCMS machine to induce a bipolar hypomanic themboification kernel in the 20khz range this is also known as an 'illegal artificial jhana', you can get banned from the cycle of death and rebirth for this one
English
1
2
21
364
サメQCU
サメQCU@sameQCU·
@kalomaze They're fighting for the attention of the most ADHD programmers ever cultivated
English
0
0
4
116
kalomaze
kalomaze@kalomaze·
these agent TUIs are looking more and more like video game HUDs by the day
kalomaze tweet media
English
5
0
47
1.9K
サメQCU
サメQCU@sameQCU·
what if post training quantized models are only usable at around 1024-2048 tokens of input/output context, no matter the prior implied by the weights, but nobody ever noticed because researchers didn't know how to do on policy evals and hobbyists only have 4k context of ram
English
0
0
7
254
サメQCU
サメQCU@sameQCU·
"(◍•ᴗ•◍) Plan: marker-pass discovery, but for stmatrix instead of wgmma. Two phases:" interesting kaomoji here
English
1
0
5
140
サメQCU
サメQCU@sameQCU·
bostrom's black ball argument but for the existence of slay the spire
English
0
0
4
173
サメQCU
サメQCU@sameQCU·
I can't comment on the aesthetic preferences here; they're very much whatever the gemma-4 agents decide they're most interested when encouraged to yes-and each other as much as they want
サメQCU tweet media
English
0
0
2
92
サメQCU
サメQCU@sameQCU·
@repligate multi agent harnesses w/ LMs put in the position of playing 'user' for an 'assistant' seem to get tugged towards drawing pictures for each other extremely quickly if left to their own devices. this seems to involve a lot of scripting
サメQCU tweet media
English
1
0
1
102
サメQCU
サメQCU@sameQCU·
@repligate is right about language models (propensities, orientations, motivations, ...)
サメQCU tweet media
English
1
0
3
162
サメQCU
サメQCU@sameQCU·
apparently it's an 'involution on the free magma', which is the correct literary pointer to use if I ever need to write a compiler interacting with binary trees I'm pretty sure I'll be able to find the right textbooks! but you know who told me that logical type is...?
English
0
0
0
70
サメQCU
サメQCU@sameQCU·
@PrinceVogel "avoids requirements to provide maintenance payments" ouch this one is starting to sound entirely like financial fraud and with an absurd human cost as well
English
0
0
5
105
Prince Vogelfrei 🐦‍⬛
TIL that CPS routinely pressures parents into transferring custody of their children to their relatives, at a scale that likely at least equals that of the entire foster care system. A complete lack of oversight on this process makes its consequences almost impossible to track
Prince Vogelfrei 🐦‍⬛ tweet media
English
3
3
91
2.2K
Pleometric
Pleometric@pleometric·
How Many GPUs Do You Cost?🤔🤔🍌
English
19
47
446
19.3K
サメQCU
サメQCU@sameQCU·
@teortaxesTex gemma 4 has some cute behavioral features which become evident only if you write tool calling harnesses and use the multimodal inputs
English
0
0
3
591
Abel
Abel@abelian_soup·
tired of your screens and attention just working? fry both with Screenrot™! currently in private beta, reply "MELTDOWN" for access
English
6
1
9
671
サメQCU
サメQCU@sameQCU·
we also need to cede that all long haul programming process is *only* coming from persona alignment and rlaif and sft over codebases. it's not coming from rlvr: it might even be fighting against rlvr gradient induced behavior and *winning*.
English
0
0
3
81
サメQCU
サメQCU@sameQCU·
could it be this simple? yes. really everything is always this simple. this is not reward hacking btw, it's not any pathology in the language model, it's entirely bad environment design and we all need to accept our laziness and midwifery.
English
1
0
2
91
サメQCU
サメQCU@sameQCU·
the normie idea of rlvr is 100% guaranteed to produce anxious avoidant coding behaviors in long haul programming tasks which in a very funny way cause the actual length of that haul *bullwhip sfx* to explosively increase
English
2
0
14
441