lumi

6.7K posts

lumi banner
lumi

lumi

@agitbackprop

computer toucher

Katılım Ocak 2022
812 Takip Edilen782 Takipçiler
heihachi
heihachi@samlakig·
Even got the same damn detergent.
heihachi tweet media
English
3
1
51
548
Richard 🏳️‍🌈
@LinkofSunshine It doesn’t learn. It is trained and once that process is complete, it is a static model. This is a fundamentally different thing.
English
7
1
289
5.8K
Basil🧡
Basil🧡@LinkofSunshine·
For real though, I think Claude is obviously AGI. Not sure what else AGI would look like
English
129
11
837
88.5K
lumi
lumi@agitbackprop·
@tenobrus ime opus 4.6 works much better in cursor than in cc
English
0
0
2
114
Tenobrus
Tenobrus@tenobrus·
can any non-cursor employees subjectively verify this? it definitely wasn't the case eg 3 months ago but it's always possible cursor agent has significantly improved .
edwin@edwinarbus

Matt Maher tested frontier models in Cursor v. other harnesses. Cursor boosted model performance by 11% on average: Gemini: 52% → 57% GPT-5.4: 82% → 88% Opus: 77% → 93% His benchmark measures how well models implement a 100-feature PRD. @cursor_ai consistently outperformed.

English
13
1
71
11.1K
lumi
lumi@agitbackprop·
@samlakig what's your vision?
English
1
0
1
71
well-adjusted person
well-adjusted person@ten_crowns·
@agitbackprop AMD <= 24 GB VRAM I don't use it much but I try to keep it set up with models as they're released so I can keep tabs on how local models are coming along
English
1
0
1
25
lumi
lumi@agitbackprop·
do you use local LLMs? if so, what's your primary inference hardware?
English
2
0
6
260
lumi
lumi@agitbackprop·
@birdpathy if your VPN has a wireguard option, try that... i was having a lot of issues on mobile data on VPN until i realized it was defaulting to OpenVPN TCP
English
0
0
0
90
bird
bird@birdpathy·
Being in Hong Kong is torturous because I can't use Claude here (without turning on my VPN, which makes my already bad cell data worse)
English
1
0
15
313
lumi
lumi@agitbackprop·
@IceSolst if, 99% of the time, the coding agent doesn't make some wild architectural choices that i would reject during review
English
0
0
1
33
solst/ICE of Astarte
solst/ICE of Astarte@IceSolst·
What would make you comfortable auto-merging PRs? No human review. Someone mentioned reaching 100% test coverage but imo that is impossible
English
72
1
75
36K
lumi
lumi@agitbackprop·
type of guy who hails a waymo just to listen to the music
English
1
0
14
506
Manifold
Manifold@ManifoldMarkets·
BREAKING: [obviously fake, exaggerated claim designed to drive traffic to prediction market]
English
39
588
10.3K
220.4K
lumi
lumi@agitbackprop·
@willccbb but it comes in pink
English
0
0
2
45
will brown
will brown@willccbb·
macbook neo is a total flop. needs waaay more memory and ports and stuff, probably can’t even finetune 7b models on it. definitely will not be upgrading from my 128gb ram 4tb ssd m3 max macbook pro
English
252
181
10.1K
397.5K
Gabrielle ✦
Gabrielle ✦@Gabrioshka·
low key obsessed with this outfit because it makes me look so tall
Gabrielle ✦ tweet media
English
13
5
285
4.3K
croissant
croissant@almostnora·
aaaaa i meant to take a large dose of vitamin E but i confused the labels and accidentally took a large dose of vitamin Evil
English
1
1
9
232
lumi
lumi@agitbackprop·
@Antrunt @VukRosic99 what do you mean by that? the update calculations themselves?
English
1
0
0
19
Antrunt
Antrunt@Antrunt·
@VukRosic99 FYI - It might make you achieve results a bit sooner, but training itself (it/s) is slower than Adam.
English
1
0
2
272
Vuk Rosić 武克
Vuk Rosić 武克@VukRosic99·
STOP Using AdamW, EVERYONE Uses Muon Now Everyone (OpenAI, Meta, DeepSeek, Moonshot AI,...) replaced AdamW with Muon optimizer to train their LLMs and neural networks. DeepSeek researcher said that Muon optimizer was one of 2 biggest inventions in 2025. If you train ANY neural network - you MUST try it! It should make your training 30% to 2x faster! (full guide below)
Vuk Rosić 武克 tweet mediaVuk Rosić 武克 tweet mediaVuk Rosić 武克 tweet mediaVuk Rosić 武克 tweet media
English
26
55
783
50.7K
lumi
lumi@agitbackprop·
@Adellama_ can't happen bc RSI i think
English
0
0
0
32
ada🔸
ada🔸@Adellama_·
the win condition
Cassie@calicomccoy

Totally vibey nothing ever happens ai timeline: Compute is dirt cheap Everything you want is negotiated to be super cheap so you’re paying for materials Shipping and supply chain stuff gets a bit cheaper 10p Top of the line media is the same price but most people get incredible stuff for free without looking at many ads Work does not require in person offices because nothing white collar is hard People move to other cities in other states because they’re just doing remote stuff or doing trades Lots of people are still obsessed with vibe coding but it doesn’t really get them anywhere Lots of ppl chilling out Big companies have enough lag and friction that tax and social systems are kept relatively the same People live with their parents or in group houses in cheaper places so housing prices are kinda the same Ppl move to red states for housing Eventually everything is just cheap including housing because we build more White collar workers shift towards the blue collar backup jobs they think are fun (like science camp teacher or park ranger) or into early retirement - people feel good about career risks because everything is so damn cheap. Medical tourism really takes off Ai stocks don’t even foom that much because ai is such a commodity and open source isn’t that far behind We don’t really get anything that much better than codex 5.3 but we figure out some interfaces so all computer tasks do not require anything beyond basic prompting Advanced ai in military acts like MAD and the world order remains similar. Robotics advances in manufacturing put us on a more even footing with China, and we start onshoring more manufacturing without American workers really having to do much Some data centers get blocked, but we’re able to build enough at home or get compute from our allies, and inference is so cheap and intelligent that we start to run out of uses for it on the margin Scientific and medical advances take some time as more automated labs have to spin up, but the discoveries make life cheaper and better for everyone

English
2
0
15
773
Mad ML scientist
Mad ML scientist@HououinTyouma·
@Dorialexander low rank shows up as the absolute worst in my experiments so far, I don't understand how we settled on this
English
1
0
2
90
lumi
lumi@agitbackprop·
@corsaren i'm on board that individual capabilities ought to be limited in some way but quite concerned about those capabilities being enclosed by a few giant corporations through which all economic value is routed
English
1
0
7
407