c. mcdonnell

138 posts

c. mcdonnell banner
c. mcdonnell

c. mcdonnell

@cmcdnd

cerebration proponent, mom

sf Katılım Ağustos 2025
583 Takip Edilen23 Takipçiler
Alpin
Alpin@AlpinDale·
My experience on the app has substantially improved
Alpin tweet media
English
1
0
31
612
Zach Krall
Zach Krall@zachkrall·
had an idea for previewing attachments
English
12
4
221
11.3K
Ethan
Ethan@torchcompiled·
At 16MB i feel like a winner would be absurdly sparse, thinking Fastfood Layers, pushing recurrence to the max, random projections which can be retrieved by storing seed, butterfly matrices, and exploiting ops like FFT. Feels like anything else would be a marginal gain.
Vuk Rosić 武克@VukRosic99

i did quick 71 experiments for 500 out of 13,000 steps for OpenAI's challenge 1. Mixture of Experts is absolute WINNER (very surprising as it shouldn't be for small LLMs) > Expert count matters most. 4 (best) > 3 >> 2. 2. UNTIED Embeddings work, tied are disaster 3. Depthwise Convolution - DEAD END Insights: 1. 4-expert MOE + leaky ReLU -> -0.048 BPB, clear winner 2. Untied factored embeddings (bn128) -> -0.031 BPB, worth combining with MOE 3. MOE + QAT combo -> preserves quantized quality for submission dead ends 1. Depthwise convolution -> every variant hurts, bigger kernels hurt more 2. Tied factored embeddings -> catastrophic, especially at small bottlenecks 3. Weight sharing -> not competitive with MOE for quality 4. Conv + anything combos — compounds the damage Next Steps 1. Validate MOE 4e + leaky at 2000-5000 steps, multiple seeds 2. Test MOE 4e + leaky + untied bn128 — the two biggest wins may stack 3. Full run (13780 steps) of best combo to see if it beats 1.2244 BPB leaderboard 71 experiments, 3 GPUs, ~500 steps each. Vuk Rosić 500 step training mainly helps us eliminate VERY BAD losers, winners need to be tested on longer training. Thank you @novita_labs for compute!

English
5
4
104
15K
c. mcdonnell
c. mcdonnell@cmcdnd·
@hoverdesign hyper constrained ML! final model can't be bigger than 16mb and training time is capped at 10min
English
1
0
1
57
Bryan Caplan
Bryan Caplan@bryan_caplan·
Don't believe in anything "bigger than yourself." Instead, become big enough for you to believe in yourself.
English
5
2
35
2.6K
Erica Levin
Erica Levin@bankof_amERICA·
Miami is cracked bc wdym this is my mid day walk view
Erica Levin tweet media
English
1
0
11
293
c. mcdonnell
c. mcdonnell@cmcdnd·
@viemccoy It's almost like we don't even have capital S Science without the faith & magic. Autism is not a fair replacement, we really need Pascals again.
English
0
0
2
55
𝚟𝚒𝚎 ⟢
𝚟𝚒𝚎 ⟢@viemccoy·
Please consider reading my newest article, Semiotic Triage: Overcoming the Type Error Tragedy. In it, I discuss how we might suspend our disbelief long enough to use faith as a generating function for enchantment, going beyond our present epistemic limitations.
𝚟𝚒𝚎 ⟢ tweet media
English
6
6
59
1.9K
Cat
Cat@CatOrman1·
when you throw a dinner party with someone you find out whether or not they’re an operator
English
30
41
1.8K
227.6K
eigenrobot
eigenrobot@eigenrobot·
everything 👏 is 👏 monocausal 👏 and 👏 specifically 👏 results 👏 from 👏 whatever 👏 shit 👏 I'm 👏 on 👏 about 👏 at 👏 any 👏 given 👏 time
English
146
1.3K
9K
0
c. mcdonnell
c. mcdonnell@cmcdnd·
@signulll They're annoying but zoom calls and panels are what we have left of French/English ritualized social theater. Cling on, we don't want them entirely gone.
English
0
0
0
31
signüll
signüll@signulll·
no offense but never has there been a moment in my life where i’ve wanted to hear a bunch of “panelists” speak at a conference lmao. like i get actively repulsed by that entire concept.
English
49
27
779
40.3K
Nate Fischer
Nate Fischer@NateAFischer·
So much AI hysteria is driven by the class envy and aspiration of public intellectuals: Basically if it’s an existential threat like the nuclear bomb, it must be controlled by priests/intellectuals and their client managerial class. But if it’s just a very powerful tool, it will simply increase the power of crass capitalists.
English
29
38
403
173.5K
Jeffrey Emanuel
Jeffrey Emanuel@doodlestein·
@IterIntellectus I just mix this into milk with my creatine and some vanilla syrup. Very painless and easy.
Jeffrey Emanuel tweet media
English
12
0
98
21.4K
vittorio
vittorio@IterIntellectus·
if you read enough fiber research you either start thinking you’re going insane or realize the entire nutrition industry has a financial incentive to keep you buying protein powder instead. you need to eat your fiber and i’m not even trying to be nice about it. 95% of Americans are *deficient* while enough fiber alone reduces all-cause mortality by 30%, colorectal cancer by 26%, and your gut bacteria literally digest your own intestinal lining when you starve them of it. fiber should be a bigger supplement category than protein but since it actually makes you healthier there’s no interest in selling it
Zero HP Lovecraft@0x49fa98

The more you look at actual nutrition and health science, the more fixated you become on fiber intake

English
336
1.1K
18.5K
1.4M
Sterling West
Sterling West@sterling_west24·
What does the e in e-girl stand for?
English
14
0
27
32.4K
c. mcdonnell
c. mcdonnell@cmcdnd·
@owroot not a completely fair comparison because it's regular vs. semilight (I don't have regular Times Now) But look at the refreshed Times- so beautiful
c. mcdonnell tweet media
English
0
0
25
745
O.W. Root
O.W. Root@owroot·
I love having all my email in Times New Roman. On the occasion that it flips into something else due to copy and paste, I am aghast. Typing emails in Times New Roman I basically feel like this image
O.W. Root tweet media
English
26
18
898
21.2K