Emy Gervais
@emy_aze
Followers
71
Following
91
Media
25
Statuses
208
All the noise I’ve heard about how I’m “ruining my wife’s reputation online” all came from the exact same demographic. All of them came from white gen-x men. I will never let anyone disrespect my wife in public, if that bothers you, you can fuck off. -Emy G.
0
0
4
Number 1 on @huggingface while being single-author is WORLD CHAMPION, in my book.
1
0
15
TRMs confirmed to be able to do chatbots!!! In your face, openAI ! Great work, Markus!
Built a multi-turn conversation memory for our TRM mesh in 250KB. "what time is it in Tokyo" → "and the weather there?" The agent remembers Tokyo, carries the timezone across turns, and applies it to the weather query. No LLM needed. 📊 Benchmark: 80% turn accuracy,
0
0
10
BACK IN MY DAY IF YOU COULDNT CODE YOUR OWN OS IN COMPSCI 101 THEY TOLD YOU TO GET THE FUCK OUT
0
0
5
There's no reason why nvidia & co. should have all the monopoly for training AIs on GPUs. CUDA is a gatekeeper!! Magnus proves that you don't need 10x of the latest h800spromax 80k GPU to train a model from scratch. Impressive stuff!
I can legitimately say I've created one of the world's smallest, fastest, most accurate models to date. 580KB, 4s inference on CPU, 5,287/sec batch GPU, 148K parameters, 99.8% accurate. Praise the TinyRecursiveModel! You just have to know how to use it properly!
1
1
1
unpopular opinion : If you can't write code in assembly, can you really call yourself a computer scientist?
0
0
3
Look to @jm_alexia for better alternatives. I know for a FACT that her command prompt was going zoooooom just yesterday.
AI that isn’t run locally, or that can’t save memory locally, has no long-term future. It’s just a stopgap until a better alternative appears.
0
0
2
First day on the job : you have to train a llm from scratch on your laptop’s integrated gpu, you have 24h to beat all the top models or your entire family dies. What do you do?
6
1
6
Physicists are the astrology witches of STEM. I see you playing with your little laser beams and crystals.
0
0
1
I don’t mean anything by saying that abusive elephant training techniques are very effective when applied to a research scientist. Again, I dont mean anything by it, just an observation.
0
0
2
Sick of this CUDA bottleneck bullshit. Do it straight in assembly and train on the CPU.
1
0
6
I can’t wait to watch the patriot games. Royal rumble and Wrestlemania are gonna suck next year (they will cover up the women in 50*c heat cuz SA). Someone had to do something about this.
0
0
1
Next fundamental physical problem to solve, I WANT MY RAM TO COME IN 16 Tb STICKS.
0
0
1
We should commit, no more hard drive! No more SSDs! No more DRIVES!! Its about taking a stand! We do it all on RAM, pure RAM memory. WE COMMIT TO RAM!! These sticks have been around forever and they never failed us!! CLOSE ALL HARD DRIVE FACTORIES AND DIVERT PRODUCTION TO RAM
2
0
4
Recursive reasoning beats multi-billion-parameter models You can now easily train your own 7M param model from scratch and outperform DeepSeek-r1 on ARC-AGI 1 We provide a simple speedrun script that handles setup, training, and eval in one go.
32
205
2K
This chick won the Grand Prix 2025 arc-prize "Best Paper" award at neurips so she kinda knows what she's talking about.
I highly recommend people not to brainstorm with LLMs. See south park s27e3 which nails it perfectly.
0
0
4
Fuck falcons, they’re such assholes. I’m talking about these mf’ers. They’re all selfish, and HUge assholes. -not a fan.
0
0
2