HNHacker News
TopNewBestAskShowJobs

E-Reverance

998 karma · joined November 11, 2020

submissionscomments

Make Quizzes with Your Friends

instagram.com·1 pts·E-Reverance·
1

Diffeomorphic Optimization

arxiv.org·2 pts·E-Reverance·
0

Parallel Language Models via Variational Decoding

arxiv.org·4 pts·E-Reverance·
0

University of Leicester to make redundant all pure math professors (2021)

reddit.com·24 pts·E-Reverance·
2

Next Concept Models

arxiv.org·4 pts·E-Reverance·
0

Correction Training for Diffusion Models

sunovivid.github.io·2 pts·E-Reverance·
0

Leading Chinese mathematicians unite behind use of AI in research

scmp.com·4 pts·E-Reverance·
0

The Unreasonable Effectiveness of Linear Prediction as a Perceptual Metric

arxiv.org·1 pts·E-Reverance·
0

Residual Matrix Transformers: Scaling the Size of the Residual Stream

arxiv.org·3 pts·E-Reverance·
0

Skate 3 in the Browser

skate.aaddpp.lol·2 pts·E-Reverance·
2

Discrete Fourier Transform faster than n log n [pdf]

github.com·2 pts·E-Reverance·
0

Integer multiplication below n log n

github.com·113 pts·E-Reverance·
83

Dust: Pretraining Transformers Without Backpropagation

qlabs.sh·280 pts·E-Reverance·
82

Single cortical neurons as deep artificial neural networks

sciencedirect.com·2 pts·E-Reverance·
0

SoftServe: A Scalable Quasi-Newton Method for Deep Learning

arxiv.org·1 pts·E-Reverance·
0

Chess, but it looks like doodles on paper

twitter.com·2 pts·E-Reverance·
0

Self-Play Pretraining with Zero Data

arxiv.org·4 pts·E-Reverance·
0

Learning to Discover Interesting Mathematics

arxiv.org·1 pts·E-Reverance·
0

Log-Depth Recurrent Language Modeling

arxiv.org·1 pts·E-Reverance·
0

Realtime Pathtracing in Octane

twitter.com·1 pts·E-Reverance·
0

94% on AIME with 1B Params

paradigma.inc·2 pts·E-Reverance·
1

86" 3D Display

lookingglassfactory.com·3 pts·E-Reverance·
0

Fluid Simulation in One Tweet

davidar.io·1 pts·E-Reverance·
0

Solving rubiks cubes "without search"

arxiv.org·4 pts·E-Reverance·
0

Jev might be Qwen finetune

nitter.click·2 pts·E-Reverance·
0

DeepSeek's Parent Company

high-flyer.cn·1 pts·E-Reverance·
0

Replicating Jev with Qwen-2.5

nitter.perennialte.ch·3 pts·E-Reverance·
0

Dart: Denoising Autoregressive Transformer

arxiv.org·1 pts·E-Reverance·
0

Discrete Beckmann Transport Models for One-Step Language Modeling and Reasoning

arxiv.org·1 pts·E-Reverance·
0

Grid-Free Monte Carlo for Time-Dependent Diffusion

arxiv.org·2 pts·E-Reverance·
0
Page 1 of 6Next →