ML Frameworks

FractalCore schedules training tasks across mobile nodes, averages their checkpoints with FedAvg, and pays them in "Liquid MBs" — plus it slices Llama-class models into phone-sized pieces.

VisionHOPE is a PyTorch visual backbone built on self-referential nested learning, where what the model remembers and how it learns co-evolve while it processes a single image.

It replaces autoregressive LLM classification with a single parallel forward pass, cutting latency from seconds to sub-20 ms while staying local.

TrenTorch teaches machine learning by having you build a PyTorch-shaped framework from scratch, plus 350+ browser-graded coding questions.

One framework to run RL post-training on diffusion, autoregressive, and hybrid models without rewriting the orchestration layer each time.

To prove that a 2.78-trillion-parameter model can run on a single CPU with 8 GB of RAM and no GPU.

nnsight gives researchers a Pythonic way to intercept, inspect, and rewrite the hidden activations of deep learning models as if they were ordinary variables.

OmniChar Studio packages a character's face, body, and wardrobe into a portable `.char` file that stays consistent across models — and trains the LoRAs to back it up.

A training framework built for the reality where inference and learning run at different speeds, with hand-optimized kernels for the largest MoE models.

To house the code, checklists, and profiling recipes for an O’Reilly guide that treats AI performance as a full-stack systems problem rather than a benchmarking contest.

DanKS is Kingsoft’s attempt to master a four-player Chinese card game using structured retrieval and PPO self-play, shipped as a complete research lineage.

ManiSkill 3 is a GPU-parallelized robotics framework built on SAPIEN that turns a single RTX 4090 into a data factory pumping out visual training data at 30,000+ FPS.

A world-model RL algorithm that reportedly needs no tuning across Atari, Minecraft, and continuous control.

Fizgig trains LoRAs on 8 GB cards, but its real pitch is the workbench: repair, remix, and profile LoRAs you already have.

DataWhale's free Chinese-language course takes you from tensors to ONNX deployment, with source-code readings of ResNet and Swin Transformer along the way.

PySR hunts for human-readable equations that fit your data, using a Julia-backed genetic search engine wrapped in a scikit-learn interface.

Most gradient boosters treat categorical features as an afterthought; CatBoost claims to handle them natively while delivering best-in-class prediction speed.

Halo scales LLM and multimodal training from one GPU to multi-node clusters by wrapping standard Hugging Face models in-place instead of maintaining a separate distributed rewrite.

Metaflow wraps the entire ML lifecycle—from notebook experiments to cloud-scale production—in a single Python framework so teams can iterate without rewriting glue code.

To demystify reasoning LLMs by coding inference-time scaling, RL, and distillation on top of a pretrained model, step by step.