Articles Tagged “Machine Learning”
10 articles found
On-Device Piano Model Autocompletes Music on iPhone
A 125M-parameter transformer generates about 108 piano notes per second on an iPhone 15, trained on 300 million note events with no cloud call.
AI Agents Reproduced 2,226 ICML Papers in Just 19 Days
A Hugging Face community challenge used AI coding agents to audit 2,226 ICML 2026 papers, verifying claims across 6,816 public reproduction logbooks.
CuspAI Raises $450M to Speed Up Materials Discovery
CuspAI raised $450M at a $2.6B valuation and launched an AI Materials Foundry with 45+ partners including NVIDIA, Meta, Samsung and Lam Research.
SAP Closes Prior Labs Deal on Tabular Foundation Models
SAP completed its Prior Labs acquisition at over 1 billion euros and will invest another 1 billion by 2030 in open tabular foundation models.
MIT's GIFT Turns 2D Designs Into CAD at 20% Compute
MIT's GIFT system converts a 2D image and text into executable CAD code using about 20% of the compute rival methods need, with no human labeling.
PyTorch 2.13 Brings FlexAttention to Apple Silicon
PyTorch 2.13 landed July 8 with FlexAttention on Apple Silicon — up to 12x faster attention on Mac GPUs and 4x lower memory for LM training.
Hugging Face Hub 1.18 Adds an AI Skills Marketplace to Its CLI
The new huggingface-hub 1.18 release introduces an 'hf skills' command, surfacing a growing marketplace of reusable AI skills for the open-source community.
Tufts Researchers Build AI That Uses 1% of the Energy and Outperforms Neural Nets
A Tufts University neuro-symbolic AI achieved 95% accuracy on complex reasoning tasks while consuming just 1% of the energy of conventional deep learning systems.
MIT Researchers Develop a Proxy Model Technique That Doubles LLM Training Speed
A new MIT method uses a lightweight proxy model to predict reasoning outputs, cutting the reinforcement learning rollout bottleneck in half.
New Self-Distillation Technique Triples LLM Inference Speed With a Single Model
Researchers achieve 3x faster LLM inference by baking multi-token prediction directly into model weights — no draft model or extra hardware required.