#машинное-обучение
-
Google open-sources TorchTPU, a native PyTorch backend for TPUs
The company has released torch_tpu, a project that lets PyTorch models and operations run directly on Google's Tensor Processing Units. The code ships under the Apache 2.0 license.
-
An AI That Gets Better at Cracking CAPTCHAs the More It Tries
Researchers built a system that solves visual CAPTCHAs and grows more accurate with every encounter — even recovering after defenders scramble the images against it.
-
Hybrid agents: thinking in latent space, acting through text
Researchers propose HybridAgent, an architecture that switches between "latent" reasoning and text-based interaction — which, they report, speeds things up and saves tokens without sacrificing quality.
-
Byte language models: long sequences as a resource, not a burden
Researchers in Shanghai show that a plain Transformer can learn directly from raw bytes without a tokenizer — and, as parameters scale, overtake subword models.
-
A Transformer Scored 44% on ARC-AGI-1 for 67 Cents
Indian researcher Mithil Vakde trained a tiny model from scratch in an hour and a half on a single GPU and beat many large language models on accuracy — with no synthetic data and no recursion.
-
Startup Inherent: the small Faraday model beat Claude and GPT at reproducing scientific papers
London-based lab Inherent, founded by DeepMind alumni, has unveiled its AI agent Faraday, which — running on a compact model — outperformed the systems from Anthropic and OpenAI at independently reproducing the results of scientific papers.
-
A Model That Knows the World No Further Than Fifth Grade
Researchers trained language models on a corpus limited to the elementary-school curriculum in order to separate the acquisition of knowledge from its retrieval. The takeaway: the filter on the training data sets a ceiling on capability.
-
A $500 Fine-Tuned 9B Model Beat Top AIs at Catalog Review
The authors of Fermisense claim that an open 9-billion-parameter model, fine-tuned with reinforcement learning, outperformed leading proprietary models on a specific task and turned out to be dozens of times cheaper to run.
-
Xiaomi-Robotics-1: How to Work Around the Data Shortage for Robots
Xiaomi has unveiled a foundation model for robots that trains on 100,000 hours of “embodiment-free” trajectories and transfers the skills to real machines.