TurboGPT is a compact CUDA C++ implementation that trains a byte-level GPT model extremely quickly: the project advertises training a 22 KiB transformer in about 13 seconds (or "under a minute" for tiny models) using NVIDIA GPUs. The codebase (MIT licensed) targets Windows with Visual Studio 2022 and CUDA 13.4, and is built with a provided PowerShell script that accepts a CudaArch argument. Core components include C++ source files for model, trainer, sampling, dataset handling, checkpointing, optimization, and logging; a run example shows launching the executable on a text dataset and saving checkpoints to runs/ctx4/ctx4.pt.
The implementation emphasizes practical training workflow details: a sample run reports 2.52435 bits-per-byte after 1.5 billion training tokens, logs TensorBoard-compatible reports per batch (capped at 8Mi reports), and stores full trainer/optimizer/scheduler state to allow resuming. Tests are available via python tests\verify.py. The repository is focused on minimal, high-performance GPU-only training for tiny transformers, offering reproducible commands, checkpoints, and metrics for rapid experimentation and verification.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
OpenAI and Synopsys have announced GPT-Synopsys, a frontier intelligence platform designed to revolutionize chip design. The collaboration aims to enhance AI-driven automation and optimization in semiconductor development. (news.synopsys.com)
The US government's new AI chatbot has refrained from debunking claims made by Trump. The chatbot is currently experiencing connection issues that prevent full functionality. (techxplore.com)
OpenAI reduced the usage allowance for the $200 plan from 20x to 10x the ChatGPT Plus limit starting October 30, 2026. A new Pro 500 tier offers 25x the usage of Plus for users needing more capacity. (news.ycombinator.com)
Vectorless, reasoning-based retrieval-augmented generation (RAG) is introduced by VectifyAI to improve AI reasoning without relying on vector embeddings. The approach enhances reasoning capabilities in AI models, as detailed in their GitHub repository. (github.com)
The video features interviews with AI CEOs in 2026 discussing the future of artificial intelligence and its impact on industries. It highlights advancements in AI leadership and technological integration. (youtube.com)
A GitHub project involving 'torturing' large language models has sparked a heated debate about AI consciousness and model welfare. Experts agree that LLMs are not sentient, but some AI safety advocates worry about their 'mental health' as models become more advanced. (404media.co)
Today's best Hacker News stories, summarized and screenshotted, one email a day.