Loading activity…
@__tinygrad__
We make tinygrad; sell tinybox for the GPU middle class. Our mission is to commoditize the petaflop.
AI
The reported setup used the default SGLang recipe, with roughly 60 tokens per second as the K3 comparison.
AI
The exabox’s refundable preorder deposit is $100,000; Tinycorp’s storefront marks preorders out of stock.
AI
Anthropic updated its Usage Policy to explicitly prohibit reselling, proxying, or unauthorized access to Claude through consumer subscriptions (Pro/Max plans), targeting services pooling subscriptions for API-like access.
AI
Anthropic allegedly lets users verify with a Chinese passport before instantly banning them; another commenter calls this a conspiracy theory.
AI
The tinybox red kimi pairs 16 TB of storage with 128 GB of GPU memory, tinygrad says.
AI
Claude says Max stays $100 a month and Ultra $200; weekly limits remain unchanged.
AI
The Information links Moonshot’s Kimi K3.1 Pro and DeepSeek’s V4.1-Flash 1017 to OpenAI’s model, Reuters reports.
AI
A user says their calculations make buying either rack hard to justify and has requested prices from several sellers.
AI
Tinygrad’s October 16 announcement invited buyers to get one.
AI
A post calls accelerator companies’ focus on inference and tokens per second at the expense of fundamentals a serious mistake.
AI
The tinygrad account says weeks six through 10 will focus on GPU usage and performance.
AI
Meta says Muse Wonder launched through its Model API and Muse Code, while Full was available to Muse Code subscribers.
AI
A post asks whether readers want to deploy exaboxes while quoting another post, without adding details.
AI
Tinygrad says AMD also sent it boxes and calls its software “quite usable” now.
AI
The tinygrad account says it used the G5/192/4TB machine for tinygrad development all summer.
AI
Tinygrad asks readers to identify the PyTorch equivalent of the UIR example it linked.
AI
The tinygrad account says the Pittsburgh class will meet three mornings a week and lectures won’t be posted online.
AI
Tinygrad said in X posts that it finished an AMD MLPerf contract with all four targets hit, criticized Tenstorrent, Qualcomm and Intel, and separately argued for a far smaller datacenter hardware design.
AI
OpenAI says it drew on an independent mathematics advisory group’s advice about how to release the results.
AI
The poster says compiler performance is graded on a curve, with the bottom performer getting 0 points and the top getting 15.
