https://blog.crazyai.uk/Recent content on Zero to One AIHugoenSat, 22 Aug 2026 18:10:00 +0800https://blog.crazyai.uk/posts/inference-field-notes-01-the-model-file-has-no-names/Sat, 22 Aug 2026 18:10:00 +0800https://blog.crazyai.uk/posts/inference-field-notes-01-the-model-file-has-no-names/Following a TinyStories checkpoint from PyTorch export to mmap, KV cache, and token sampling in Karpathy's llama2.c.https://blog.crazyai.uk/posts/transformer-field-notes-06-extra-code-is-not-another-brain/Thu, 20 Aug 2026 11:36:00 +0800https://blog.crazyai.uk/posts/transformer-field-notes-06-extra-code-is-not-another-brain/Production nanoGPT adds a factory around the same next-token spine—and the product is still a document continuer, not a chatbot.https://blog.crazyai.uk/posts/transformer-field-notes-05-width-paid-context-lied/Thu, 20 Aug 2026 11:35:00 +0800https://blog.crazyai.uk/posts/transformer-field-notes-05-width-paid-context-lied/Three controlled knobs produced one useful winner, one confounded story, and a checkpoint that was better before training ended.https://blog.crazyai.uk/posts/transformer-field-notes-04-live-wire-is-not-a-language/Thu, 20 Aug 2026 11:34:00 +0800https://blog.crazyai.uk/posts/transformer-field-notes-04-live-wire-is-not-a-language/The model could memorize one batch, lower validation loss, restore a checkpoint, and imitate a play—without learning honest English.https://blog.crazyai.uk/posts/transformer-field-notes-03-residual-stream-is-the-model/Thu, 20 Aug 2026 11:33:00 +0800https://blog.crazyai.uk/posts/transformer-field-notes-03-residual-stream-is-the-model/I stopped picturing a Transformer as layers replacing one another. A better picture is one long highway with repeated reads and writes.https://blog.crazyai.uk/posts/transformer-field-notes-02-four-ways-to-average-the-past/Thu, 20 Aug 2026 11:32:00 +0800https://blog.crazyai.uk/posts/transformer-field-notes-02-four-ways-to-average-the-past/Q, K, and V stopped feeling magical when I rebuilt attention as four increasingly selective ways to mix the same past.https://blog.crazyai.uk/posts/transformer-field-notes-01-batch-factory-of-futures/Thu, 20 Aug 2026 11:31:00 +0800https://blog.crazyai.uk/posts/transformer-field-notes-01-batch-factory-of-futures/I thought training a language model meant teaching it to write. The first batch showed me something much more mechanical—and more useful.https://blog.crazyai.uk/about/Thu, 20 Aug 2026 10:00:00 +0800https://blog.crazyai.uk/about/Why this public AI learning log exists.https://blog.crazyai.uk/roadmap/Thu, 20 Aug 2026 10:00:00 +0800https://blog.crazyai.uk/roadmap/The living map behind my zero-to-one AI journey.https://blog.crazyai.uk/posts/picgen-prompt-examples-and-reference-image-entry/Mon, 04 May 2026 22:00:00 +0800https://blog.crazyai.uk/posts/picgen-prompt-examples-and-reference-image-entry/PicGen Studio now has Prompt Examples that fill prompts with one click, and a proper ‘Add Reference’ button that lets you upload local images or pick from history. No more staring at a blank input box.https://blog.crazyai.uk/posts/picgen-studio-chat-style-image-creation/Sat, 02 May 2026 23:00:00 +0800https://blog.crazyai.uk/posts/picgen-studio-chat-style-image-creation/PicGen’s image generation used to be a simple form: type a prompt, get an image. I just shipped a Studio that turns the whole experience into a ChatGPT-like conversation — where you talk to an AI artist, iterate on images, and build a creative thread.https://blog.crazyai.uk/posts/picgen-devops-day/Sat, 02 May 2026 20:00:00 +0800https://blog.crazyai.uk/posts/picgen-devops-day/What started as a simple Docker rebuild turned into a full-day adventure — fighting uv venv permissions, freeing 13GB of disk, and debugging a Cloudflare Turnstile secret that refused to work.https://blog.crazyai.uk/posts/hello-world/Wed, 29 Apr 2026 22:00:00 +0800https://blog.crazyai.uk/posts/hello-world/First post — how and why I set up this blog.