Vision Transformers
2,039 posts
Vision Transformers
@vitransformer
Building in ML with blogs ๐
- New blog post by @AmanGokrani: Everyone says Claude Code "just works" like magic. He proxied its API calls to see what's happening. The secret? It's riddled with <system-reminder> tags that never let it forget what it's doing. (1/6) [๐ link in final post with system prompt]
- New blog post: We've never enjoyed working on Kernels more than this. We have some very fast AI-generated kernels with a simple multi-agent system. They're running close to or even surpassing Pytorch shipped kernels. (1/6) [๐ link in final post]
- Replying to @vitransformerFor builders: Spam <system-reminder> tags Front-load context before work Add reminders in tool results Use dynamic context injection The magic isn't intelligence. It's obsessive focus management. (5/6)
- Replying to @vitransformerMillion dollar question: Why does ONLY Anthropic use <system-reminder> this heavily? Special training signal? We don't know. But now you can steal their pattern. [Blog ๐] medium.com/@outsightai/peโฆ [Claude Code system prompt ๐] gist.github.com/agokrani/919b5โฆ (6/6)
- Blessed by sensei !!
New blog post: We've never enjoyed working on Kernels more than this. We have some very fast AI-generated kernels with a simple multi-agent system. They're running close to or even surpassing Pytorch shipped kernels. (1/6) [๐ link in final post] - Replying to @vitransformerThe punchline: "Tiny reminders, at the right time, change agent behavior" It's not a smarter model. It's a model that never lets itself forget. Even GPT-4 would work better with Post-it notes everywhere. (4/6)
- Replying to @vitransformerThese tags are EVERYWHERE: In prompts. In tool results. In user messages. Even in bash outputs. Example: One todo update generates 4+ reminders. It's engineered paranoia. Claude literally can't drift off-task. (2/6)
- Replying to @vitransformerBefore ANY work starts: Summarizes context (50 chars) Extracts topic (2-3 words) Validates commands for injection Spawns specialized sub-agents Every step has pre-flight checks + embedded reminders. (3/6)
- look ma i made it to batch 4 graduation @lossfunk
- week 1 x delta @theresidency our acceptance made us dedicated to ship and hereโs what this week full of ups and downs made us build : - launching the first ever The Kernel Company website: thekernelcompany.com -released blog on optimised FLUX.1-Kontext-dev - performance
- look mom i gave a talk!!Can AI explain itself? Lossfunk resident @vitransformer talks difference between explainability and interpretability, and how a ProtoPNets open the black box of neural networks. Video link below ๐
