Marco D'Agostini

@madacol.com

81 Followers 253 Following

madacol.com I love to automate stuffs, mostly using web pages nowadays

A WhatsApp conversation with an LLM agent, showing the agent become incoherent: ... 🤖 Do you have diabetes, or do you take a GLP-1 medicine such as Ozempic, Wegovy, Mounjaro, or similar? no 🤖 Have you had upper abdominal or chest surgery, or do you regularly take opioids or medicines with strong anticholinergic effects?numerusformassistant to=web.run_analysis code 腾讯天天中彩票? Wait need no tool? Let's continue; I mistakenly? Actually send final only. No need tool.
This image is a screenshot of a WhatsApp conversation, depicting a technical workflow for updating a bot's test suite. The chat includes the following key elements: * **Bot Message:** A message with a robot emoji explains that a 1-line JavaScript code block needs to be updated because a new rule requires a minimum of 5 lines for it to render as an image. * **File Edit:** An "Edit" notification displays the file path: `/home/mada/whatsapp-llm-bot/tests/sendBlocks.test.js`. * **Code Diff:** A visual diff shows a single line, `console.log("hello");`, being replaced by a multi-line `greet` function and a function call within a `markdown` constant. * **Bash Command:** A final message shows a command block for running tests using `pnpm test`, with a 60-second timeout.
Large language models are often used to answer queries grounded in large text corpora (e.g. codebases, legal documents, or chat histories) by placing the entire corpus in the context window and leveraging in-context learning (ICL). Although current models support contexts of 100K-1M tokens, this setup is costly to serve because the memory consumption of the KV cache scales with input length. We explore an alternative: training a smaller KV cache offline on each corpus. At inference time, we load this trained KV cache, which we call a Cartridge, and decode a response. Critically, the cost of training a Cartridge can be amortized across all the queries referencing the same corpus. However, we find that the naive approach of training the Cartridge with next-token prediction on the corpus is not competitive with ICL. Instead, we propose self-study, a training recipe in which we generate synthetic conversations about the corpus and train the Cartridge with a context-distillation objective. We find that Cartridges trained with self-study replicate the functionality of ICL, while being significantly cheaper to serve. On challenging long-context benchmarks, Cartridges trained with self-study match ICL performance while using 38.6x less memory and enabling 26.4x higher throughput. Self-study also extends the model's effective context length (e.g. from 128k to 484k tokens on MTOB) and surprisingly, leads to Cartridges that can be composed at inference time without retraining.
source
screenshot of a web music player