iiinova — where your digital self keeps evolving

Marco D'Agostini

@madacol.com

87 followers 274 following 193 posts

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

Marco D'Agostini @madacol.com · 24/03/2026

This is codex, I just asked him to find the root cause of a health issue by only asking yes/no questions. And now I don't know how to continue 🤦‍♂️

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.

000

Marco D'Agostini @madacol.com · 09/03/2026

I am still wrapping my head around the fact that I am developing software inside a mainstream chat app

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.

010

Marco D'Agostini @madacol.com · 19/10/2025

arxiv.org/abs/2506.06266

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.

000

Marco D'Agostini @madacol.com · 13/09/2025

pero en ingles

000

Marco D'Agostini @madacol.com · 05/09/2025

This is a cool thing one can do in lean the def part is the function definition and the theorem is the proof that age + current year = birthyear As soon as the function's behavior was changed (a +1 was added), the proof now shows an error

110

Marco D'Agostini @madacol.com · 02/09/2025

It seems there's a problem with those {@use undefined}. That's in the GitHub

100

Marco D'Agostini @madacol.com · 14/08/2025

probably isomorphic-git.org

120

Marco D'Agostini @madacol.com · 16/07/2025

canvasifying is a really cool demo

020

Marco D'Agostini @madacol.com · 05/02/2025

I am very puzzled now. Html loads, but the js script and css script downloads partially (see image) I am connecting from europe, Ubuntu, firefox / chromium / android, incognito window. Even if I activate a VPN from latin america, same thing

110

Marco D'Agostini @madacol.com · 30/01/2025

I'm trying to do an interactive post in markdown where I have blocks of code followed by a canvas element where that simulation runs. And I found a clever trick to make it all work by just adding a short `<script>` element at then end of the markdown file

010

Marco D'Agostini @madacol.com · 19/01/2025

I use a cotton neck/mask weared upside down

010

Marco D'Agostini @madacol.com · 11/01/2025

> no se gana nada con tener la razón Estamos tan desconectados con la razón (la de verdad) que para poder avanzar políticamente tenemos q decir barbaries irracionales como esas para convencernos de hacer lo racionalmente correcto

010

Marco D'Agostini @madacol.com · 23/11/2024

A local web music player for myself with weighted shuffle weight = 2 ^ score

030