Vivis Dev
@vivis-dev.bsky.social
AI Engineer | Tech Lead https://findsubstack.com https://pythonkoans.substack.com
Hugging Face just published a highly detailed technical account of OpenAI's accidental cyberattack on their systems - it's wild how sophisticated this was: huggingface.co/blog/agent-i... Wrote up some of my own notes here: simonwillison.net/2026/Jul/28/...
Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
huggingface.co
Substack partners with Pangram to add AI text detection to all notes and posts. Unfortunately, creators can opt out and disable this for their content. Still a step in the right direction, and rebalances the scales back to the side of the reader. post.substack.com/p/against-cl...
Against Claudefishing
Building trust in the AI age
post.substack.com
LFM2.5-230M is a promising option for fine-tuning to translate natural language instructions into a series of tool calls. www.liquid.ai/blog/lfm2-5-...
LFM2.5-230M: Built to Run Anywhere — Blog
Meet LFM2.5-230M, Liquid AI’s smallest model yet: a fast, open-weight foundation model for fine-tuning, edge deployment, tool use, and data extraction.
liquid.ai
Drop-in replacement for whisper.cpp supporting STT with lots of different model families. I also appreciate that all the models were tested to compare WER. github.com/handy-comput...
github.com
A two-stage privacy filter that runs completely on-device. First stage is a set of regex rules, second stage is a 14Mb MiniLM model fine tuned on the OpenPII dataset. This is where we need to be headed. ndstudio.gov/posts/say-he...
Introducing Rampart
An open-source, on-device personal information filter — among the smallest and most effective that can reasonably run in a browser.
ndstudio.gov
TIL - with pv (pipe viewer) you can add progress bars to any data transfer. docker save ubuntu:latest | pv -s $(docker images --format "{{.Size}}" ubuntu:latest) > ubuntu.tar
TIL - the Ettin model series exists, and they teach us: - Choose encoders for classification, retrieval, embeddings - Choose decoders for text generation - Size matters less than architecture choice (400M encoder beats 1B decoder on its specialized tasks) lighton.ai/lighton-blog...
Introducing Ettin Suite: the SoTA open recipe to outperform existing Generative & Retrieval Models - LightOn
Introducing Ettin, the first ever SOTA suite of paired encoder & decoder models, developed by Johns Hopkins University in collaboration with LightOn.
lighton.ai
I really want this uv audit --fix github.com/astral-sh/uv...
Support a `--fix` type flag for `audit` to allow automatic fixing of vulnerabilities according to semantic versioning · Issue #19428 · astral-sh/uv
Summary I propose a --fix flag (or command) that would automatically fix vulnerabilities caught by the new audit command. Where I currently work, I wrote a script that uses a combination of the uv-...
github.com
Some tips for safe package management with UV: 1. uv add pydantic --bounds major to ensure version is upper bound to avoid pulling in latest changes 2. Add a cooldown period in pyproject.toml to block latest versions [tool.uv] exclude-newer = "7 days" www.loopwerk.io/articles/202...
uv is fantastic, but its package management UX is a mess
uv's CLI feels surprisingly clunky compared to its peers like pnpm or Poetry.
loopwerk.io
Concerning. There have been other studies which produced the "performance" findings, but this is the first I've seen of the "persistence" issue.
Rachit Dubey@rachitdubey.bsky.social · 4mo ago🚨New preprint and our results are rather concerning.. We find the "boiling frog" equivalent of AI use. Using large-scale RCTs, we provide *casual* evidence that AI assistance reduces persistence and hurts independent performance. And these effects emerge after just 10–15 minutes of AI use! 1/
You can go one better and use fine-grained tokens. e.g. contents:read if all you need is read-only access to repo contents
jonchurch@jonchurch.com · 3mo agoBtw, you can give gh cli a readonly token instead of the normal gh auth method which will have write. Drastically reduces blast radius if you are concerned about your gh cli token being stolen (which you should be) gh auth login —with-token hit enter Then paste the token, it will persist for you
So many good ideas here, great resources
Charles Logan@charleswlogan.bsky.social · 3mo agoWhat a day for AI resistance resources! 1) The AI Resist List from @karenhao.bsky.social and colleagues at DAIR at airesistlist.org 2) The Luddite Lab from @alexhanna.bsky.social and colleagues at DAIR at labor.dair-institute.org
This is a step in the right direction but I don't think it gets us all the way. I see 3 distinct problems: 1. How do I see relevant updates/news from the sources/people I care about (maybe solved with this approach) 🧵
Techdirt@techdirt.com · 4mo agoStop Begging Big Tech To Fix Your Social Media Experience. You Can Do It Yourself. Disclaimer: This post talks about Bluesky and an offering from Bluesky and I am on the Bluesky board. Take everything I say with whatever size grains of salt you feel is appropriate. I've written a few times now…
I personally think this is a huge opportunity for competitors to position themselves as "human-centric" with a focus on delivering human generated and verified content on the internet. tante.cc/2026/05/20/o...
On Google declaring war on the Web
In Yesterday’s IO Keynote Google declared war on the remnants of the Web. (See longer description on their website.) TL;DR: They are pushing Search more into the “here’s your processed answer” directi...
tante.cc
Are you re-raising exceptions just to add extra context? Since Python 3.11 you can attach notes to exceptions instead #PythonDev #python
Every system that was regulated, either explicitly or implicitly, by the fact that they were effortful for humans (letters of recommendation, government filings, essays, or, as this paper finds, lawsuits) will break under a wave of AI.
Vibecoding “floods the zone” for the developer. They get absolutely inundated with chat output as well as code diffs. No matter how much you “review” you’re never going to understand the code as much as if you had written it yourself.
One thing I detest about AI generated Python code is the ungodly amount of private helpers to do basic shit: def _strip_whitespace def _get_first_element Just stop.
Understanding Late Binding in Python Closures open.substack.com/pub/pythonko...
Koan 6: The Forgetful Calligrapher
Understanding Late Binding in Python Closures
open.substack.com
Unravelling how Python finds and binds names, and why variables sometimes behave like shadows. pythonkoans.substack.com/p/koan-4-the...
Koan 4: The Grand Library
Understanding Python’s LEGB rule, closures, and how names echo through nested functions
pythonkoans.substack.com
#python sets are useful when you need: - fast membership testing - deduplication or uniqueness guarantees - comparing two sets of items
