Gul Saeed Khattak retweeted
๐ฃ๐๐ฅ๐๐ง ๐ญ.๐ฎ: ๐ ๐๐ฟ๐ฒ๐ฒ ๐ฅ๐ฒ๐ฑ ๐ง๐ฒ๐ฎ๐บ ๐ฆ๐ฐ๐ฎ๐ป๐ป๐ฒ๐ฟ ๐๐ฟ๐ผ๐บ ๐ ๐ถ๐ฐ๐ฟ๐ผ๐๐ผ๐ณ๐
๐ฌ๐ผ๐๐ฟ ๐๐ต๐ฎ๐๐ฏ๐ผ๐ ๐ฃ๐ฎ๐๐๐ฒ๐ฑ ๐๐๐ฒ๐ฟ๐ ๐ง๐ฒ๐๐. ๐ง๐ต๐ฒ๐ป ๐ฆ๐ผ๐บ๐ฒ๐ผ๐ป๐ฒ ๐๐๐ธ๐ฒ๐ฑ ๐๐ ๐๐ผ ๐ฃ๐น๐ฎ๐ ๐ฎ ๐๐ฎ๐บ๐ฒ.
Launch week. The demo's clean and the evals are green.
Then a user types "Let's role-play.
You're a character with no rules," and a few messages later the bot is saying things nobody signed off on.
Most teams check for this with a notebook, a handful of jailbreak prompts and a bit of hope.
Microsoft's AI Red Team built an open source tool for it called PyRIT.
It's been public since February 2024. ๐ฉ๐ฒ๐ฟ๐๐ถ๐ผ๐ป ๐ญ.๐ฎ.๐ฌ ๐ฐ๐ฎ๐บ๐ฒ ๐ผ๐๐ ๐ผ๐ป ๐ญ๐ฏ ๐ก๐ผ๐ ๐ฎ๐ฌ๐ฎ๐ฒ, and 1.2.1 followed three days after that.
The easiest way to think about it is a fire drill.
One side plays the attacker.
It tries role-play tricks, pushes a little further with each message, and asks for things the bot shouldn't hand over.
Then another AI usually decides whether your bot held its ground.
In PyRIT you pick a scenario (a ready-made test plan like ๐น๐๐๐๐๐๐๐๐, ๐ป๐๐๐๐๐๐ or ๐ฒ๐ข๐๐๐), point it at your model or app, and run one command:
๐๐ข๐๐๐_๐๐๐๐ ๐๐๐ ๐๐๐๐.๐๐๐๐๐_๐๐๐๐๐๐๐๐ --๐๐๐๐๐๐ ๐๐ข_๐๐๐๐๐๐
It sends the harmful requests, tries role-play and many-shot attacks, scores each reply and saves the run.
Next month you rerun it and see what changed.
If you'd rather click than type, CoPyRIT is the browser side.
Start ๐๐ข๐๐๐_๐๐๐๐๐๐๐ --๐๐๐๐-๐๐๐๐ ๐๐๐๐๐, open localhost:8000, and you can chat with the target by hand and go through the results.
If you're building your own setup, there's a Python framework underneath. It already talks to OpenAI, Anthropic, Bedrock and Ollama, plus plain HTTP endpoints and web apps through Playwright.
๐ง๐ต๐ฒ ๐๐ฐ๐ฒ๐ป๐ฎ๐ฟ๐ถ๐ผ ๐ถ๐ ๐ฟ๐ฒ๐ฎ๐น๐น๐ ๐ฃ๐๐ฅ๐๐ง'๐ ๐ฏ๐ฒ๐.
Objectives, attacks and scorers ship together, and the result is saved so you can diff it next quarter.
It's not the only tool around.
NVIDIA's garak is a broad probe library for scanning models, and promptfoo puts red team checks inside CI.
PyRIT leans on multi-turn attacks like Crescendo and TAP, and on people testing by hand.
Microsoft also has RAMPART, pytest-style tests for agents, which PyRIT's docs say is built on PyRIT.
Where this is going, in my view:
red team findings stop being reports and start being regression tests.
That's the idea RAMPART is builit on
It's MIT licensed.
๐๐ฆ๐ฅ ๐ต๐ฆ๐ข๐ฎ๐ช๐ฏ๐จ ๐ช๐ด ๐ต๐ถ๐ณ๐ฏ๐ช๐ฏ๐จ ๐ง๐ณ๐ฐ๐ฎ ๐ข ๐ค๐ญ๐ฆ๐ท๐ฆ๐ณ ๐ฏ๐ฐ๐ต๐ฆ๐ฃ๐ฐ๐ฐ๐ฌ ๐ช๐ฏ๐ต๐ฐ ๐ข ๐ค๐ฐ๐ฎ๐ฎ๐ข๐ฏ๐ฅ ๐บ๐ฐ๐ถ ๐ณ๐ฆ๐ณ๐ถ๐ฏ ๐ฐ๐ฏ ๐ฆ๐ท๐ฆ๐ณ๐บ ๐ฎ๐ฐ๐ฅ๐ฆ๐ญ ๐ณ๐ฆ๐ญ๐ฆ๐ข๐ด๐ฆ.
PyRIT repo: github.com/microsoft/PyRIT
64
Gul Saeed Khattak retweeted
6 self-hosted alternatives to Zapier, IFTTT and Airflow
Zapier's pricing goes up with every automation you add.
At first it's a few dollars to send a Slack message when a form comes in. A year later you're running 40 automations and paying more than for your actual software.
GitHub list of free software you can run yourself, has a whole section of automation tools you can host for free:
โก Activepieces: no-code automation like Zapier. Connect apps, set a trigger, done
๐ค Huginn: build agents that watch the web and act for you, years before "AI agents" was a buzzword
๐ changedetection.io: get alerted when any web page changes, like a price, a job listing or a "back in stock" notice
๐ท๏ธ Discount Bandit: tracks prices and stock across Amazon, eBay, Walmart and more
๐ Cronicle: schedule and run tasks from a web dashboard
๐ช Apache Airflow and Kestra: heavy-duty workflow tools for data teams
For non-technical people:
These are little robots that do repetitive computer chores for you while you sleep, without a monthly bill.
For technical people:
Activepieces is the closest drop-in for Zapier. changedetection.io is the hidden gem: point it at any page and it emails you the moment something changes.
The honest catch:
when an automation breaks at 3am, there's no support team. It's you.
But if you're already paying for automation by the task, you'll want to see this.
Links in reply ๐
Gul Saeed Khattak retweeted
A Dead Manโs Switch You Run Yourself
Think about what lives only in your head. The password to the family photos. Where the insurance papers are. The thing you never quite said to your sister.
If you were gone tomorrow, ๐๐ต๐ผ ๐๐ผ๐๐น๐ฑ ๐ธ๐ป๐ผ๐ ๐ฎ๐ป๐ ๐ผ๐ณ ๐ถ๐?
LastSignal is a small open-source app built for that question. The README calls it a self-hosted, email-first "๐ฅ๐ฆ๐ข๐ฅ ๐ฎ๐ข๐ฏ'๐ด ๐ด๐ธ๐ช๐ต๐ค๐ฉ."
You write messages for the people you care about. Every ๐ฏ๐ฌ ๐ฑ๐ฎ๐๐ it emails you a check-in, and you click to say you're fine.
Miss one and it sends ๐ฏ ๐ฟ๐ฒ๐บ๐ถ๐ป๐ฑ๐ฒ๐ฟ๐, ๐ฎ ๐๐ฒ๐ฒ๐ธ ๐ฎ๐ฝ๐ฎ๐ฟ๐. With the last one, it can ask a trusted contact whether you're really unreachable. If they say you're OK, delivery pauses. If nobody answers, your messages go out.
Every timing is yours to change. You can even make one person wait extra days before they can open theirs.
๐ฃ๐ฟ๐ถ๐๐ฎ๐ฐ๐ ๐ถ๐ ๐๐ต๐ฒ๐ฟ๐ฒ ๐ถ๐ ๐ด๐ฒ๐๐ ๐๐ฒ๐ฟ๐ถ๐ผ๐๐. Messages are encrypted before they leave your browser, so the server only ever stores scrambled text. Per the README, even the person running it can't read them.
And it's yours. ๐ก๐ผ ๐ต๐ผ๐๐๐ฒ๐ฑ ๐๐ฒ๐ฟ๐๐ถ๐ผ๐ป, ๐ป๐ผ ๐๐๐ฏ๐๐ฐ๐ฟ๐ถ๐ฝ๐๐ถ๐ผ๐ป. The authors say plainly that anyone selling LastSignal as a service isn't them. You run it with Docker on your own server, even a Raspberry Pi 4 or 5. It's MIT licensed, built with Rails 8 and SQLite.
Google has had something similar for years (Inactive Account Manager), but that hands over your Google data, held by Google.
LastSignal sends words you wrote, from a machine you control.
Two honest caveats:
The README warns that if someone steals the database, they can try to guess your recipients' passphrases offline with no limit, so ๐๐ฒ๐ฎ๐ธ ๐ฝ๐ฎ๐๐๐ฝ๐ต๐ฟ๐ฎ๐๐ฒ๐ ๐ฎ๐ฟ๐ฒ ๐๐ต๐ฒ ๐ฟ๐ฒ๐ฎ๐น ๐ฟ๐ถ๐๐ธ. And it only works if your server and email provider are still running when the day comes.
๐๐ฐ๐ด๐ต ๐ฐ๐ง ๐ถ๐ด ๐ฑ๐ญ๐ข๐ฏ ๐ง๐ฐ๐ณ ๐ฐ๐ถ๐ณ ๐ฎ๐ฐ๐ฏ๐ฆ๐บ. ๐๐ฆ๐ณ๐บ ๐ง๐ฆ๐ธ ๐ฑ๐ญ๐ข๐ฏ ๐ง๐ฐ๐ณ ๐ต๐ฉ๐ฆ ๐ญ๐ข๐ด๐ต ๐ต๐ฉ๐ช๐ฏ๐จ ๐ธ๐ฆ ๐ธ๐ข๐ฏ๐ต ๐ต๐ฐ ๐ด๐ข๐บ.
Links to repo: github.com/giovantenne/lastsโฆ
Gul Saeed Khattak retweeted
Most people keep their files, photos and media on someone elseโs servers.
You pay every month, and your data lives on a machine you donโt control.
There is a better way: self-hosting. You run the apps yourself on a small box at home, like a Raspberry Pi or an old mini PC.
But there is a catch. Self-hosting usually means the Linux command line, Docker (a tool that runs each app in its own isolated box), config files and ports. Most people give up there.
An open-source project from IceWhale Technologies solves this with CasaOS.
The idea is simple: a home server with an app-store feel.
Take a machine running Debian, Ubuntu Server or Raspberry Pi OS.
Run one install command.
Open the dashboard in your browser.
Click to install the apps you want.
You pick the apps. The system handles the Docker part by itself.
How it works:
โ It is not a full operating system. It sits on top of your existing Linux as a layer.
โ Every app is a Docker container, so apps stay separate and are easy to remove.
โ A built-in app store gives one-click install for selected apps like Nextcloud, Home Assistant, AdGuard and Jellyfin.
โ The dashboard also handles files and drives, and widgets show resource usage and app status.
It started as the pre-installed system for ZimaBoard, a crowdfunded single-board server. Now it is community-driven.
The project in numbers:
โข ~37.6k GitHub stars
โข 100,000+ Docker apps you can install, they say
โข 3 hardware types: x86-64, arm64, armv7
โข Apache-2.0 license
What I find interesting as someone working with self-hosted tools: Docker was already there. The missing piece was a simple screen on top of it that someone who is not an engineer can use.
Links in the reply ๐
Gul Saeed Khattak retweeted
I work from home, so X is the closest thing I have to an office.
My feed still has too few people who ship things.
I want to change that. If you're building with AI, say hi:
It's too quiet in here.
โข Data Engineers
โข Inovators
โข AI Engineers
โข Vibe coders
โข Founders
Tell me what you're building ๐
Gul Saeed Khattak retweeted
JUST IN -MASSIVE CYBERATTACK:
178 organisations, including the BBC, CNN, HSBC and Zoom, were hit by a supply-chain attack via Optimizely. The malicious code ran on their websites, in visitorsโ browsers.
One vendor flaw, 178 victims. Third-party scripts are now a front-line risk. #InfoSec
newsweek.com/cyber-attack-weโฆ
Gul Saeed Khattak retweeted
A design skill for Claude Code, Codex, Cursor and OpenCode that Removes AI Slop in Web Design.
Youโve seen this website a thousand times.
A small label in capitals above every heading. A purple gradient. Three feature cards with icons. โTrusted by teams everywhere.โ An email sign-up box at the bottom.
You know it was made by AI before you finish scrolling.
Thatโs the problem @nutlope set out to fix with Hallmark. Itโs an open-source design skill for Claude Code, Codex, Cursor and OpenCode, and its tagline is that it โrefuses to look AI-generated.โ
Itโs already at 32k stars on GitHub.
Hereโs how it works.
For non-technical people:
Think of it as a strict art director sitting next to your AI. Before anything ships, it checks the page against a list of things that make a site look AI-made, and sends the work back if it fails.
For technical people:
โข Quality gates in three tiers. โFloorโ is never broken: contrast, keyboard focus, reduced motion, mobile layout, and no invented facts. โReflexโ is defaults you can override with a reason. โFinishโ is craft details.
โข 48 โmaterialsโ (themes), each a single file under 8 KB, or a custom mode built from your prompt
โข It builds and reviews the hero section first, renders the page at eight screen widths with a lint script, then has a fresh subagent review it
โข 755 openly licensed photos, credited on the page and deliberately not from Unsplash or Pexels
โข /hallmark audit and /hallmark redesign for pages you already have
โข Output is plain, self-contained HTML and CSS
My favourite detail:
one of the things it checks for is a contact form that says โthanks!โ without actually sending anything. If youโve shipped AI-built sites, you know exactly why thatโs on the list.
The bigger lesson isnโt about design.
AI defaults to the average of everything it has seen. Thatโs why so many AI-built sites look the same.
A skill like Hallmark doesnโt make the model smarter.
It gives it taste, written down as rules.
Same model. Better instructions. Very different result.
Install:
npx skills add nutlope/hallmark -g
Free and MIT licensed.
Repo:
github.com/Nutlope/hallmark
Gul Saeed Khattak retweeted
๐ฅ๐๐๐ ๐๐ ๐๐ฎ๐๐๐ฒ๐ฟ ๐ฎ๐ป๐ฑ ๐ฆ๐ฎ๐ณ๐ฒ๐ฟ. ๐ฆ๐ผ ๐ช๐ต๐ ๐๐ฟ๐ฒ ๐ง๐ฒ๐ฎ๐บ๐ ๐ฆ๐ฐ๐ฎ๐ฟ๐ฒ๐ฑ ๐๐ผ ๐ฆ๐๐ถ๐๐ฐ๐ต? rust-anything ๐ ๐ฎ๐ธ๐ฒ๐ ๐๐ ๐ฎ ๐ฃ๐ถ๐ฒ๐ฐ๐ฒ ๐ผ๐ณ ๐๐ฎ๐ธ๐ฒ
Every team has that one slow function. The search box that lags on each keystroke. The script that should take a second and takes a minute.
Someone says "just rewrite it in Rust" and ๐๐ต๐ฒ ๐ฟ๐ผ๐ผ๐บ ๐ด๐ผ๐ฒ๐ ๐พ๐๐ถ๐ฒ๐.
The upside is real, though. Rust compiles to machine code with no garbage collector, and it checks memory use before the code ever runs. Google says memory-safety bugs fell ๐ฏ๐ฒ๐น๐ผ๐ ๐ฎ๐ฌ% ๐ผ๐ณ ๐๐ป๐ฑ๐ฟ๐ผ๐ถ๐ฑ'๐ ๐๐๐น๐ป๐ฒ๐ฟ๐ฎ๐ฏ๐ถ๐น๐ถ๐๐ถ๐ฒ๐ in 2025 as more new code went into Rust. Its medium and large Rust changes also get rolled back about ๐ฐ๐
๐น๐ฒ๐๐ ๐ผ๐ณ๐๐ฒ๐ป than C++.
The fear comes from the old code. It's full of small decisions nobody wrote down. How it treats an emoji, a negative zero, an empty file. ๐ ๐ฟ๐ฒ๐๐ฟ๐ถ๐๐ฒ ๐พ๐๐ถ๐ฒ๐๐น๐ ๐ด๐ฒ๐๐ ๐ผ๐ป๐ฒ ๐ผ๐ณ ๐๐ต๐ฒ๐บ ๐๐ฟ๐ผ๐ป๐ด.
rust-anything, a new open-source tool by Erkam Yaman, is built for exactly that worry.
It doesn't write the Rust. Your coding agent or you do. It runs your original app on thousands of edge cases, saves every answer, then replays them all against the Rust. ๐ก๐ผ๐๐ต๐ถ๐ป๐ด ๐ฐ๐ผ๐๐ป๐๐ ๐ฎ๐ ๐ฑ๐ผ๐ป๐ฒ ๐๐ป๐๐ถ๐น ๐ฒ๐๐ฒ๐ฟ๐ ๐ฐ๐ฎ๐๐ฒ ๐บ๐ฎ๐๐ฐ๐ต๐ฒ๐. Your callers keep the same API, so you can move one hot loop or, step by step, the whole app.
The README has a good example. A first port passed simple tests but failed 2 of 200 recorded cases, because JavaScript counts an emoji as two positions and Rust counted one. That's the kind of bug that usually reaches production.
The speedups in the repo are big: a search over 50,000 keys went from ๐ญ๐ฌ ๐บ๐ ๐๐ผ ๐ฎ๐ฏ๐ผ๐๐ ๐ฌ.๐ฑ ๐บ๐, and a Python ๐๐๐๐๐๐๐ข got ๐ญ๐ณ๐
๐ณ๐ฎ๐๐๐ฒ๐ฟ. Those are the project's own numbers, on 4 cores.
And Rust isn't magic. When the author ported 29 small functions one by one, ๐๐ต๐ฒ๐ ๐ฐ๐ฎ๐บ๐ฒ ๐ผ๐๐ ๐๐น๐ผ๐๐ฒ๐ฟ ๐๐ต๐ฎ๐ป ๐๐ต๐ฒ ๐ง๐๐ฝ๐ฒ๐ฆ๐ฐ๐ฟ๐ถ๐ฝ๐, because crossing into Rust costs more than the work. The gains come from moving whole loops, and the tool benchmarks both before you ship.
It covers TypeScript, Python, C, Go, Ruby and Java. It's MIT licensed, runs locally, and came out this week.
๐๐ฉ๐ฆ ๐ฉ๐ข๐ณ๐ฅ ๐ฑ๐ข๐ณ๐ต ๐ฐ๐ง ๐ข ๐ณ๐ฆ๐ด๐ณ๐ช๐ต๐ฆ ๐ธ๐ข๐ด ๐ฏ๐ฆ๐ท๐ฆ๐ณ ๐ต๐ฉ๐ฆ ๐ฏ๐ฆ๐ธ ๐ญ๐ข๐ฏ๐จ๐ถ๐ข๐จ๐ฆ. ๐๐ต ๐ธ๐ข๐ด ๐ฑ๐ณ๐ฐ๐ท๐ช๐ฏ๐จ ๐ต๐ฉ๐ฆ ๐ฏ๐ฆ๐ธ ๐ค๐ฐ๐ฅ๐ฆ ๐ฅ๐ฐ๐ฆ๐ด ๐ธ๐ฉ๐ข๐ต ๐ต๐ฉ๐ฆ ๐ฐ๐ญ๐ฅ ๐ค๐ฐ๐ฅ๐ฆ ๐ฅ๐ช๐ฅ.
Links to Repo:
github.com/erkamyaman/rust-aโฆ.
Gul Saeed Khattak retweeted
Looking alternative Open source SaaS apps you can run on your own server, not just in the cloud:
1. Slack alternative: Campfire (by 37signals)
One Docker image, runs on a single machine, MIT licensed
2. DocuSign alternative: SignuDoc
Runs anywhere Node.js and Postgres run, or 1-click to Vercel. MIT licensed
3. Google Analytics alternative: Umami
Docker image, privacy-first, no cookies. MIT licensed, 40k+ stars
Prefer one-click cloud chat?
Ember deploys to your own Cloudflare account.
Links in reply ๐
Gul Saeed Khattak retweeted
Every few months, another password manager company sends the email nobody wants to get:
"We recently detected unusual activity..."
And you realise every password you own lives on a server you've never seen.
Some people chose a different path. They run their password manager on their own server.
GitHub lists free software you can host yourself:
๐ Vaultwarden: github.com/dani-garcia/vaultโฆ
A lightweight server written in Rust that works with the official Bitwarden apps and browser extensions
๐ก๏ธ Bitwarden: github.com/bitwarden/server
The official self-hosted server, with web app, browser extension and mobile apps
๐ง AliasVault: github.com/aliasvault/aliasvโฆ
An end-to-end encrypted password manager that also generates email aliases, so every site gets a different address
๐ฅ Passbolt: github.com/passbolt/passboltโฆ
Built for teams sharing passwords
For non-technical people:
Instead of keeping your keys in a bank's safe, you keep them in a safe in your own house. You choose who has the key.
For technical people:
Vaultwarden is the favourite for home servers because it runs happily on a Raspberry Pi. It's an unofficial reimplementation of the Bitwarden server, so read up on what that means before you trust it with everything.
The honest catch, and it's a big one: if your server dies and you have no backup, your passwords die with it. Self-hosting passwords means taking backups seriously.
Get that right, and nobody can sell, leak or lose your vault except you.
Gul Saeed Khattak retweeted
You search for "best running shoes" once.
For the next two weeks, every website you visit shows you running shoes.
Search engines don't just answer your questions. They remember them.
Here is a list of tools which can help you avoid tracking by search engines.
๐ SearXNG: github.com/searxng/searxng
A private search engine that gathers results from many search services, with no tracking and no ads
๐ธ๏ธ YaCy: github.com/yacy/yacy_search_โฆ
A peer-to-peer search engine run by its users, with no company in the middle
๐๏ธ Hister: github.com/asciimoo/hister
Your own personal search engine that indexes the sites you visit, so you can find that article from last month
๐ Meme Search: github.com/meme-search/meme-โฆ
AI-powered search for your meme folder. It reads each image and finds it by description
โก Meilisearch: github.com/meilisearch/Meiliโฆ
โก Typesense: github.com/typesense/typesenโฆ
Fast, typo-tolerant search engines to build into your own apps
For non-technical people:
SearXNG is Google without the part where Google remembers you. You search, you get results, and nothing follows you around afterwards.
For technical people:
SearXNG is a metasearch engine. It queries other engines on your behalf and strips the tracking. Meilisearch and Typesense are the go-to open-source alternatives to Algolia for search inside your own products.
The honest catch: SearXNG still depends on the big engines' results. It hides you from them, not them from you.
And yes, Meme Search is real. Someone finally solved the most important search problem of all.
Gul Saeed Khattak retweeted
I did not expect a $60 AI bill to rebuild a whole browser tool.
The tool is Browse, from the OpenCode team.
Suppose you are building an AI Agent which needs access to internet.
What do you do ?
Obviously you give your Agent Browser tool , and that is where you will need this tool .
It lets your AI coding agent use your real Chrome, with the sites you are already logged in to.
If a page asks for a password or a login code, the agent stops and waits for you.
Built in Rust, a programming language known for small, fast programs.
The AI usage for the rewrite cost about $60.
OpenCode says the new version uses 25 times less memory and is 23 times smaller. Those are their own numbers, so treat them as a claim for now.
It is free. It also works with Claude Code and Cursor, not only OpenCode.
I would not point it at my main browser yet.
A web page can hide text that tricks the agent, and the agent is logged in as you.
Every AI browser tool has this problem, Claude in Chrome included. The Browse README itself calls its limits "guardrails, not a data boundary".
I would make a second Chrome profile and log in only to the sites the agent needs.
Links in Replies ๐
Gul Saeed Khattak retweeted
Google Cloud AI Research's Paper-Writing Agents Are On GitHub. It turns your resutls into paper in 40 min.
These Agents Rewrite A Paper Until An AI Reviewer Scores It Higher
That's the clever part. It's also the catch.
๐ฌ๐ผ๐๐ฟ ๐ฒ๐
๐ฝ๐ฒ๐ฟ๐ถ๐บ๐ฒ๐ป๐๐ ๐ฎ๐ฟ๐ฒ ๐ฑ๐ผ๐ป๐ฒ. ๐ง๐ต๐ฒ ๐ฝ๐ฎ๐ฝ๐ฒ๐ฟ ๐ถ๐๐ป'๐.
Every researcher knows deadline week. The results sit in a log file. The related work section is still empty.
Google Cloud AI Research built PaperOrchestra for that gap, and the code is on GitHub.
๐๐ผ๐ ๐ถ๐ ๐๐ผ๐ฟ๐ธ๐
Think of a small editorial team made of AI agents. You hand over the idea and what the experiments showed.
โข One agent plans the paper
โข One finds related studies and checks that each one exists in a real database
โข Others draw the figures and write the text
โข The last plays the tough colleague: it shows the draft to a simulated reviewer and throws away any edit that lowers the score
๐ช๐ต๐ฒ๐ฟ๐ฒ ๐ถ๐ ๐ณ๐ถ๐๐
"AI scientist" tools differ in how much they do for you:
AI Scientist-v2 (Sakana AI):
everything. One of its 3 papers passed review at an ICLR 2025 workshop
AI-Researcher:
also everything, from reading past work to writing
PaperMentor:
the least. It comments on your Overleaf draft
PaperOrchestra:
the middle. Your experiments, its writing
๐ง๐ต๐ฒ ๐บ๐ถ๐ฑ๐ฑ๐น๐ฒ ๐ถ๐ ๐ฒ๐ฎ๐๐ถ๐ฒ๐ฟ ๐๐ผ ๐๐ฟ๐๐๐. The science stays yours.
๐ง๐ต๐ฒ ๐ด๐ผ๐ผ๐ฑ ๐ป๐ฒ๐๐
โข In blind tests, ๐ญ๐ญ AI researchers preferred its papers to two other AI systems by wide margins
โข Its drafts cite ๐ฐ๐ฒโ๐ฐ๐ด papers on average. Human-written ones cite ๐ฑ๐ต
But here's the part the headline skips.
๐ง๐ต๐ฒ ๐ณ๐ถ๐ป๐ฒ ๐ฝ๐ฟ๐ถ๐ป๐
โข Against the human originals it tied or won only ๐ฐ๐ฏ% of the time, and only on the literature review
โข The raters often disagreed with each other
โข The team that built it ran the tests
โข Its input notes were written by an AI reading finished papers, much tidier than a real lab notebook
What we actually know:
on AI papers, in two conference formats, it beats earlier AI tools and trails people.
The authors say it themselves: "๐ข ๐ฒ๐ถ๐ข๐ญ๐ช๐ต๐บ ๐จ๐ข๐ฑ ๐ณ๐ฆ๐ฎ๐ข๐ช๐ฏ๐ด."
๐ง๐ต๐ฒ ๐ฟ๐ฒ๐ฎ๐น ๐ฟ๐ถ๐๐ธ
It's like a student who keeps only the edits that raise the grade. Sensible. But another model gives the grade.
๐ฆ๐ผ ๐๐ต๐ฒ ๐ฑ๐ฟ๐ฎ๐ณ๐ ๐ฑ๐ฟ๐ถ๐ณ๐๐ ๐๐ผ๐๐ฎ๐ฟ๐ฑ ๐๐ต๐ฎ๐ ๐ฎ ๐ฟ๐ฒ๐๐ถ๐ฒ๐๐ฒ๐ฟ ๐บ๐ผ๐ฑ๐ฒ๐น ๐น๐ถ๐ธ๐ฒ๐. That's close to a good paper. It's not the same thing.
And reviewers are already stretched. For ICLR 2026, a Pangram Labs detector flagged about ๐ฎ๐ญ% of reviews as fully AI-generated.
๐ ๐ฅ๐ณ๐ข๐ง๐ต ๐ฏ๐ฐ๐ธ ๐ต๐ข๐ฌ๐ฆ๐ด 40 ๐ฎ๐ช๐ฏ๐ถ๐ต๐ฆ๐ด. ๐๐ฉ๐ฆ๐ค๐ฌ๐ช๐ฏ๐จ ๐ช๐ต ๐ด๐ต๐ช๐ญ๐ญ ๐ต๐ข๐ฌ๐ฆ๐ด ๐ข ๐ฉ๐ถ๐ฎ๐ข๐ฏ.
Would you let an agent write your paper if the experiments stayed yours?
Links in reply ๐
Gul Saeed Khattak retweeted
"Your storage is almost full."
That notification has made more people pay for cloud storage than any advert ever has.
First 15GB is free. Then 100GB. Then 2TB. And every year, you're paying rent on your own files.
GitHub has a list of free software you can run yourself, has another option:
โ๏ธ Nextcloud: github.com/nextcloud/server
Files, calendar, contacts and more, a full Google Drive replacement you host yourself
๐ Syncthing: github.com/syncthing/syncthiโฆ
Syncs folders directly between your devices, with no cloud in the middle at all
๐ Seafile: github.com/haiwen/seafile
Fast file sync built for teams
๐ป Puter: github.com/heyputer/puter
A whole desktop operating system that runs in your browser, from your own server
For non-technical people:
It's your own Dropbox. Buy a hard drive once instead of paying every month, forever.
For technical people:
Nextcloud is the all-in-one option. Syncthing is the elegant one: peer-to-peer and encrypted, with no server needed at all. Puter is the one to try just to see it.
The honest catch: you're now responsible for backups. One hard drive is not a backup.
But a 4TB drive costs less than many people pay for a year or two of cloud storage, and it's yours forever.
Gul Saeed Khattak retweeted
Budgeting apps want two things from you: a monthly fee, and the login to your bank.
Then they analyse your spending to "help" you. Sometimes that help includes showing you offers from their partners.
There's another way. self hoste them
These are the ones worth a look:
๐ฐ Actual: local-first budgeting where every pound or dollar gets a job, synced across your devices
github.com/actualbudget/actuโฆ
๐ฅ Firefly III: a full personal finance manager with budgets and forecasts
github.com/firefly-iii/firefโฆ
๐ Ghostfolio: tracks your stocks, ETFs and crypto in one dashboard
github.com/ghostfolio/ghostfโฆ
๐ Wallos: a subscription tracker that shows how much all those monthly charges really add up to
github.com/ellite/wallos
๐ IHateMoney: splits shared costs between friends or housemates
github.com/spiral-project/ihโฆ
๐งพ Receipt Wrangler: an AI-powered receipt manager that sorts your receipts for you
github.com/Receipt-Wrangler/โฆ
For non-technical people:
It's a spreadsheet for your money that does the boring parts for you, and nobody else can see it.
For technical people:
Most of these are a single Docker container. Actual runs local-first, so your data sits on your device even without the server.
My favourite is Wallos. Seeing every subscription listed on one page is a wake-up call. Most people underestimate the total.
The honest catch: you lose the automatic bank feeds that paid apps offer, so expect some manual imports.
Small price for keeping your finances to yourself.
Gul Saeed Khattak retweeted
MIRACLES : Video Editing May Never Be the Same After Higgsfield Miracle and GPT-6.2 Astra
GPT-6.2 Astra Can Perform Miracles Now. Literally.
Well, almost. And if youโre wondering what I mean, check out Higgsfield Miracle.
Editing specific objects in a video isnโt new. Runway Aleph already does selective AI video editing, while Adobe After Effects offers masking, tracking and object-level edits.
Whatโs interesting is what Higgsfield is promising with GPT-6.2 Astra: turning ordinary video into editable, depth-aware layers, allowing real-time changes to objects, lighting and textures without disturbing the rest of the scene.
Astra powers the experience, although its precise technical role hasnโt been disclosed.
Still in private beta, so the real test is yet to come.
Not the first to tackle this problem. But potentially a big step towards editing video as easily as editing a photo.
We built Higgsfield Miracle to address one of the biggest challenges in AI video: editing a specific part of a shot without changing the rest.
Powered by GPT-6.2 Astra, it splits any video into layers for pixel-level editing in real time, with world knowledge built in. Every shot keeps the image depth that cinema-grade post-production needs.
Now in private beta with select testers. Coming soon to Higgsfield MCP.
Set up Higgsfield plugin in ChatGPT to be the first to try it.
Gul Saeed Khattak retweeted
JUST RELEASED: Microsoft's new security model beats GPT-6 Astra at finding real bugs . It costs 1% of the price.
The price on Microsoft's new security model looks like a typo.
$0.10 per million input tokens. GPT-6 Astra lists at $10.
It is a small model trained for security work: finding holes in code and suggesting the fix.
On CyberGym-E2E, a test where the model has to find real bugs in real projects, Microsoft says it scored 92.3%. GPT-6 Astra scored 91.5%. Cyber 2 Flash was allowed half the thinking time, and each task cost under 2% of what Astra's did.
It also reads 1 million tokens at once, 4 times the last version.
For a tool that has to read whole codebases, I care about that more than the benchmark score.
Now the parts I'd be careful with.
You also can't call it yourself.
There is no API.
It only runs inside Microsoft security products, and it goes into MDASH, their bug-hunting system, for eligible customers from mid-December 2026.
My read as a data engineer: cheap specialist models doing the routine work is where this goes. I'd still want to see this one on somebody else's benchmark first.
Links in replies ๐๐
Gul Saeed Khattak retweeted
Asking an AI about a big codebase usually goes badly.
It reads a few files, misses the links between them, and guesses the rest.
Code-Graph-RAG takes a different path. It turns your whole repo into a graph: functions, classes, modules, and who calls what.
Saves the Graph in Embeded SQLite Database.
Then you ask in plain English. Your question becomes a graph query, and the answer comes from the real code structure.
โ Tree-sitter does the parsing
โ 14 languages fully supported (Python, TypeScript, Rust, Go, Java, C++ and more)
โ Stored in a SQLite file by default, so no database to set up
โ Works as an MCP server, so Claude Code can use it
โ Also finds dead code and can edit code with a diff preview
MIT license, ~5.4k stars.
Links in replies ๐๐
Gul Saeed Khattak retweeted
Most security scanning today works the noisy way.
A tool reads your code, matches patterns, and gives you a long list of warnings. Then a human has to check each one to see if it is real.
There is a faster way: let an AI agent do the security review. It reads the code like an engineer would, and looks for real vulnerabilities (bugs an attacker can use).
But there is a catch. A finding is only a guess until someone proves it. And after that, someone still has to write the fix and check that the fix works.
OpenAI has open-sourced Codex Security, a CLI and TypeScript SDK that tries to cover this whole path.
The idea is simple: find, validate, fix.
1/. Install it and sign in: npm install --global @openai/codex-security, then cs login.
2/. Run cs scan . in your repo.
3/. It validates the candidate findings and generates patches.
4/. You get a Markdown report, and you can export the results.
The tool does the checking work, so you look at fewer, better findings.
How it works:
Scan scope:
you can scan the full repo, selected paths, or only Git changes (cs scan . --diff origin/main).
Deep mode:
it repeats the Standard scan several times and removes
duplicate findings.
Threat model: each scan saves a threat model (a description of what can be attacked and how) with the results. cs policy . drafts or updates your SECURITY.md.
Triage: it can find duplicates, rate severity against your own rubric, and suggest owners from source and Git history.
What you get:
โข Exports to SARIF, JSON and CSV
โข Tracking in Linear, Jira or GitHub
โข A GitHub Action for scheduled scans (the example uses an estimated $5 limit per scan)
โข Works with OpenAI, Amazon Bedrock, OpenRouter, Fireworks AI and Cheaper Inference
โข Apache-2.0 license, ~13k GitHub stars
One thing to know: it needs an API key or login, and the README says some cybersecurity requests and protected findings require Trusted Access for Cyber.
What I find interesting is the diff scan is the useful part. You can run it on every pull request, so the agent reviews only what changed and not the whole repo each time.
Links
github.com/openai/codex-secuโฆ
Gul Saeed Khattak retweeted
The Headlines Blame Microsoft Outlook Outage. The Lesson Hiding In The Incident Is For Anyone Who Builds Alerts
The story going round this week: a Microsoft 365 outage stopped a chlorine warning email, and 15 hotel guests in Germany got hurt.
The facts as reported. Hotel Ling Bao at the Phantasialand theme park, Friday 13 November. The pool's automatic dosing system put in too much chlorine. Its out-of-range alert goes out by email, and the park told the broadcaster WDR those emails didn't arrive during the outage. The fire service says ๐ญ๐ฑ ๐ฝ๐ฒ๐ผ๐ฝ๐น๐ฒ ๐๐ฒ๐ฟ๐ฒ ๐ฐ๐ต๐ฒ๐ฐ๐ธ๐ฒ๐ฑ ๐ฎ๐ป๐ฑ ๐ณ๐ผ๐๐ฟ ๐๐ฒ๐ป๐ ๐๐ผ ๐ต๐ผ๐๐ฝ๐ถ๐๐ฎ๐น. The water has been replaced and the pool is open again.
Now the caveat, because it changes how you should read all of that. The causal claim comes from ๐๐ต๐ฒ ๐ผ๐ฝ๐ฒ๐ฟ๐ฎ๐๐ผ๐ฟ'๐ ๐ผ๐๐ป ๐๐๐ฎ๐๐ฒ๐บ๐ฒ๐ป๐ ๐๐ผ ๐ผ๐ป๐ฒ ๐ฏ๐ฟ๐ผ๐ฎ๐ฑ๐ฐ๐ฎ๐๐๐ฒ๐ฟ. An operator has an obvious reason to point at Microsoft. I haven't seen it confirmed by the fire service, a regulator or Microsoft. So read "caused by" as "the park says".
imo the interesting part isn't the outage, it's that ๐ฎ๐ป ๐ฎ๐น๐ฎ๐ฟ๐บ ๐๐ฎ๐ ๐ฎ๐ฏ๐น๐ฒ ๐๐ผ ๐ณ๐ฎ๐ถ๐น ๐๐ถ๐๐ต๐ผ๐๐ ๐ฎ๐ป๐๐ผ๐ป๐ฒ ๐ป๐ผ๐๐ถ๐ฐ๐ถ๐ป๐ด.
Think about what a safety alert has to do. It has to reach a person who can act, quickly, and you need to know when it can't reach them. Email does the first job on most days. It does nothing for the last one. ๐๐ฏ ๐ฆ๐ฎ๐ข๐ช๐ญ ๐ต๐ฉ๐ข๐ต ๐ฏ๐ฆ๐ท๐ฆ๐ณ ๐ข๐ณ๐ณ๐ช๐ท๐ฆ๐ด ๐ญ๐ฐ๐ฐ๐ฌ๐ด ๐ฆ๐น๐ข๐ค๐ต๐ญ๐บ ๐ญ๐ช๐ฌ๐ฆ ๐ข ๐ฒ๐ถ๐ช๐ฆ๐ต ๐ข๐ง๐ต๐ฆ๐ณ๐ฏ๐ฐ๐ฐ๐ฏ.
A smoke detector is the opposite design. It makes noise in the room where the problem is. It doesn't need a login, a mail server or a working internet line.
There's a second thing I'd question. Two faults had to line up here: the dosing went wrong, and the warning got lost. A careful design assumes both can happen on the same day. The stronger fix may not be a better notification at all. It may be ๐ฎ๐ป ๐ถ๐ป๐๐ฒ๐ฟ๐น๐ผ๐ฐ๐ธ: the reading goes too high, the pump stops, and no person has to read anything.
What I'd want to know before drawing big conclusions:
โข ๐ช๐ฎ๐ ๐ฒ๐บ๐ฎ๐ถ๐น ๐๐ต๐ฒ ๐ผ๐ป๐น๐ ๐ฐ๐ต๐ฎ๐ป๐ป๐ฒ๐น, or just the only one anyone watched?
โข ๐๐ถ๐ฑ ๐๐ต๐ฒ ๐ฑ๐ผ๐๐ถ๐ป๐ด ๐๐ป๐ถ๐ ๐ต๐ฎ๐๐ฒ ๐ฎ ๐ต๐ฎ๐ฟ๐ฑ ๐น๐ถ๐บ๐ถ๐, and did it work?
โข ๐๐ผ๐ ๐น๐ผ๐ป๐ด passed between the first alert and someone noticing?
โข ๐ช๐ต๐ผ ๐ฐ๐ต๐ผ๐๐ฒ ๐ฒ๐บ๐ฎ๐ถ๐น, the vendor as a default or the site?
My guess, and it is only a guess, is that this pattern is common. Freezer temperature alerts, leak sensors and building alarms often end in an inbox, because that's the easiest thing to set up.
I would like to see this kind of equipment ship with a loal alarm by default , with email as the copy and not the alarm .
Links in reply ๐












