Gemini has a new embedding model out called gemini-embedding-exp-03-07 - I've added support for that in llm-gemini 0.14, notes here simonwillison.net/2025/Mar/7/g...

RAWR! I want to share that I've incorporated #AI #agent into my threat detection tool for #GitHub enhancing its analytical capabilities with an extra layer of analysis. What's exciting is the continuous improvement. I will refine it with personalized feedback loops for reporting suspicious accounts.

A 1900 newspaper page - extremely dense columns of text

I upgraded my llm-anthropic plugin to use the new API feature where you can pass URLs to images and PDFs, and in testing it got a great result from Claude 3.7 Sonnet against this 1900 newspaper scan from the Library of Congress
simonwillison.net/2025/Mar/1/l...

Completely understand. I'm still trying to get mine off the ground but I've so busy with work and other projects and life in general.

I now have a working web service for #GitHub Watchdog. It's already displaying the latest #malware campaign.

I just got off work and wanted to see what the watchdog has been barking at. It found a really bizarre campaign on #GitHub. Seems benign at the moment but this reminds me of #scammers testing boundaries before sending a mass #phishing campaign in the #email world.

github.com/BearHuddlest...

I've been working on my GitHubWatchdog tool some more by reviewing the results. Then a thought came to me. Why don't I make a "Threat Research" AI agent that can analyze my findings, highlight them, and provide ideas for future heuristic methods. This is really exciting for me! Wanted to share.

After observing the watchdog tool barking at suspicious users on #GitHub for a few days, I noticed ALL malicious repository I've discovered has a disk usage of 2 to 3 kilobytes. I'm writing up a new heuristic method to use this to my advantage.

For fun, I've created a tool, GitHub Watchdog, that'll bark at suspicious users on GitHub. I'm slowly adding more heuristic methods to scan for malicious repos. I'm seeing results. All true positive! Very interesting disk_usage... Possible new behavior pattern?

$ go install golang.org/dl/go1.24.0@latest $ go1.24.0 download Downloaded 0.0% ( 0 / 74636413 bytes) ... Downloaded 50.0% (37318206 / 74636413 bytes) ... Downloaded 100.0% (74636413 / 74636413 bytes) Unpacking go1.24.0.linux-arm64.tar.gz ... Success. You may now run 'go1.24.0' $ go1.24.0 version go version go1.24.0 linux/arm64

🥳 Go 1.24.0 is released!

📰 Release notes: go.dev/doc/go1.24

📦 Download: go.dev/dl/#go1.24.0

#golang

I've discovered an intriguing approach to overcoming most captcha challenges by utilizing agent chaining models. This innovative technique combines essential components such as a vision model, reasoning model, and web API. It opens the way for advanced web scraping.

@github.com The past three days, I've found over 3000 malicious stargazers staring 120+ repositories hosting malwares on the platform. I would love the mass report these but there is no option. Who can I reach out about this?

🎉 Go 1.24 Release Candidate 3 is released!

🔐 Security: Includes security fixes for the go tool and the crypto/elliptic package.

🏖 Run it in dev! Run it in prod! File bugs! go.dev/issue/new

🗣 Announcement: groups.google.com/g/golang-ann...

🚚 Download: go.dev/dl/#go1.24rc3

I'm building a Go web scraper to detect suspicious website patterns. Thinking about phishing, malware, & scams. Challenges: rate limits, JS execution, & legality. Anyone tackled this before? Thoughts on effectiveness? 🤔 #GoLang #WebScraping #CyberSecurity #Infosec

Screenshot of a GitHub repository page titled "FonePaw-Screen-Recorder-Download." The repository is public, with one branch and no tags. The main content shows a README file with the title "DOWNLOAD LINK" in blue and a password "2025" mentioned below. The repository description highlights keywords such as "fonepaw screen recorder mac crack," "fonepaw screen recorder free," and others. Tags like "fonepaw-screen-recorder-mac" and "fonepaw-screen-recorder-free-trial" are visible on the right side. The page indicates 25 stars, 0 forks, and a recent update of the README file made 5 hours ago by the user "akinotaxebunch19."

Screenshot of a GitHub user profile for "nanasita." The profile features a circular profile picture of a woman with long hair wearing a white shirt, set against a natural background. Below the username, it shows 5 followers and 9 following. The account username "st_prasol" is displayed, with the note "Joined 7 hours ago." On the right side, under "Popular repositories," there are six listed repositories with their names, descriptions, and star counts: AI-Based-Music-Generator (Public): An AI-powered tool that generates music tracks based on user preferences (6 stars). Fitness-Tracker-App (Public): A fitness tracking application that logs workouts, calories, and progress (5 stars). Video-Editing-Tool (Public): A desktop application for basic video editing, including trimming and adding effects (5 stars). Password-Manager (Public): A secure password manager that encrypts and stores passwords locally (3 stars). Crypto-Price-Alert (Public): A tool that monitors cryptocurrency prices and sends alerts when certain conditions are met (2 stars). AI-Resume-Analyzer (Public): An AI-powered tool that analyzes resumes and provides feedback on structure and content (2 stars). At the bottom, a contributions chart displays activity across the past year, with 32 contributions indicated. A blue badge on the bottom right displays "2025."

Screenshot of a GitHub user profile for "Antontrey." The profile displays a circular profile picture of a bald man wearing sunglasses and a black shirt, with the username "antontrey." Below the name, it mentions "Software Engineer" and shows 14 followers and 17 following. The location is listed as "Ukraine," and an "Achievements" section features a colorful badge with a smiling face emoji. There is an option to "Block or Report" the user. On the right side, a list of repositories is displayed with public visibility. Each repository name is short and appears to be randomly generated text, each with 2 stars and marked as updated 11 hours ago. The repositories include: hcfgghj (Public) ghvyhf (Public) guyuguy (Public) fgyg (Public) gfghv (Public) vvgyv (Public) jhbghvhgc (Public) gxf (Public) kiu (Public) Each repository includes a dropdown menu for further actions and a star icon. The layout features search, type, and sort filters at the top.

Screenshot of a GitHub profile for the user "xeroxic" with pronouns "she/her." The profile has a circular avatar displaying a turquoise and white pixelated design. Below the username, the name is listed as "FFFF," with 11 followers and 0 following. The location is shown as "Hanoi," and the local time is listed as "06:44 - 13h ahead." The account was created 2 days ago. On the right side, under "Popular repositories," there are six repositories, each labeled as public and with 11 stars: vault-keeper: Tagged as using the Go programming language. data-repair-hub crypto-wizard balance-tracker ledger-command-center token-crafter At the bottom, a contribution graph displays "3,923 contributions in the last year," with a dense activity pattern across most months. The year "2025" is shown on a blue badge on the right side. The "Contribution activity" section is visible below the graph.

I've found various types of suspicious #GitHub accounts thanks to a tool I created. Some new fake #Stargazers, some possibly #hijacked users, others hosting #malware, and even freshly minted #bots.

I was able to develop a tool to detect the Stargazer Goblins on GitHub. Just reported as many the GitHub API rate limit allowed. #GitHub

Just created a new GitHub account. Time to start working on my first public project.

calling it a night, got pretty far building my personal go web server

After experimenting with DeepSeek-R1 and GPT o3-mini, hands-on experience is now a significant differentiator in AI research. Developing expertise through projects & code helps bridge the gap between theory & practical applications. #AIresearch #SoftwareDevelopment

I'm excited about this. I can't wait to get my hands on it and create some agentic AI for my personal workshop.

At first, I relied on so much that when I decided to stop using it, I was at a lost. I just sat there waiting for "Chatpty" to appear after writing the func name. It was a bad habit to break.

Now, I only use the chat function for rubber ducking or quickly look up a niche terminal command.

Just started developing my own web hosting service in Go. Let's see how this project fares throughout 2025.

A shot in the dark, anyone hiring a full stack developer?

Panel 1 Gustopher arrives in the kitchen. In his hand is a list that says "Cast" at the top. He shouts excitedly. Gustopher: DAD! I was cast in the Christmas play! Mewbert is also there. Panel 2 August inserts a casserole into the oven. The casserole is in the type of casserole dish that everyone has seen but no one can find it in any stores. This is the sort of observation that may pay off in the future but it probably won't. August: Congratulations. Gustopher: I GET TO LIGHT THE STREETS ON FIRE! Panel 3 August is perplexed at the inclusion of arson in a Christmas play. Surely the school wouldn't do that, especially not after the incident of 1995. He checks the cast list. August: What? That can't be- Ah... Panel 4 August: You're a street lamp lighter. Gustopher: Oh... I guess that's cool too. Mewbert: Mew (Possibly meaning "My condolences small one. But do not give up hope. I truly believe that within you is the ability to give the best performance possible. You will light the lamps in such a glorious manner, the audiences will believe the streets have been set ablaze by the power of your performance. Enough talk, feed me and pet me.")

Christmas Play - Gator Days

Congrats!

Go REST API, thinking of creating multiple endpoints to do singular operations. This would allow granular POSTS for the clients to utilize. Hm... I could add params for even more granularity...

Ah yes, the Enshittification model.

No coding today. On PTO.

At the highest level, I like to think of LLMs as an assistant and not an expert. If you use a LLM as a source of truth, you're in for a bad time.

Kind of wild that the Llama 3.2 1B and 3B models can support up to 128k content with such small footprint.