Yo Sub Kwon

we're on a crazy moonride right now! #bitcoin

Breaking: Anthropic just cut off Max plan access for third-party Claude tools like OpenClaw. If you're a power user running agents across multiple subscriptions, this changes your math immediately. A quick video reaction to what happened, why Anthropic did it, two stopgap options if you need to keep things running today, and why neither of them is a long term answer. The free ride on subscription arbitrage is over. Check out our blog post to read what comes next: permaship.ai/blog/anthropic… #openclaw #claude #anthropic #opensource

@karpathy People have been building crazy Obsidian brains. I personally index mine in all kinds of ways and one of the best use cases is drafting email/signal/telegram/slack messages. It actually has real context and information to intelligently respond and actually knows the way you think

LLM Knowledge Bases Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So: Data ingest: I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them. IDE: I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides). Q&A: Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale. Output: Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base. Linting: I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into. Extra tools: I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries. Further explorations: As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows. TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.

Claude Code source leaks (again) and this time it's got a gacha-buddy system built in where every user deterministically (by user ID) gets a buddy. It gets a procedurally generated sprite and its own personality.

@dunkhippo33 github.com/permashipai/ne… like this?

This rules/judgment separation is a completely new programming paradigm I never faced before. I suspect we will soon see developer products that help with this abstraction. 5/5

In the long, long ago, when you would run code you just wrote and it actually worked on the first attempt, you would be jubilant. That scenario will never happen again. Now, I expect it should work on the first attempt.

14 of the 17 flights I've booked so far this year had delays. The only thing that has been saving the connections I haven't missed is the connecting flight being delayed as well. Suppose this is the new normal.

I know a lot of complaints in the AI world lately are about how poorly Opus, Codex, or some other model is performing in the moment, which can have a lot of variability. For some things, I'm sure it makes a big difference, but generally for software development, it really doesn't make much of a difference at all. Whether people intuitively understand it or they have read research, it's a fact that a workflow outperforms one-shot model usage. The more robust your workflow, the less important the actual quality of an individual model working within it. A really simple workflow might be build with Claude, and critique with Codex. Repeat. Complex workflows are often created by individuals tailored for their own purposes and may involve sub-agents, a bunch of customized prompts, or a series of loops that check all steps of a PRD are completed and that the QA, Security, or any post-review agents say everything is good. The more your trust your evals and and workflow, the less concerned you can be about the day-to-day performance of a model, and the less you have to be worried about model drift or how well your systems will perform when you change to new models. If your output is generally acceptable after it goes through your process, then you can optimize more for cost and speed.

I have open-sourced a project: Nexus. It's a project I've been working on to make AI proactive, not just responsive. As a CEO, I juggle priorities and the real bottleneck was never just execution, it was judgment. I didn't want AI to wait for me to hand it tasks, but something that could understand my goals, identify and prioritize valuable work on its own, and act accordingly. AI was already very much involved in my processes. Eventually, it reached the point where it could independently surface meaningful opportunities, propose work, execute it, and follow through. That's now Nexus: github.com/PermaShipAI/ne… Essentially, any number of specialized agents continuously look for work they believe would be valuable based on overarching objectives. A commanding agent will evaluate those proposals against values, OKRs, and other strategic priorities, requesting revisions or additional review when needed before approving or rejecting them. Once approved, work is routed to an executor, whether that’s Claude, Codex, Gemini, PermaShip, or another system. From there, Nexus can verify quality, assess outcomes, and determine whether follow-on work should happen next. It has already handled a wide range of work for me: everything from smaller tasks like building and deploying a website with DNS and email, to large, complex multi-repo projects with staged feature releases. Over the past few months, it has been responsible for thousands of PRs. Since then, it has expanded far beyond software development alone. I believe this could be quite useful to a lot of people.

Today we’re announcing that @NitraFinance has raised a combined $187 million in financing as we build the AI-native operating system for healthcare practices. We're also announcing that Dr. Richard Park, founder of CityMD and healthcare legend, will be joining Nitra’s Board.

@levelsio This is exactly where software is heading. It goes against all my experience as a cybersecurity expert to have self-evolving code agents in production, but software is becoming something closer to that of an organism. Your next step will be moving away from Claude altogether.

This week I decided to just permanently switch to running Claude Code on the server mostly on bypass permissions mode: c() { IS_SANDBOX=1 claude --dangerously-skip-permissions "$@"; } And for the first time in my life I think I've actually managed to outrun my todo list What happened is I simply blasted through my to do list of features I had to build and bugs I had to fix I've never shipped so fast and Claude Code almost made no mistakes, and when it did it they were tiny that weren't fatal (important because I'm mostly working on the server in production now) Before I was always known to ship fast (also because I always work alone) but while I shipped new things would always build up on my features/bug board (my users can submit them there) But this is the first week where I've been fast enough to outrun them The board is actually empty! As other people have written on here the real bottleneck is becoming myself and my creativity, not how fast I can ship. Because I think I ship faster now than I can come up with new ideas, or maybe my brain will adjust to this new speed (probably) Also I feel another limit is becoming my own mental context window, as in how many things, features, bugs, projects, I can keep in my mind in parallel while building on all of them. It's a lot and I haven't reached that limit yet but I feel I might be close I also noticed that you start going really fast the more you let it just go loose, before I was slow because I didn't trust it and I was scared it would destroy my code, now I just let it go. As @karpathy wrote, things feel like they've changed a lot around December last year when models became good enough to really code with and I feel the same When I see other friends code with Claude Code I often notice they're slow because they still check everything, which is good of course, but I feel the better way would be to create some tests and just let it run freely and see if it can pass those For me the tests are mostly just me checking out if the new feature on the site works or not, and in 99% cases it just does, and then I ask it to improve it further Because I run Claude Code on the server in production, I don't have to wait for deployment anymore (although that took only 3 seconds anyway before, that still adds up), now it's wait for it to be done coding, I refresh the site and I test it, that feedback loop is how I work and it's made me WAY faster Anyway here's what I did this week and the majority of these things were requested by people on the bug board, I'd say this is about 10x my normal output: 📸 Photo AI - Built new image viewer and mobile image viewer - Added batch remix, multi-photo import, filtering by model in gallery - Security overhaul: phased out insecure ?hash= login, migrated to session tokens - Fixed Google login loop, multi-model selection, talking scripts - Added custom audio upload for talking videos - Created dynamic model selector from server endpoint 🏡 Interior AI - Revived [ Add furniture ] feature (started 6 months ago, image models now good enough) - Added custom style upload for redesigns - Built own Gaussian Splat viewer for 3D - Made /remove_bg endpoint for furniture backgrounds - Migrated 3D walkthrough to new World Labs API - Added .skp file support, paint color masking, empty room button 🎒 Nomads - Launched weekly AI-generated newsletter from chat - Built profile edit modal, moved profile editing from /settings to profile page - Added TikTok/YouTube links, status bar, server-side API tracking - Added hundreds of new profile tags and traits - Fixed timezone filters, broken links, user avatars 🗺️ Hoodmaps - Revived write mode (before was only read for last few years because db was rekt) - Built heatmap mode using sentiment-scored tags (50K+ tags) - Fixed root cause: tags not entering DB due to wrong PRAGMA (should be WAL) - Added good/bad area detection with admin grid controls - Set up Claude Code Telegram bot for live changes - Enabled CF cache, fixed health check, fixed Brussels 📕 MAKE book - Built auto ePub/PDF generator cron worker - Added dynamic generation with personal customer watermarks - Added image compression for file size 💾 Pieter .com - Added Wikipedia text-only reader for Kindle - Exploring Windows 3.11 emulator using v86 (to replace Em-DOSBox) - Added product recommendations on homepage - Installed Wall Street Raider (1986) 👩‍💻 Remote OK - Installed Chatbase AI customer support bot - Added "report not remote" link on job posts 🏨 Hotelist (3 todos) - Fixed hotel URLs and city range bugs - Added iron amenity

You know what's eerie.. Software that builds and improves itself. Even if you've known for years it was coming, actually experiencing it is wild. I wake up every morning to updated software that's better and more powerful than yesterday. Scary and incredible.

If you're building software and you're obsessing over which model to use while you're vibe coding a dozen Claude Codes across 4 licenses, you probably need to take a step back for a minute and re-think your workflow.

Engineer productivity shouldn't be measured in "lines of code" or "tickets closed." It should be things like - How many autonomous tasks can run in parallel - How little human babysitting is required - How quickly the system can improve itself

By 2030, AI is going to render most white-collar jobs obsolete. Software engineers, lawyers, accountants, and every occupation that's behind a computer will be on a catastrophic decline. Robots on a huge incline.

@IndexVentures There is an account @AdriaHua that is posing as one of your staff and trying to trick accounts into downloading malware under the guise of meeting with them. All the correspondence seems to be AI-automated.

@levelsio I agree that it's exciting. It's also terrifying. I think $2T SaaS industry gets completely annihilated along with hundreds of millions of human employees and replaced by AI systems. It's ridiculous how easy it is now as a single individual to build a powerful new technology.

It's so exciting to be in tech right now He's 100% right, it's like the early days of the internet around 1995 again Everything was new and you had endless opportunities, you could build the first web browser, or mail app, or video streaming tech, etc But instead of internet, it's now the early days of AI and 1995 is 2025 Most people don't realize it but this is a VERY special time with SO many opportunities right now, all you need is to build And times like these only happen every ~30 years if at all, jump on it!

Tinkering on an agentic system that modifies its own code to add functionality when required. Who else is working on this kind of stuff? How are you handling continuous deployment? Rollbacks? Figuring out when the code is ready for merge or requires a human to be looped in?

I'll browse the litepaper this weekend and see if it's anything like what I've envisioned.

Maybe they price differently, or it's only for women, or it's Spanish-speakers, luxury vehicles only, or has USB charging available. The possibilities are quite wide for anyone who can find a competitive way to brand their app and enroll drivers and customers.

I just heard about trip.dev, the rideshare protocol, today. I'm not endorsing it as I don't know much about it yet, but I've been saying for many years now that the underlying tech for companies like Uber and Twilio needs to be turned into a protocol or