I thought I was seeing fewer arXiv papers on the front page of Hacker News (HN) these days, and I wanted to check if that was real.

So I asked Claude to run a quick analysis: track the share of arXiv stories on HN over time. It queried the BigQuery HN dataset, bucketed the stories by month.

This is the 16th year we’ve been teaching the Stanford Lean LaunchPad class. This year, from the first hour of the first class, we realized we were seeing something extraordinary happen. It was both the end and beginning of a new era.

Teams showed up to the first day of class with MVPs (Minimal Viable Products) looking like finished products that previous classes had taken weeks or months to build. After the class, as the instructors sat processing what just happened, we realized there’s no going back.

I’ve been writing about how AI is going to change startups, but the shock of seeing 8 teams actually implementing it was mind blowing. And not a single team thought they were doing anything extraordinary.

I’ve been getting more and more curious about the risk from Anthropic’s Claude Mythos Preview. So I pulled the system card, a whoppingly inefficient 244-page document that devotes just seven pages to the claim that the model is too dangerous to release. In fact, the 23MB of PDF I had to download was 20MB of wasted time and space. Compressing the PDF to 3MB meant I lost exactly nothing.

Foreshadowing, I guess.

Spoiler alert: the crucial seven pages out of 244 do not contain the word “fuzzer” once. That’s like a seven page vacation brochure for Hawaii that leaves out the word beaches.

Also, the crucial seven pages out of 244 do not contain the expected acronyms CVSS, CWE or CVE, they do not have comparison baseline, an independent reproduction, or the word “thousands.” I’ll get back to all of that in a minute.

The flagship demonstration document turns out to be like the ending of the Wizard of Oz, a sorry disappointment about a model weaponizing two bugs that a different model found, in software the vendor had already patched, in a test environment with the browser sandbox and defense-in-depth mitigations stripped out. Anthropic failed, and somehow the story was flipped into a warning about its success.

An autonomous robot ping-pong player dubbed Ace has achieved a milestone for AI and robotics in Tokyo by competing against and sometimes defeating top-level human players at table tennis, a feat that could presage an array of other applications for similarly adept robots.

Ace, created by the Japanese company Sony's (6758.T), opens new tab AI research division, is the first robot to attain expert-level performance in a competitive physical sport, one that requires rapid ‌decisions and precision execution, the project's leader said. Ace did so by employing high-speed perception, AI-based control and a state-of-the-art robotic system.

If you’ve ever spent any length of time in a tall building—either because you live or work in one—you probably know this feeling: The elevator’s cars always seem to go in the wrong direction. If you want to go down, there’s an elevator going up, and vice versa.

The phenomenon intrigued physicists George Gamow and Marvin Stern in the mid-1950s.

Shellscape is an online web app that simulates a terminal environment for learning Linux shell commands. It has 31 levels of increasing difficulty that work entirely on the frontend without needing any virtual machines or installations.

In Serbia, we say “zimi-zami-zum”. When “zum” is said, everyone shows either a fist (like rock) or a relaxed palm (like paper). This is played by 3 players. The odd one out (eg. the lone palm vs. 2 fists) is “it”. We can scale this up to more players by deciding that the majority is eliminated and the minority moves to the next round. Repeat this process until there are fewer than 3 people remaining.

Mozilla says it used an early version of Anthropic's Claude Mythos Preview to comb through Firefox's code, and the results were hard to ignore. In Firefox 150, the team fixed 271 vulnerabilities identified during this effort, a number that would have been unthinkable not long ago. Instead of relying only on fuzzing tools or human review, the AI was able to reason through code and surface issues that typically require highly specialized expertise.

A curated list of modern alternatives to classic command-line tools.

Faster, prettier, smarter replacements for the Unix utilities you use every day.

TUI SSH client Suite: list-based host selection, parallel sessions, multi-hop proxying, and port forwarding. SSH/SCP/SFTP etc. Written in Go.

Claude Code has gotten extremely good at finding security vulnerabilities, and this is only the beginning.

The landscape of AI is not merely filled with news. It is filled with teams. You have the doomers, the accelerationists, the skeptics, the it’s-a-bubble oracles, the anti-bubble counter oracles, and so on. It would be convenient for my sanity—and, perhaps, the sanity of my readers—if I simply joined one team and never removed the jersey. But I don’t think any aforementioned tribe has a monopoly on good arguments. I think the doomers are right about the risk of the technology, and the accelerationists are right about the promise of the technology, and the skeptics are right that the doomers and accelerationists can both overstate their cases.

In 2022, I made a New Year’s resolution to switch from Chrome to Firefox, and from VS Code to Neovim.

My goal was to reduce my dependence on GAFAM tools, and it turned out to be a good decision considering this.

It took some time to adjust, but I am now a happy Firefox user on both desktop and mobile.

That said, it still has some issues, for example, the tab system on Android. I wish there were an easier way to search through tabs or select multiple tabs to close them, instead of closing everything at once.

However, my experience with Neovim was very different. I can say I really tried to adopt it, as I used it for four years before deciding to abandon it.

My first session with Claude Code was practically magical. I was speaking to my computer, telling it with natural language what I wanted it to do, and it was able to just do it. It did ( and still does ) feel like a completely new form of input, a new way to control my machine. I have misgivings about using AI in this way, but I still think this is a great tool for sufficiently low-level tasks. I’m waiting eagerly for the day that I can spin up a local LLM that can perform this function as well as Claude Code does.

I'm as anti-genAI as it gets. And yet, this past month, I have used generative coding to complete a project. It works. I hated making it.

These days, Wandering Thoughts has some hacked together HTTP request rate limits. They don't exist for strong technical reasons; my blog engine setup here can generally stand up to even fairly extreme traffic floods (through an extensive series of hacks). It's definitely possible to overwhelm Wandering Thoughts with a high enough request volume, and HTTP rate limits will certainly help with that, but that's not really why they exist. My HTTP rate limits exist for ultimately social reasons and because they let me stop worrying and stop caring about certain sorts of abuse.

One of the biggest problems with measuring AI progress is the ambiguity of measuring intelligence itself.

AGI is treated as a milestone we have yet to cross, but there is no central definition of AGI.

Depending on who you ask, AGI is achieved when a system:

• can fool humans into thinking it is one of them, in other words, pass a Turing Test
• demonstrates creativity (Springer)
• can develop new skills (DeepMind)
• solves unfamiliar tasks (DeepMind)
• is generally capable across domains (IBM)
• is superior to humans in intelligence (Scientific American)
• outperforms humans economically (OpenAI Charter)
• can independently solve complex problems without human oversight (DeepMind)

Even with the lack of consensus, I can confidently say we have AGI, because the criteria above has been met.

Retrieval-Augmented Generation (RAG) has become the dominant paradigm for grounding Large Language Model (LLM) agents in domain-specific knowledge. The standard approach requires selecting an embedding model, designing a chunking strategy, deploying a vector database, maintaining indexes, and performing approximate nearest neighbor (ANN) search at query time. We argue that for domain-specific knowledge grounding --- where the vocabulary is predictable and the corpus is bounded --- this entire stack is unnecessary. We present Knowledge Search, a two-layer retrieval system composed of (1) grep with contextual line windows and (2) cat of pre-structured fallback files. Deployed in production across 20 specialized LLM agents serving three knowledge domains (Traditional Chinese Medicine, Christian spiritual classics, and U.S. civics), our approach achieves 100% retrieval accuracy with sub-10ms latency, zero preprocessing, zero additional memory footprint, and zero infrastructure dependencies.