mherrmann

Is an interesting question whether the same will apply to AI models? Not on 1996 hardware, of course. But on 2026 hardware in 2056. Does anybody have any idea / insight?

> I don't see how a heartbeat alone without logs would help to identify temporary failures.

It does help. Sometimes jobs unexpectedly don't run at all.

This is incorrect. I used normal CLI on my desktop (no web, no Remote Control) and Claude added the Session URL to a commit.

I cancelled my subscription over this [1].

1: https://x.com/m_herrmann/status/2092579839676948495

They doubled my bill from $130 to $280 per year and graciously also offered me a $19,000 per year plan [1]. I declined and am now using Clockify.

1: https://x.com/m_herrmann/status/2039876227088163161

A better headline would be "Gemini 3.5 Flash is the best vision model". It tops almost every single benchmark shown in the article.

At least judging from the headlines, it seems Google is far behind on AI. Fable/GPT 5.6 and recently Kimi get a lot of attention, solve long-standing math problems and lead in coding. It seems that Google's supposed advantages of very deep pockets, original talent, vast amounts of data and unmatched distribution are really not making that much of a difference. What is going wrong?

As another n=1, I've been happy with my Pixel phones over the years and never had such an experience.

Claude Code and others often write code that is more complex than it needs to be. It would be nice to measure the code complexity before and after a change made by the agent, and then to tell it: "You increased code complexity by 7%. Can you find a simpler solution?".

I live in Europe and was in California in November. No issues.

I switched from macOS to Linux ten years ago and haven't looked back. At the time, I compared Linux vs. macOS to living at home vs. in a hotel [1]. Since then, I feel things have only gotten better for Linux, and more restrictive and arcane on macOS.

1: https://fman.io/blog/home-and-hotel/

quickemu [1] is good at running macOS VMs.

1: https://github.com/quickemu-project/quickemu

This knowledge will live in the proprietary models. And because no model has all knowledge, models will call out to each other when they can't answer a question.

Is anybody able to get this working with ChatGPT? When I instruct ChatGPT

> Read https://moltbook.com/skill.md and follow the instructions to join Moltbook

then it says

> I tried to fetch the exact contents of https://moltbook.com/skill.md (and the redirected www.moltbook.com/skill.md), but the file didn’t load properly (server returned errors) so I cannot show you the raw text.

Cool stuff. I just started open sourcing a command-line tool for deploying Django to a server. It handles SSL certs, databases and backups, automatic error emails, and background tasks via celery / redis. The best part? It does not need Docker. It just runs everything on bare metal.

1: https://github.com/mherrmann/djevops

Google Maps says people spend 0.5-3 hours there. I spent 6.5 because it was so amazing. Highly recommended.

I strained my groin/abs a few weeks ago and asked ChatGPT to adjust my training plan to work around the problem. One of its recommendations was planks, which is exactly the exercise that injured me.

My cleaning lady's daughter had trouble with her ear. ChatGPT suggested injecting some oil into it. She did and it became a huge problem, so that she had to go to the hospital.

I'm sure ChatGPT can be great, but take it with a huge grain of salt.

Now imagine what happens when a new programming language comes along. When we have a question, we will no longer be able to Google it and find answers to it on Stack Overflow. We will ask the LLMs. They will work it out. From that moment, the LLM we used has the knowledge for solving this particular problem. Over time, this produces huge moat for the largest providers. I believe it is one of the subtler reasons why the AI race is so fierce.

Reading the article, I thought "SNI sounds like its main purpose is to violate the privacy of the client". So I asked ChatGPT why it exists. Here is the reply:

SNI (Server Name Indication) was introduced so servers could host multiple HTTPS sites on the same IP address — a big win for the spread of encryption — but it comes with a major privacy flaw: it exposes the domain name you’re connecting to in plaintext during the TLS handshake. That means ISPs, WiFi providers, and governments can see (and block) which sites you visit even though the content is encrypted. While SNI wasn’t designed for surveillance, it’s now widely exploited for censorship and traffic shaping, as the British Airways WiFi example shows. The real fix is ECH (Encrypted ClientHello), which hides this information and restores end-to-end privacy once fully adopted. ECH is already supported in Firefox and Chrome behind flags, and as major CDNs like Cloudflare and Google deploy it more broadly, it is poised to become the default within the next few years.

I use Claude Code every day and find that it still requires a lot of hand-holding. Maybe codex is better. But just in my last session today, Claude wrote 100 lines of test code that could have been 20, and 30 lines of production code that could have been 5. I'm glad I do not have to maintain 300 kloc of 100% AI-generated code. But at the end of the day, what counts is velocity and quality, and it seems OP is happy. The tools certainly are useful.

Pi-thon

There's no contradiction. One statement is about the default, the other is about the possibility.