The thesis that closes this arc: LLMs have reached, or are very close to reaching, the ceiling of the transformer architecture. And the good news is that nothing more than that is needed.
Last week I published a long post about Jev, TypeSafe’s hosted classifier that became all the rage in September. The short version of that post: Jev is a classifier, something that has existed for decades, and the “193x faster” amazement only exists because too many people were using chat LLMs as expensive, slow classifiers. I reproduced Paulo Câmara’s essay The student who marks answer A, introduced the open alternatives Clef and Laya, and to compare the models side by side I leaned on the Decision Index, a public ranking that already existed.
Yesterday, Monday, I was woken up at 8 a.m. by a loud, insistent beep. I knew it was the NAS before I opened my eyes. I ran to check, and nearly lost it: two drives marked as critical, in an array that can never have more than one.
Some folks from LUA Vision, the Brazilian company behind Genesys PI, reached out on LinkedIn offering test access to their model. That kind of thing usually doesn’t interest me, but this particular pitch, a Brazilian lab training a model from scratch instead of reselling a foreign model with a Portuguese layer on top, made me curious enough to run the test myself.
Every major open source project today ships with a CODE_OF_CONDUCT.md in the repository. The idea was sold that this existed to “protect minorities” and “create safe spaces.” In practice, what I watched happen over the last decade was something else: an instrument to cancel, persecute, and ban anyone who thought differently from whoever was in control. And the people leading those hunts, as a rule, were not the ones maintaining the projects’ code.
This is my thesis, and I know it’s unpopular. It took me ten years of patience to be able to write this piece. I stayed quiet watching this theater repeat itself year after year, waiting for the right moment to document everything I witnessed up close, and today I get to watch, with real pleasure, my own predictions come true one by one.
TL;DR: ai-memory 2.0 is out, with the open OKF format, local embeddings turned on by default, and real support for several agents and a whole team working on the same project in parallel.
Last week I wrote about the Discord censorship and Brazil’s Digital ECA law, the Digital ECA (“Estatuto Digital da Criança e do Adolescente”, the digital version of Brazil’s Child and Adolescent Statute): the ANPD (Brazil’s data protection authority, now also the country’s de facto internet regulator) ordered the Go Live feature shut down nationwide because end-to-end encryption prevents content surveillance, and in the same package came the first sentence enhancement in Brazilian history for committing a crime “using a VPN”. After that article, the question I got the most was the obvious one: “OK, so what do I do?”
Claude now ships with a stamp on it. Since August 2, 2026, Anthropic has been invisibly marking the text of its newest models, with the older ones migrating over the following months.
No off switch. The backlash came fast: a wave of subscription cancellations, with cancellation screenshots making the rounds on X.
Before you cancel on reflex, it’s worth knowing what the mark actually is. It works quite differently from what most people picture, it exists for a concrete reason, and it vanishes with an almost comic ease.
I’ve run four models: Qwen 3.8 Max, GLM 5.3, Gemini 3.7 Flash, and a 27B Qwen 3.8 running locally on my RTX 5090. One of them closed in on the leading group. Another pulled off the biggest jump this test has ever recorded. A third got caught cheating mid-test. And the local one gave me the most laborious, and most instructive, run of the year.
I spent 1 week and hundreds of dollars to run a real-world programming benchmark (real project back to back) through 30 LLMs to assert which is best and why.
ai-memory run keeps the same programming session while switching among Claude Code, Codex, and other harnesses, with searchable workstreams and integration with ai-jail and ai-usagebar.
This post has two parts. First, the news: Microsoft’s Majorana 2 announcement and why the physicists in the field remain with one foot (both, actually) behind. Then, the educational part: a decent explanation of Shor’s algorithm, because a lot of people repeat “quantum computers break RSA” without the slightest idea of the mechanism, and the mechanism is too beautiful to leave out.
Everything below is on my GitHub, with binaries, releases and install instructions.
In this post: clock-tui (with the google-calendar-tui and ghpending widgets), github-visualize, frank_geary, frank_scanlation, frank_lyrics, frank_type, distrobox-gaming and aitrepreneur-docker.
This week a curious controversy showed up in the open LLM world: Rio 3.5 Open 397B, presented by the City of Rio / IplanRIO in the Rio 3 Open announcement as an open model.
The accusation: in practice, the published checkpoint, meaning the model weights file, was a direct blend of Nex-N2-Pro and Qwen3.5-397B-A17B, without initial credit to Nex.
Another round of my coding benchmark. This time it’s three open source entries: Kimi K2.7 Code, GLM 5.2, and the MiniMax M3 that got open weights but that I can’t run at home no matter how hard I squeeze. Before the numbers, the usual context for anyone who parachuted in here, plus an update to the data center soap opera, because there’s news.
Anthropic shipped Claude Fable 5 this week, and before we get to my benchmark numbers (spoiler: a technical tie with Opus), the whole soap opera deserves a retelling, because the context matters more than the model. And it’s a good one, the whole nine yards: a secret model “too dangerous to release,” a closed consortium of big tech, an IPO on the horizon, and aggressive marketing.
Did you know many manga artists compose key moments as two-page spreads? If you read on typical fansub sites, you are probably missing many of them because most of those pages only show one long, vertical, single-page scroll.
Prettify Manga Reader restores more of the artist's original intent. It turns ugly, ad-bloated manga websites into a modern dark-mode reader with fitted pages, joined two-page spreads, keyboard navigation, and clean controls. It is generic by design, so it works across most manga sites instead of being hardcoded to one domain.
Opus 4.8 doesn’t feel much different from Opus 4.7, not in daily use and not in the benchmark. 95/100 against 97/100, inside the noise. It’s the fastest Opus I’ve measured, but the day-to-day experience is the same.
Grok 4.3, anecdotally, struck me as a bit more literal and strict about following the prompt. In the benchmark it improved substantially over the previous generation (Grok 4.20). It still doesn’t come close to Opus or GPT, but at least now it starts to be usable. 72/100, Tier B.
MiniMax M3, same story. The previous version (M2.7) was unusable, and the new one is finally at least usable. 78/100, Tier B.
All three land roughly in the band of a Sonnet 4.6 or a DeepSeek V4. One or two generations behind the new Opus and GPT.
I have a strong opinion about this, so let me just plant the flag: no open source project is ready to be published without three things. In order of importance:
Installation surface. The new user has to be able to install and try the tool with the least friction possible. One command, ideally.
Tests and automated CI. So contributions are easier and there’s common ground on what’s minimally acceptable to land in the project.
Documentation. The short, easy-to-read version (the README) and a more detailed set in docs/, so new contributors understand the architectural decisions.
I’m a subscriber to the MANGA Plus by Shueisha app on Android’s Google Play Store. For those who don’t know, this is Shueisha’s official channel (the Japanese publisher behind Shonen Jump) for reading their manga legally, with chapters releasing weekly nearly the same day they’re published in Japan. Two dollars a month gets you the entire catalog with an official English translation.
The problem is that MANGA Plus has no official website for reading. The subscription is locked to Android. There’s an iOS version too, but no web. A tablet helps, but nothing beats the comfort of a big, static screen when you want to read a whole One Piece arc in one go.
Most people just trust Google and leave everything there. Anyone who follows this blog knows I don’t trust anyone: if it’s in the cloud, it’s not mine. That’s why I keep a NAS at home to back up everything that matters. I don’t trust Google Drive, OneDrive, Dropbox, or Gmail.
After a while it clicked: Linux has had native email support forever. The modern protocols (IMAP, SMTP) are open standards. And the local storage format, Maildir, has been around and working well since the 90s. Thunderbird isn’t needed for any of this.
Every time I mention coding agent harnesses, somebody shows up asking: “but why don’t you use Pi?” or “why not Oh-My-Pi?”
I hate that question when it comes like that.
Not because I have anything against Pi. I don’t. The problem is that almost nobody explains the why. What is the concrete benefit? What problem does it solve better than Claude Code or OpenCode? Which use case is clearly better there? Which limitation does it remove?
In last week’s post, “Wrapping Up My AI Marathon: Success or Failure?”, I closed the first phase of my marathon. More than 600 hours of heavy coding-agent use, dozens of projects, hundreds of thousands of lines of code and documentation, and a pile of practical conclusions. Not marketing. Practical.
This post is a roundup of what I learned in that lab. A few of my tools, a few habits, a few recommendations. Nothing mystical. No “revolutionary new way to program.” Old engineering with new tooling.
Anyone following the blog already knows I use a whole bunch of different LLM vendors. The main ones are Claude and GPT, via Claude Code and Codex. I use each one’s harness because I’m locked into their subscription plans (Pro, Plus, Max), which come out way cheaper than paying per-token credits like OpenRouter. Even so, I also use OpenRouter to reach other models for testing, and I even have a Z.AI GLM subscription for small stuff.
The problem with juggling several plans is keeping track of how much I’ve burned on each. I’d been using claudebar, a Waybar widget that shows your Anthropic plan usage. I like it, but I wanted to see all my vendors in one place, glance at it and know where I stand. And I didn’t want to clutter my Waybar with a separate widget per vendor. So I built my own: ai-usagebar.
Five days ago I wrote a long post on coding-agent memory where I recommended agentmemory as the answer. After a week running it in personal production, I’m walking it back. This post explains what went wrong and the open source project I started building to fix it: ai-memory.
TL;DR: No. Across all three rounds of experiments I ran, mixing “strong frontier planner + cheap executor” loses to just using Opus 4.7 alone in a mature harness. Solo Opus 4.7 in opencode delivers Tier A (97/100) in 18 minutes for ~$4 pay-as-you-go. No multi-agent combination beats that on quality, and no combination is cheaper at the same time. The exception is Codex GPT 5.4 xHigh + medium executor, which drops from ~$16/run to ~$1-3/run while losing 3 quality points. Useful if you only have GPT in your provider stack. For everything else, let the frontier model decide when to delegate on its own, especially if you’re on a Plus/Pro/Max subscription.
This is possibly one of the most relatable benchmarks for real programmers, to compare most of the more popular open source and commercial LLMs available. You will be surprised by those findings and insights.