adlrocha

The prouder I am of a post, and the more research work it required, the less impact it has.

GLM-5.3 and Qwen3.8-27B shipped on the same day.

GLM5.3 and Qwen3.8-27B performance improvements previous generation is explained through post-training...

I am late this week, and I just finished writing my post for tomorrow.

Actually this could also apply easily to marriage.

The best thing that you can do in business is to do what Munger did with Buffet.

You can see clearly how some companies are ADHD.

Turns out I am not rrhh e only one that prefers reading open model outputs.

I don't know if it is a skill or a curse, but I seem to be able to eventually make all agents agree with me :) They call it psycophancy, I call it perseverance.

The releases of GLM5.3-Flash and Qwen3.8-Flash yesterday marked AI independence day.

There is no way you can't love @posthog's marketing.

I would never have thought I would see myself tweeting this: "I find way more pleasant to read GPT-sol outputs than Opus'.

A monopoly through vertical integration is brewing

And this, kids, was the day that Arch’s new package manager was born (Or not, but it would be the start of a nice story) > DHH: Update: I heard back!

I've been having a weird feeling these past few weeks.

So you are telling me that with a bit of pruning I'll be able to run GLM5.3-Flash at home?

This is fucking huge!

I approve > tobi lutke: I’m thinking about banning Claude code at Shopify until they change their mind and read AGENTS.md and .agents/skills etc.

From a notebook to a transcript, and focusing on the depth layer...

August Off Series Vol.

Or how to keep your agent’s context fresh as mint as they loop...

A round-up of my recent theses, and a wrap for the season....

How to REAP your experts to fight the upcoming open-weight crunch...

So it goes. > zerohedge: Tech CEOs: Just Kidding About That Jobpocalypse https://www.zerohedge.com/ai/tech-ceos-just-kidding-about-jobpocalypse

I finally read how the X algorithm works.

Been there, done that.

Y gracias a @PabloGrueso, este año asistiré a mi primera #TRGCON.

My plan was simple: hedge my family’s access to intelligence by building a local AI box on my desk.

On surviving the Great AI Capital Shift, and our last chance to make useful work...

Ok, fucking excited with what we are about to build

I am not proud of tomorrow’s post.

I am the only one that hadn’t had the time to test Fable?

So it goes.

So it goes.

A good analogy that I sometimes use to think about LLMs is that they are like bloom filters over all their training knowledge

Finally > ClaudeDevs: Claude Desktop is now available on Linux (Ubuntu and Debian) in beta.

So it goes. > : Igor Os:

More disgusting LLM-generate mouthfuls from today...

Before LLMs, whenever I had to do something in the computer more than twice I wrote a script for it.

A lot can be learnt if you read all the output of you agent’s work.

New batch of excuses.

You see here two ends of the spectrum pulling to their own camps.

The math says that I shouldn’t be building my own AI rig, and why I am building one anyway...

The Tesla flywheel and why we won’t get the Humanoid butler anytime soon...

Mondays are quiet.

So it goes. > Hedgeye: The LLM Token Expenditure Index is now back to its January level, when the agentic AI market exploded

This week has been a fun (and long) one to write.

Fable’s release and on strengthening the case for AI independence...

So it goes. > sdmat: Anthropic

On the need of becoming AI independent.