Preslav Rachev

I’m considering getting one of the smaller reMarkable Paper Pro Move devices to try it out. I was a big fan of the RM2, and to this day feel sorry I sold it in a rush. People have had mixed feelings about the Move, but were equally wrong about the SuperNote (which I found a complete disappointment)

Large language models are modern day’s Schrödinger’s Cat. They see non-deterministic, but down to the core, they are very, very deterministic. Anything based on pseudo-randomness is deterministic by nature.

but I’m curious if they really want to task an LLM to pilot a drone (there are far better models for that) or for something else. Or, is it simply a media ploy to scare the nation.

When one says the US is going to use Anthropic’s or OpenAI’s LLMs for military purposes, what do they really mean? I’ve been using these tools for a few years now, and certain use cases pop to mind ... 1/2

1995: From Batman Forever’s cinematic design to HTML tables 1995 begins with web designers creating cinematic experiences using images and browser tricks, and ends with the arrival of table support in Netscape Navigator — giving true control over layout.

“If 1994 was when the Web became a publishing medium, then 1995 was when the Web truly marked itself as a unique expressive medium. The Web became a place — a destination — rather than a mere repository for documents.” — @ricmac.cybercultural.com

cybercultural.com/p/1995-web-d...

Writing about Agentic Engineering Patterns I’ve started a new project to collect and document Agentic Engineering Patterns—coding practices and patterns to help get the best results out of this new era of coding agent development …

@simonwillison.net is the first I've heard use the term "agentic engineering", which feels like a much better description of how I use tools like Claude Code (as opposed to "vibe coding", which implies no human architecting the result). Great stuff in this post:

simonwillison.net/2026/Feb/23/...

spec: generic methods for Go · Issue #77273 · golang/go Proposal: Generic Methods for Go A change of view. Background For clarity, in the following we use the term concrete method (or just method when the context is clear) to describe a non-interface me...

🚀 "spec: generic methods for Go" has been accepted!

You will soon (1.27?) be able to declare (on concrete types only) methods that introduce type parameters, i.e. type parameters other than the ones (if any) that come from the method's receiver.

github.com/golang/go/is...

#golang

It works, but it requires a lot of time and patience.

Don’t Fight the Weights When your context goes against a model’s training, you struggle to get the output you need. Learn to recognize when you’re fighting the weights so you can do something different.

It helps to think of LLMs as bell curve statistical optimizers. Most code they've been trained on, would likely simply concatenate strings using plain operators. And, this is what they pick most of the time.

Teaching them new tricks is called "fighting the weights." www.dbreunig.com/2025/11/11/d...

I know the feeling. You subscribe to 100 blogs, where you only really read 1%. This is one area we've been trying to improve with the release of @feedle.world - instead of having to subscribe to hundreds of RSS feeds, we create ones per topic, sourcing posts from thousands of blogs and podcasts.

Plenty of businesses failed before AI, and plenty will fail with it. The tooling may have improved, but reality is as harsh as ever. A bad idea with AI is still a bad idea, just shipped faster.

Но пък доброто е, че всеки път се получава нещо ново, и по-добро от предното.

Twenty years is proper perspective. I've landed in a similar place - the agent needs to be on a tight leash with clear boundaries, not given free rein. The moment you stop reviewing its output is the moment it quietly introduces chaos.

Twenty years of context is exactly what makes the difference. The bit most people miss is that "remaining in control" also means owning the context the agent works with - if your codebase knowledge lives on someone else's server, you've already ceded control before the first prompt.

What you see from most folks leaving Claude Code to work alone, is some random mid web app that looks like every other. Remarkable nonetheless, but also, very mid, unless you really step in the product owner shoes, and do a ton of AI hand-holding.

I’ve been programming for over 20 years now, the last three of which, using AI tooling almost exclusively. Here is the bottom line:

It’s not.

You can achieve great results with agentic tooling, but only if you consider the agent your junior programmer assistant.

You must remain in control.

Welcome to the #golang club!

FTR, if you write code that you'd like your future self to understand, yes, more lines do actually make it easier. Go was not meant for writing ergonomics, but for that of reading (especially by newcomers).

Plus, AI agents have finally solved the writing part for us.

You can use Opus for Plan mode and Sonnet for everything else by typing "/model opusplan" in Claude Code. It'll even show up in your "/model" list after that.

WHAT IS THIS SORCERY AND HOW DID I NOT KNOW ABOUT IT BEFORE??!??!?

#buildinpublic #claudecode

Think about it the next time you ask ChatGPT to write a LinkedIn post for you. We know you did.

The result is a "JPEG of thought," and like all forms of digital compression, taking the raw data and compressing it is easy, but going from a compressed result to the raw, natural beauty of the original is nearly impossible.

Through what's called "semantic ablation," LLM models systematically discard rare "tail" data to gravitate toward the center of the Gaussian distribution, maximizing probability at the expense of unique signal.

Have you ever asked yourselves, how ChatGPT can be syntactically perfect and the same time intellectually mid? It's a statistical problem, really.

---
#chatgpt

Not, if you care about your memory usage: "For production scenarios, Go offers balance of performance and resource efficiency, making it ideal for cloud-native and containerized deployments. Java provides marginally better latency characteristics but at the cost of 12x higher memory consumption."

True, but the thing is, you see those two sides right away. With most other languages I've used, they present themselves as 80/20: you get the 80% "gateway hook" in 20% of the time, but scaling beyond that always results in hacks and trade-offs.

#Golang is fair in that regard. Ugly, but WYSIWG.

a man wearing sunglasses and a colorful shirt is making a surprised face and saying wow . ALT: a man wearing sunglasses and a colorful shirt is making a surprised face and saying wow .

Yeah, sure, you can use Claude Code to generate Go code for you, but has it ever occurred to you to ask it to generate tailor-made code generators that take your existing code and instantly give it superpowers?

Case in point, type-safe URL routing, validation, serialization, sum types etc

#golang

Don't give up. It's always been like that. It's the natural evolution of things. We just need to adapt and find a higher level of problems to solve.

For the record, you can manually invoke these stats in a CC session, buy uisng the `/usage` command. By comparing the two, I'd at least know if I am going too fast, and I'd introduce a break in my plan.

---
#programming #claudecode

My problem with Claude Code limits these days aren't the 5-work blocks, but reaching my weekly limit way too soon. And, while I haven't found and easy way to fetch my current weekly limit programmatically, I quickly coded an xbar script that would show me the % elapsed till reset, and I'd compare.

Още един добър цитат: "Близките на загиналите още не са „приели смъртта“ – не са погребали и оплакали своите родственици – и ми се струва, че е редно за малко да замълчим."