iiinova — where your digital self keeps evolving

New LLMs announced and I ... couldn't care less. If you're not a vibe-coder, truth is that you can have a productive workflow that depends more on your harness configuration, with newer models only competing for token efficiency, in a race to the bottom, where they belong.

131

Alexandru Nedelcu @alexn.org · 15h

Agentic programming started being viable at Opus 4 / GPT 5. We've had a year of experience with it, not a lot, but things are moving fast. Where are the great apps built with it? Great games? Do you have anything to show for it, other than personal utilities? Why not?

120

Alexandru Nedelcu @alexn.org · 29/09/2026

#GraalVM Native Image is great in principle. Problem is it doesn't mix with runtime reflection well, as #Java and #Kotlin libraries are a minefield of runtime reflection; even with all facilities (reachability metadata repo, tracing agent, good tests) I still bump into issues.

240

Alexandru Nedelcu @alexn.org · 28/09/2026

The Web is the ultimate platform because: ‣ Browser is always available, users have no installation requirement; ‣ Browser is the best sandbox; ‣ It can be quite fast. People don't target native apps, as they suck for a myriad reasons, only one of which being the extra effort.

030

Alexandru Nedelcu @alexn.org · 27/09/2026

English isn't optimal for AI, being too abstract (every word is an abstraction), a bottleneck in the way of achieving real gains. Soon you'll communicate with grunts, barks, chirps, or body language, like showing teeth, chest pounding, touching or grooming yourself. #ai #llm

020

Alexandru Nedelcu @alexn.org · 27/09/2026

Age of AI is bringing with it some of the dumbest takes imaginable. Newer samples: - Generated code can be assembly; - You no longer need unit tests, or DRY, or abstractions. This is the result of “cargo cult programming”, and of doing stuff based on vibes, which precedes AI.

220

Alexandru Nedelcu @alexn.org · 23/09/2026

I like #OpenCode because you can just instruct the AI to build you plugins for adding functionality and the experience is pretty good (also, #EffectTS FTW). Here's another simple plugin that adds the Brave Search API as a `websearch` (tool) provider: github.com/alexandru/...

github.com

GitHub - alexandru/opencode-brave-search: OpenCode v2 plugin adding the Brave Search API as a `websearch` (tool) provider

OpenCode v2 plugin adding the Brave Search API as a `websearch` (tool) provider - alexandru/opencode-brave-search

000

Alexandru Nedelcu @alexn.org · 22/09/2026

My article reached HN 😎 alexn.org/blog/2026/0...

Screenshot of a Hacker News page. Top story: “AI Has No Wisdom and Neither Will You” from alexn.org. It has 178 points, posted by dimonomid 1 hour ago, with 223 comments. Above is the orange Hacker News navigation bar with new, past, comments, ask, show, jobs, and submit links.

040

Alexandru Nedelcu @alexn.org · 21/09/2026

My #OpenCode configuration now has a custom plugin that evaluates shell commands for safety via #Jev. It evaluates if shell expressions are within the agent's permissions, according to its configuration & system prompt. github.com/alexandru/...

Code editor showing a JSON configuration file. The snippet defines a "shell-safety" plugin with options like endpoint and model, and configures "Explorer" and "Librarian" agents.

000

Alexandru Nedelcu @alexn.org · 18/09/2026

“I haven't edited code by hand since…” That's like saying you stopped reading & writing, and that the agent is doing all the writing (and thinking) for you. It's all very novel, but I'm pretty sure that all people going down this path will regret it.

130

Alexandru Nedelcu @alexn.org · 15/09/2026

On every single HN thread talking Java there's at least one top comment lamenting type erasure. .NET generics reification is only needed b/c of runtime reflection. This doesn't come up in languages with type classes at all. Nobody asks for reification în Rust or Scala.

041

Alexandru Nedelcu @alexn.org · 15/09/2026

"If you don't pay, you are the product" has stopped making sense for software long ago. Even if you pay, your data is still getting sold. If you don't want to be the product, you need to read the fine print and also prefer Open Source from volunteers (which you should sponsor).

020

Alexandru Nedelcu @alexn.org · 13/09/2026

It's 2026, we have AI reading and writing for people, we have powerful voice-to-text engines that work better than ever before, such that we can speak instead of type, and yet, one of the best things you can do for your career is (still) 🥁 ... touch-typing.

100

Alexandru Nedelcu @alexn.org · 12/09/2026

VS Code is better than Zed in every possible way that matters. I wish it wasn't so. Even the often exaggerated problem of VS Code, being built on top of a browser, is a pro, as its design is better & has more rendering abilities. And Zed sucks at startup too.

101

Alexandru Nedelcu @alexn.org · 09/09/2026

This is how a setup with delegation to cheaper models looks like, pics show a single session. It was harder to balance back in the days of Haiku 4.5. A dumb model can do damage that the smarter model won't fix, but nowadays, Luna, or DeepSeek are quite capable. #LLM #AI

A table titled "Agents" listing four agents sorted by cost. Orchestrator is the most expensive at $100.04, followed by Explorer at $18.53, Junior at $10.77, and Librarian at $4.06. The table includes token usage and message counts.

A table titled "Models" listing usage stats for two OpenAI models. Row 1: gpt-5.6-sol with 3.3M input, 243K output, 188.6M total, 187ms/1K, and cost $98.88. Row 2: gpt-5.6-luna-fast with 55.7M input, 2.3M output, 885.3M total, 99ms/1K, and cost $33.37.

010

Alexandru Nedelcu @alexn.org · 08/09/2026

I've had my own skepticism or dread regarding LLMs, however, I can now say that w/ LLMs ... programming is more fun. That's because we've been dealing with a lot of shit that's now automate-able. E.g., ever tried publishing a library to Maven Central? If so, does it still work?

120

Alexandru Nedelcu @alexn.org · 06/09/2026

Commenting on the latest LLM models, as that's what boring people do... I dislike GPT-6 Astra. Compared with GPT 5.6 Sol, improvements are only marginal, if any, while being significantly more expensive (in my actual use, not benchmarks). AI may have peaked.

120

Alexandru Nedelcu @alexn.org · 04/09/2026

Many "best practices" that are absolute BS in software development, and I have to suffer from AI slop now following them. Ex: don't use magic strings and numbers. The string is more clear than the name replacing it, it never changes, but if it does, it's just search&replace ffs

040

Alexandru Nedelcu @alexn.org · 03/09/2026

Tech news is dominated by LLMs these days, but it's same old, same old, and I'm actually more excited about the new #Scala 3.9 LTS 😍 www.scala-lang.org/ne...

scala-lang.org

Scala 3.9 LTS released!

060

Alexandru Nedelcu @alexn.org · 03/09/2026

Funfix Database — v0.5.2 release. Provides components built on top of RDBMSs. Currently, it provides a very useful "delayed queue". #Java, #Kotlin and #Scala library. Plays well with virtual threads & Scala version is powered by Cats-Effect. github.com/funfix/dat...

github.com

Release v0.5.2 · funfix/database

This patch release improves Scala interoperability, expands Scala documentation, and updates the build and release tooling. Changes OfferOutcome.fromJava now throws IllegalArgumentException when i...

000

Alexandru Nedelcu @alexn.org · 01/09/2026

No paradigm does modules and components better than OOP. Pick your favorite non-OOP language; admit that its abilities for expressing components, even modules, suck, and any defense you can have rely on hacked-in virtual dispatch tables (poor man's OOP).

020

Alexandru Nedelcu @alexn.org · 30/08/2026

Funfix Tasks v0.5.0 #Java #Kotlin #Scala This is a library exposing `Task` and `Fiber`, JVM datatypes meant primarily for JVM interop. NOTE: Trying to have scheduled releases for all projects I care about, so polishing up the maintenance infra. github.com/funfix/tas...

github.com

Release v0.5.0 · funfix/tasks

This release changes the asynchronous task API and improves cancellation handling. Breaking changes AsyncFun<T> now receives a single public Continuation<? super T> and returns void. Task.fromAsync...

000

Alexandru Nedelcu @alexn.org · 28/08/2026

Continuations4s version 0.2.0: just a maintenance release. #Scala #Scala3 This is a micro-library providing continuations, for Scala 3, supported across JVM, WASM and Native, extracted from the Gears library. github.com/funfix/con...

github.com

Release v0.2.0 · funfix/continuations4s

Continuations4s provides low-level delimited continuations for Scala 3 on JVM, Scala.js with WebAssembly JSPI, and Scala Native. Version v0.2.0 is just maintenance release. There are no library sou...

000

Alexandru Nedelcu @alexn.org · 27/08/2026

Migrated the project at work to #Scala 3. ~600,000 lines of code. Used an AI-driven loop; took a week with continuous interventions from me; superficial changes, few problems discovered. The project already had a codebase preparing for it (no problematic deps, -Xsource:3-cross).

080

Alexandru Nedelcu @alexn.org · 23/08/2026

I may change my tune on significant indentation in #Scala 3. I've witnessed GPT 5.6 Luna (high) refactoring code, then struggling to balance the parens & braces, making a complete mess. Understanding & modifying the structure of your program have different challenges...

110

Alexandru Nedelcu @alexn.org · 21/08/2026

To add in AGENTS.md if you're using GPT models: "- Banned words: seam"

010

Alexandru Nedelcu @alexn.org · 17/08/2026

The web is filled w/ connoisseurs of the fine differences b/w various LLM models. Story goes that many can't tell good code from bad. Well, I often can't tell the difference (e.g., DSv4 vs Sol vs Terra vs Luna) b/c too often they yield the same shit.

230

Alexandru Nedelcu @alexn.org · 11/08/2026

We still need math. It doesn't matter if AI is able to prove math theorems. The analogy of human calculators being replaced by digital ones misses the point that people able to do calculations in their head are still faster than you at problem solving or spotting mistakes.

010

Alexandru Nedelcu @alexn.org · 08/08/2026

Plenty of career angst going around these days. Thinking that via automation we end up doing the same work with less is a mistake. Human desires are infinite, the pie is always growing. We have better things to do than reinventing the same CRUD apps.

020

Alexandru Nedelcu @alexn.org · 03/08/2026

"Don't be a meat proxy" Complete agree with this article, and I'd add that this trend is poisoning online forums. So if you're one of those people, you're making your work environment and the Internet a worse place. gruhn.me/blog/2026-08...

gruhn.me

Don't be a meat proxy

061

Alexandru Nedelcu @alexn.org · 03/08/2026

#OpenCode Go is a good subscription that I've been recommending, but, with the latest updates for DeepSeek, I'm not sure that they know what their value proposition actually is (DeepSeek's official API is cheaper). That, plus they just broke all my flows. Thanks, but no, thanks.

000

Reposted by Alexandru Nedelcu

odersky.bsky.social @odersky.bsky.social · 27/07/2026

New blog post that describes the ideas behind our award-winning paper on securing AI agents: martinodersky.substack.com/p/tracked-ca...

martinodersky.substack.com

Tracked Capabilities for Safer Agents

This post describes the ideas behind our paper “Securing Agents with Tracked Capabilities”

0123

Alexandru Nedelcu @alexn.org · 23/07/2026

Automated formatting of code is possible & encouraged, but automation always has unintended consequences. Learning to format is a low effort to learning order & taste. Reasoning requires order. Scalafmt or AI won't fix the disorder of your code or your mind or your poor taste.

010

Reposted by Alexandru Nedelcu

Alexandru Nedelcu @alexn.org · 21/07/2026

"On Performance, GCs and Rust" Consider this claim: Rust programs have better performance than Java/Go/.NET programs. Do you agree with it? #Programming #Languages #Java #JVM #Rust alexn.org/blog/2026/0...

alexn.org

On Performance, GCs and Rust

On programming and personal projects

352

Alexandru Nedelcu @alexn.org · 21/07/2026

alexn.org

On Performance, GCs and Rust

On programming and personal projects

352

Alexandru Nedelcu @alexn.org · 20/07/2026

In a free market, motivation doesn't matter, only the price does. And the only thing that can threaten competition & commodification of #AI is state regulation. Reject the FUD related to open-weights LLMs by Chinese companies. You only need trust for proprietary services.

020

Alexandru Nedelcu @alexn.org · 16/07/2026

Argentina's performance was spectacular. Others get demotivated when conceding a goal, yet their motivation only grew. I'm rooting for Spain next, they are more technical, but it's going to be hard to win against such strong motivation, a force of nature. #worldcup #football

000

Alexandru Nedelcu @alexn.org · 16/07/2026

I've never seen an engineer "attempt" to override $HOME and then delete it. What a silly thing to do. "The Daily WTF" notwithstanding, meat-space having stories of blunders, I've never worked with engineers that get this silly. But hey, at least this is fast & cheap 😏

Tweet by Tibo saying:

"On file deletions. We’ve investigated a handful of reports where GPT-5.6 unexpectedly deleted files.

What we have found is that this most commonly occurs when:
- Full access mode is enabled and codex is run without sandboxing protections, including without auto review being enabled
- The model attempts to override the $HOME env var to define a temporary directory.
- The model makes an honest mistake and mistakenly deletes $HOME instead."

040

Alexandru Nedelcu @alexn.org · 16/07/2026

I have a Macbook Pro 2019 with an Intel processor. Given the forced obsoletion after macOS Tahoe, I deleted macOS from it. Installing Linux would be difficult due to T2 chip, so I went for Windows 11 and ... it's not bad. WSL is decent, and 32-bits games still work 🤷‍♂️

110