Alexandru Nedelcu

#DeepSeek v4.1 Flash is an excellent worker, but it's benchmaxxed, and in agentic work it's inferior to GLM-5.3 or GPT-6.1 Sol. It has the glaring problem that it's not following its prompt, making it a terrible orchestrator. For me this casts doubt on all benchmarks I've seen.

#AI is so bad at ... writing plain text (READMEs, ADRs, code/configuration comments, etc.) that I'm having issues trusting them with anything else. Like really bad, I don't understand how people put up with it.

You have carefully structured Markdown files with rules? Telling the #AI to add another rule will make it take a dumb at the beginning of the file, treating that rule as the most important in the world; or it will place it in the wrong section, guaranteeing confusion.

#AI is so bad at ... writing plain text (READMEs, ADRs, code/configuration comments, etc.) that I'm having issues trusting them with anything else. Like really bad, I don't understand how people put up with it.

Google is locking down #Android in the name of safety, despite the fact that mallware is mostly distributed through their own Play Store. Google decides what you run on your devices. It won't block mallware, but it will block FOSS or their competition. keepandroidopen.org/

Complaints about solving simple algorithms in interviews for software engineering jobs are evergreen. Algorithms test your logic and communication skills. Software engineers are expected to solve problems that are easy enough for high schoolers to solve 😉

What if "you being the bottleneck" is actually the recipe for creating good software that's actually used? You all had a full year of agent-driven development, so what do you have to show for it, other than personal utilities that nobody else will touch?

Microsoft's #Copilot coding agent sucks, as it has a silly system prompt that's required and you can't override it. Screenshot shows 3 injected rules preventing efficient delegation. Note agent has no idea how many tool calls it will use, or it would make a script / 1 tool call.

I don't believe that AI will be worse for human cognition than social media. For one, you end up creating instead of mindlessly consuming short-form pre-digested content. Making builders out of people sure sounds more fun & productive. Also, my own learning has accelerated.

There will be negative consequences, of course, like with all software we've ever built. For example, people relying on their GPS app no longer learn to read maps or to navigate their city without GPS apps. OTOH, travelling being a skill, it prevented many from exploration.

When you hear about a 10x difference in throughput b/w a #Rust port versus a battle-hardened GC-managed runtime (#JVM, #dotNET, #Go), it can only be an apples vs oranges comparisson. Whether it's avoiding work via caching or algorithmic complexity...

A shared idea these days is that SOTA models improve faster than you can improve your custom harness setup, so there's no point in having a custom harness. This misses the point. (1) Correcting an agent's behavior remains relevant for cheaper models, which make better loops.

“The hate on Go is stupid” No it isn't. It's fine to hate tools. We build on our intuition, intuition is experience, intuition is good. Ofc, expertise requires opinions that are context dependent. E.g., I understand why some people like Go for their projects. But hating it is fine, healthy even.

My son is in high-school. I keep telling him that to be a good software engineer, he needs 2 abilities: - Reasoning (math, obviously) - To read and write prose; to express his ideas and thoughts with words, efficiently. This has always been true, and will remain so.

When working on personal projects, it's a good idea to be frugal, to have a sustainable setup. Best "offers" right now are these: ‣ OpenCode Go ($10-$20) for $60-$120 worth of access to DeepSeek v4.1 Flash ‣ Mistral Pro ($15) for €255 worth of access to GLM-5.3

It's important to be able to live w/ chosen LLMs at API prices. I also have a ChatGPT Plus ($20) subscription, but I'm close to unsubscribing, as it's designed to upsell you on Pro ($100 or $200), and GPT-6 Astra or even GPT-6.1 Sol are not sustainable at API prices.