Jaime Pascual 👨‍🍼

I had caps lock on when telling AI to "KEEP WORKING" after a closed connection... :S

Added a 3D view to the LLM pareto frontier with speed as the third axis (avg across OpenRouter providers). Next stop: Comparing this over time. Demo link: llm-pareto.kaidev.io?view=3d&tour=1

Cerrar todas las pestañas de Chrome era el “he terminado por hoy” de antes. Ahora es cerrar todas las sesiones de agentes.

Looking to cross-reference more data! What other benchmarks should I add to the mix? Which ones do you use?

This whole runaway AI narrative got me thinking. What if they’re not really worried about human extinction but something much more mundane. Civil and criminal liability. We’ve already seen models used in crimes, sometimes by the labs themselves.

Spent the whole weekend fighting a model: abliteration, hydra heads, LoRA, answers leaking into the thinking channel… Then I started talking to AI, trying to understand what went wrong… it just found it. Pre-converted for my exact setup. I am both relieved and furious.

This setup yields me around 40-70 tokens per second with a max capacity of 120k context length. The one I'm actually on now does 100-150, even 200 tokens per second, and 240k context length. Also playing with fine-tuning it for specific tasks.

Now I can say thank you without losing tokens 😌only context

Qwen3.8-27B locally with lm studio + Open WebUI + proxy to the outside workd + some tokens and https for security = hell yeah!! Next step: lest make all aitonomous tasks cheaper than ever, security and code reviews, log scanning and summary, business metrics summary,… sky is the limit!

Power metal 🤘

👏👏👏

We are now orchestration team builders

Learn something new everyother day about business, clients, media, the context, anything and everything, this will expand your vision and your toolset. Do not invest time only on n the project/s, invest also on yourself.

Throwback to a gem of an article that starts with one of the hardest lines in computer literature: > “Every line of code is written without reason, maintained out of weakness, and deleted by chance” Jean-Paul Sartre’s Programming in ANSI C. programmingisterrible.com/post/1392226...

Exactly this is what I feel this days (months)

Every time opus says "compressing"

Me: Hey LLM can you reduce this code? LLM: We could use this well stablished, defined and already used on the code base function, but the actual code is minimal now (3 lines). Yo know what is also minimal not having those 3 lines.

I spent the last year learning a hard truth: high-quality work and hitting every metric doesn’t always lead to success. Macro shifts or a client’s subjective mood can easily displace months of hard work. I used to think the effort guaranteed the outcome, but the reality is much more fragile.

traverse

This is getting out of hand. Using an LLM doesn’t make you a developer; it makes you a child with a crayon. You’re doodling on paper and expecting high results, but your work belongs on a fridge at most. I’m tired of the delusion.

For once I did com up with and idea, execute it with AI and then publish it, no second thoughts, just raw willingness to benchmark something and then publish it. github.com/KaidevLabs/l... Something stupid and silly but useful to me.

Two Factor Authentication more like 30 minute login

My client red flags

When does your brain start hitting you with an unstoppable amount of ideas? For me, it’s usually the third day of vacation, and it doesn't let up until day five. Anyone else get the "holiday brain" spark?

Skills is just a way to spend more tokens per call change my mind

Ouch

We spend a bunch of tokens now so that we can spend more tokens later to fix the broken tokens now.

Crystal clear!!

OpenCode users hit me with your best workflows, plugins, practices,…

25 years ago… DS was good but this was peak design

Get this man a medal or something!!

Nvim + opencode + gemini => 👌

We are all Leroy the Masochist

Has anyone coined the term "vibe testing" yet? Can I claim it? Jokes aside, how do you keep practicing TDD in this world of LLMs and AI? For me, it starts with asking for the test first, but 100% of the time the assistant or agent just goes ahead and passes the test (moves to green) automatically.

Reading a lot of different opinions and budgets... (on Telegram) Now my question went from budget to "I'm even using the right tool?" What to do you use the most? VSCode, Cursor, Claude code, opencode,... 👂

Whats your AI monthly budget? I’m talking agents, chats, Copilot,…

So simple it’s genious

Growing is understanding that rewriting a complex system from scratch to "fix complexity" is just hiding the complexity until it catches up with you.

At least someone knows what they’re doing