Georg Unterholzner

Here's a glimpse of how software engineering is going to look soon. I wanted to spend more time with my newborn daughter and less time sitting in a terminal watching an AI type. So I built kipppunkt/build. Today I'm sharing it. github.com/kipppunkt/bu...

When I first saw it working, me reviewing from my phone while the AI iterated on my feedback, I understood this is the way forward. We don't need to reinvent the wheel. We have decades worth of tools and practices. We just need to let AI meet us where we already are.

Here's a glimpse of how software engineering is going to look soon. I wanted to spend more time with my newborn daughter and less time sitting in a terminal watching an AI type. So I built kipppunkt/build. Today I'm sharing it. github.com/kipppunkt/bu...

"Isn't that just a ralph loop?" Kinda, but not really. Ralph loops are basically waterfall. You hit go and hope for the best and get no feedback until it's done. They fail the same way waterfall always has.

Here's a glimpse of how software engineering is going to look soon. I wanted to spend more time with my newborn daughter and less time sitting in a terminal watching an AI type. So I built kipppunkt/build. Today I'm sharing it. github.com/kipppunkt/bu...

AI-generated unit tests suck. Claude overshoots. GPT-5 is lazy af. No model has proper common sense. I tinkered around for a long time to get GitHub Copilot to actually work as expected. The result is this workflow.

It uses many of GitHub Copilot's capabilities (instruction files, prompt files, some tools & configs) for prompt and context engineering. With that, it's reliably able to one-shot entire test suites in my day-to-day work. See for yourself (select 2x speed):

AI-generated unit tests suck. Claude overshoots. GPT-5 is lazy af. No model has proper common sense. I tinkered around for a long time to get GitHub Copilot to actually work as expected. The result is this workflow.

You'd be mistaken if you thought pigeon droppings on the floor were just a random annoyance. Quite the opposite! They're a real-time heatmap showing you the probability of being hit by pigeon poop at any given spot. And you thought all that statistics knowledge was useless!

Basic question but I can't figure it out. How are these two design patterns called? In the first, the component structure is fixed, and behavior is controlled *only* through props serving as config. In the second, an extension point via the children prop allows for more flexibility.

I was frustrated with the lack of clear advice on setting up copilot_instructions.md. Most guides are vague, or worse, generic AI slop that *hurts* Copilot's performance. So I experimented and found better ways to make Copilot *actually helpful*. 🧵

Agent mode *bonus* tip: Enable auto-approval for tools by adding "chat.tools.autoApprove": true to your settings.json. Just be aware that you’re giving Copilot a free pass. Use it wisely! 😅

I was frustrated with the lack of clear advice on setting up copilot_instructions.md. Most guides are vague, or worse, generic AI slop that *hurts* Copilot's performance. So I experimented and found better ways to make Copilot *actually helpful*. 🧵

Most advice skips this: Be model-specific. If Copilot repeatedly makes the same mistakes (e.g., overly verbose code or missing edge cases), include correctional hints in your file. Example: "Focus on simplicity, avoid over-comprehensive solutions."

I was frustrated with the lack of clear advice on setting up copilot_instructions.md. Most guides are vague, or worse, generic AI slop that *hurts* Copilot's performance. So I experimented and found better ways to make Copilot *actually helpful*. 🧵

Most devs either skip copilot_instructions.md or fill it with generic fluff. But irrelevant instructions waste Copilot’s context space and reduce output quality. Focus on non-obvious details: unique constraints, domain rules, and project specifics.

Large Language Models (LLMs) with "reasoning" capabilities are super hyped, but new research from Apple shows their "thinking" is often an illusion. Beyond a certain complexity, they fail completely. Key points in the 🧵.

This isn't just about benchmarks—it's about fundamentals. Current "thinking" models are impressive but eventually fail, and users won't know at which task or complexity level. Scaling model size won't solve this; we need new paradigms for robust reasoning.

Large Language Models (LLMs) with "reasoning" capabilities are super hyped, but new research from Apple shows their "thinking" is often an illusion. Beyond a certain complexity, they fail completely. Key points in the 🧵.

Researchers even handed models step-by-step algorithms (e.g., for Tower of Hanoi). The result: identical failure points. LRMs couldn’t reliably execute prescribed logic—suggesting their "reasoning" is often pattern-matching, not general intelligence.

Large Language Models (LLMs) with "reasoning" capabilities are super hyped, but new research from Apple shows their "thinking" is often an illusion. Beyond a certain complexity, they fail completely. Key points in the 🧵.

The paper identifies three distinct "reasoning regimes" in puzzles like Tower of Hanoi and River Crossing. Frontier LRMs (Claude 3.7 Thinking, DeepSeek-R1) outperform standard LLMs only in the medium complexity zone. Beyond this, their accuracy collapses regardless of compute.

Large Language Models (LLMs) with "reasoning" capabilities are super hyped, but new research from Apple shows their "thinking" is often an illusion. Beyond a certain complexity, they fail completely. Key points in the 🧵.