Aldehir Rojas

Coffee enthusiast

[Aldehir Rojas @aldehirrojas]

@iamMrDuncan I didn't find caching to be very useful. I ended up using mine in a separate storage pool for fast storage.

[Aldehir Rojas @aldehirrojas]

Opus 5 and now Fable 5.1 love to write scripts to perform edits instead of calling an edit tool. This feels like a shift towards “code mode.”

[Aldehir Rojas @aldehirrojas]

I am so sick of LLMs spamming whitespace trimming all over my code. I can't seem to find a reliable way to make them stop.

[Aldehir Rojas @aldehirrojas]

Websites that only let you log in via SMS or Email can fuck right off.

[Aldehir Rojas @aldehirrojas]

I’m going to miss Fable 😞

[Aldehir Rojas @aldehirrojas]

I replaced oh-my-zsh with my own minimal zsh configuration. It performs exactly the same with faster loading.

[Aldehir Rojas @aldehirrojas]

@iamMrDuncan On the local AI train, huh?

[Aldehir Rojas @aldehirrojas]

Easy enough to vibe code a tampermonkey script to add a DiffsHub link in a GitHub PR.

[Aldehir Rojas @aldehirrojas]

@badlogicgames @SIGKITTEN There was bug in llama.cpp that was recently fixed regarding this: https://t.co/RLGp23vaPv. I cannot speak for other inference engines though.

[Aldehir Rojas @aldehirrojas]

@badlogicgames The criticism here should be directed at llama.cpp for not streaming JSON argument values with certain models like Qwen.

[Aldehir Rojas @aldehirrojas]

@mitsuhiko @hampsonw Maybe you already know this, but these reasoning models are trained for "interleaved reasoning" and perform better with thinking on, since traces are retained between tool calls to help "guide" them.

[Aldehir Rojas @aldehirrojas]

To AI issue slop generators: I hope you feel embarrassed when the fix turns out completely different than your “comprehensive” report.

[Aldehir Rojas @aldehirrojas]

@mitsuhiko @antirez llama.cpp supports tool call parameter streaming for most models. The most common issue I've noticed is clients failing to send reasoning back from prior responses. It's not part of the OpenAI Chat Completions spec, so I understand why, but it destroys model performance.

[Aldehir Rojas @aldehirrojas]

@simonw The original template drops reasoning between tool calls, which hurts performance. llama.cpp has a custom template that fixes this, and I believe it makes Gemma 4 31B a really good coding assistant for its size.

[Aldehir Rojas @aldehirrojas]

Apparently I have a talent for hitting nerves. Might be time to run my messages through AI before I hit send.

[Aldehir Rojas @aldehirrojas]

I still see people not using gpt-oss to its fullest by passing along the CoT when tool calling. I did some simple testing to show how bad it performs without it. https://t.co/U5HtMYgXHn

[Aldehir Rojas @aldehirrojas]

@stolinski Don’t do it. You won’t stand VS Code ever again. The vim extension just doesn’t hit the same.

[Aldehir Rojas @aldehirrojas]

@hkhawaja Ollama provides structured output support in the form of JSON, but that doesn’t help in this scenario. I’m not aware of Ollama supporting custom grammars.

[Aldehir Rojas @aldehirrojas]

GPT-OSS 20B and Cline/Roo Code don’t work well together out of the box. Using llama.cpp and a simple grammar, GPT-OSS can be made usable. https://t.co/AQRGMHctYv

[Aldehir Rojas @aldehirrojas]

The 120B model doesn’t seem to exhibit the same issues, or at least not as frequently.

[Aldehir Rojas @aldehirrojas]

It makes me wonder if OpenAI recommends constrained decoding even in the absence of tools. If so, then I don’t see why this couldn’t be baked into inference servers.