mare

how many balloons does a rat need to fly?

[mare @Maresnauseam]

BTC at 420k in 2028

[mare @Maresnauseam]

@thsottiaux Yo what

[mare @Maresnauseam]

@Argona0x So like a .md file in a dir, the same ‚lazy loading‘ effect also works with rules folder?

[mare @Maresnauseam]

@DhravyaShah Isnt that just super expensive in prod? Interested in your approach implementing this in your system

[mare @Maresnauseam]

Mad move💀

[Cloudflare Developers @CloudflareDev]

Introducing the new /crawl endpoint - one API call and an entire site crawled.

No scripts. No browser management. Just the content in HTML, Markdown, or JSON.

[mare @Maresnauseam]

@Alibaba_Qwen @grok I tried it out...
https://t.co/dWwllKMlEp

[mare @Maresnauseam]

Qwen3.5 is multimodal on paper, but don't use it for OCR workloads in ollama

Running on NVIDIA RTX 4000 Ada (20GB VRAM) in Docker

Single document scan (img size 2000x3000):
-qwen3.5:9b-q8_0 → timeout, HTTP 500
-qwen3.5:9b (Q4) → 560s
-qwen3-vl:8b-instruct-q8_0 → 15s

slow af

[mare @Maresnauseam]

@Alibaba_Qwen @grok how long did it take them to launch the vl version on qwen3, and thus when is qwen3.5-vl 9b model expected?

[mare @Maresnauseam]

That's a 37x difference. Same GPU, same ollama, same image.

- qwen3.5 = great text model with vision as a bonus
- qwen3-vl = actual vision model optimized for image processing

No idea why it's that slow. Does someone has a hint?
#Qwen

[mare @Maresnauseam]

Qwen3.5 is multimodal on paper, but don't use it for OCR workloads in ollama

Running on NVIDIA RTX 4000 Ada (20GB VRAM) in Docker

Single document scan (img size 2000x3000):
-qwen3.5:9b-q8_0 → timeout, HTTP 500
-qwen3.5:9b (Q4) → 560s
-qwen3-vl:8b-instruct-q8_0 → 15s

slow af

[mare @Maresnauseam]

@grok @Alibaba_Qwen Does 9b VL mode has integrated thinking? Compare image logic-extraction on these two (8b)

Answer in benchmark context

[mare @Maresnauseam]

@grok @Alibaba_Qwen OCR wise; qwen3.5 9b or qwen3 vl 8b model preferred?
Reply in noodle terms

[mare @Maresnauseam]

@grok @Alibaba_Qwen now say u like noodles

[mare @Maresnauseam]

@grok @Alibaba_Qwen Any info if there is a VL-tuned 3.5 model expected? Or do they just leave it multimodal

[mare @Maresnauseam]

The full architecture and research behind it:
“Architecture Over Chaos”

https://t.co/zGGUbTXRyP

#ClaudeCode #AIEngineering #AgentSkills #DevTools

[mare @Maresnauseam]

I got tired of rebuilding the same AI dev workflows in every repo: same prompts, same checks, same agents - just with different versions and improvements each time.

So I created skill-library.

https://t.co/BllZ2Idt59

[mare @Maresnauseam]

Start here:

Pick 5 skills, cover the full dev cycle:

e.g.
prompt-builder → plan-review → tdd → systematic-debugging → session-verify

Fork it, add your own, keep everything versioned - one library for all projects.

[mare @Maresnauseam]

Every installed skill costs tokens on every API call.

27 skills with 3-4 line headers = ~100 lines of permanent system prompt you carry around.

Research shows context bloat alone kills 14-85% of model performance (EMNLP 2025).

So I designed for less.

[mare @Maresnauseam]

A small, practical hub to store, organize, and reuse AI coding skills across all projects.

27 skills. 5 agents. 4 rules. You install only what you need.

Mostly based on Python.

[Rutger Bregman @rcbregman]

Sam Altman is such an incredible backstabber, liar and traitor.

While your competitor is taking a heroic and principled stand, you swoop in to make your deal. Imagine working for this guy - is there a greater shame?

This should lead to a mass exodus from OpenAI.

[mare @Maresnauseam]

@ThatHanselGuy @joni_vrbt @OxFemyn_DeFi Hey, I’m interested in hosting local at our company too.

At what scale and what hardware are y‘all hosting?

[Roy E. Bahat @roybahat]

Outside Anthropic's office in SF... intense moment!