Sunday, 3 May 2026
This dated roundup collects the most interesting AI and technology developments found for Sunday, 3 May 2026.
Tags: hardware, reasoning, opinion
Best Local Vision-Language Models?
What are in your opinion the best local vision models to get a good despription of picture for a 16 GB GPU? At the moment I use qwen3 vl 8b thinking q8 but I wonder, if there is a better model around? Often the models is not really to recognize the right kind of clothes and background.
Read moreResearch & Products
@WesRoth: Meta Launches Ads MCP — Claude and ChatGPT Can Now Manage Meta Ad Accounts
Meta launched its Ads MCP and CLI, creating a direct bridge for frontier AI models like Claude and ChatGPT to access and interact with the Meta Ads ecosystem through natural language.
Read moreSide-by-side comparison of Qwen-Image, ERNIE Base/Turbo, and FLUX.2 Dev across 8 custom styles (single RTX 5090)
Hey folks. I've been playing around at home picking which open-source image model to settle on for some prototyping work, and ended up doing a fun little side-by-side that maybe someone else will find useful. Same prompt and same seed across four models, with eight different style presets (AI generated). Completely amateur — no benchmarking rigor, just curiosity and a free weekend.
Read moreWhat if ChatGPT launched in 1998
Saw the Wikipedia premium screenshot and it got me thinking… Enjoy ✌️
Read moreTags: hardware, crypto_defi
xAI Token Processing Collapsed 90% — From 6 Trillion to 0.6 Trillion Per Week
Tomasz Tunguz reported that xAI's token processing volume collapsed from 6 trillion tokens per week in late 2025 to 0.6 trillion by April 2026 — a 90% decline that explains the 11% GPU utilization.
Read moreTags: funding, crypto_defi
How have you used Claude skills and integrations to achieve meaningful productivity increases at a dev org level?
I have been tasked with doing just that.
Read moreTags: ai_agents, reasoning
Tried running Claude Code with local LLMs via Ollama — ended up subscribing to Pro anyway. But now I can't disconnect from the local server.
I've been experimenting with using Ollama to run Claude Code locally with models like Gemma 4, thinking I could avoid API costs. However, I quickly realised these models aren't really optimised for Claude Code's agentic workflows — they tend to get stuck in thinking loops and don't follow Claude Code's expected output structure well. So I ended up subscribing to Claude Pro anyway.
Read more