What I found interesting this time

by: Artur Dziedziczak

May 11, 2025

“Trying out Llama.cpp’s New Vision Support,” n.d. https://simonwillison.net/2025/May/10/llama-cpp-vision/#atom-everything

Llama now accepts visual prompts. This will be a huge deal for IoT and house automation.
#vllm

“Updated Rate Limits for Unauthenticated Requests,” n.d. https://github.blog/changelog/2025-05-08-updated-rate-limits-for-unauthenticated-requests/

Microsoft made pathetic move to secure their monopoly in code version control market. Disgusting.

“REVEAL: Multi-Turn Evaluation of Image-Input Harms for Vision LLMs,” n.d. https://arxiv.org/pdf/2505.04673

“Carousels with CSS,” n.d. https://developer.chrome.com/blog/carousels-with-css

Apparently soon you will be able to ship browser native carousels in #html and #css.

“Saying ‘Hi’ to Microsoft’s Phi-4-Reasoning,” n.d. https://simonwillison.net/2025/May/6/phi-4-reasoning/#atom-everything

Great article that shows the system prompt of ’reasoning’ used in the Phi-4 model.
I still don’t fully understand how this reasoning works. However, from what I see, the LLM asks itself the same question multiple times and considers the result that is the most probable.
Sadly, this again causes it to not come up with novel solutions. Rather, it still operates within the known space / it’s limited by token probablility."

“Feed a Video to a Vision LLM as a Sequence of JPEG Frames on the CLI (Also LLM 0.25),” n.d. https://simonwillison.net/2025/May/5/llm-video-frames/#atom-everything

A plugin for OpenAI GPT-4.1 that allows users to turn video into summarized text frame by frame sounds interesting, but when I checked the output of the example, it all falls apart.
- The dog’s color is wrong.
- The dog doesn’t sniff the cupcake.
- It does not lick human fingers.
The output is highly hallucinated, not only because frames taken out of context do not represent what is really happening, but also because the AI tends to hallucinate.

“AI Generated Vulnerability Reports Sucks,” n.d. https://hackerone.com/reports/3125832

Apparently Curl got a lot of #vulnerabilities patches generated #by LLMs. Maintainers are really angry about it as most of those do not even apply as valid patch…