Interesting new benchmark from Hugging Face testing how well vision LLMs can handle long video inputs (generally after they've been split into many thousands of images) – my notes here: https://
simonwillison.net/2025/Jul/23/ti
mescope/
…
LLMS
-
Hugging Face Benchmark: Vision LLMs Long Video Input Performance
By
–
-
Scientific Papers Enhanced With LLM-Assisted Demos for Accessibility
By
–
Aside from everything else interesting about this paper, I appreciate that more scientific papers (aided by LLM help?) are now including little demos and experiments to help non-specialists get the points they are making. (And no, you cannot identify the hidden signals)
-
API and Chatbot Risks: Accessibility Versus Safety Concerns
By
–
All these risks are unfortunately in my opinion much more prevalent with APIs and chatbots (non-open weights) because they are way easier to use by a larger number of people (ex chatgpt has hundreds of millions of users with very little limitations of what you can do with it and
-
RoPE: Rotary Positional Embeddings in Transformer Models
By
–
📢Inside RoPE: Rotary Magic into Position Embeddings
— Satya Mallick (@LearnOpenCV) 23 juillet 2025
This week, we take a comprehensive look at Rotary Positional Embeddings (RoPE), an advanced technique used in Transformer-based models to enhance long-context understanding. RoPE addresses the limitations of traditional… pic.twitter.com/HIZH4LwKePInside RoPE: Rotary Magic into Position Embeddings This week, we take a comprehensive look at Rotary Positional Embeddings (RoPE), an advanced technique used in Transformer-based models to enhance long-context understanding. RoPE addresses the limitations of traditional
-
Subconscious Threads Enable Long-Horizon Reasoning Beyond Context
By
–
Beyond Context Limits Subconscious Threads for Long-Horizon Reasoning
-
Subliminal Learning: Hidden Signals in Language Models
By
–
Subliminal Learning: Language models transmit behavioral traits via hidden signals in data Cloud et al.: https://
arxiv.org/abs/2507.14805 #ArtificialIntelligence #DeepLearning #MachineLearning -
RAGFlow: Open-Source RAG Engine for Deep Document Understanding
By
–
RAG engine for deep document understanding! RAGFlow is an open-source RAG engine for deep document understanding and streamlined knowledge workflows from complex data formats. 100% Open Source
-
LLMs Progress: Six Months in Language Model Development
By
–
I have a tag! https://
simonwillison.net/tags/pelican-r
iding-a-bicycle/
… I also pulled a while bunch of them together in this talk a couple of months ago https://
simonwillison.net/2025/Jun/6/six
-months-in-llms/
… -
ICML Statement on Hidden Subversive LLM Prompts
By
–
ICML’s Statement about subversive hidden LLM prompts We live in a weird timeline…
-
New Research: LSM-2 Foundation Model Learns from Noise
By
–
The big idea: – Foundation models shouldn’t fear noise.
– They should learn from it. LSM-2 is a step towards that future where AI doesn’t just tolerate real-world messiness… It thrives in it. Read the complete research here: