SupraLabs Experimentally Releases ~30M-Parameter Any2Any Multimodal Model

SupraLabs has released an experimental multimodal Transformer called "Supra-A2A-Nano-Exp". Its most notable feature is that it treats text, images, and video all as a single, unified token stream.

Conventional multimodal models typically combine multiple components such as vision encoders, diffusion models, and cross-attention, but this model eliminates all of them and achieves multimodal processing through pure "next-token prediction".

The model has roughly 30 million parameters and a context length of 384 tokens. Images are converted from 64×64 pixels into 64 tokens via VQ-VAE and processed in the same vocabulary space as text. While it isn't practical at this scale, it's a fascinating approach that demonstrates the concept that "multimodality can be expressed through language modeling alone".

Nous Research Releases Self-Improving AI Agent "Hermes Agent"

Nous Research has released "Hermes Agent" on GitHub. It's an AI agent framework with self-improving capabilities, designed to autonomously enhance its performance through task execution.

Detailed architecture and training methods are available in the GitHub repository. Self-improving agents aim to learn from their own experience and produce better results on subsequent runs — an approach that represents one of the key directions in AI agent research.

While practicality and stability remain to be seen in future evaluations, Nous Research is a team with a proven track record in the open-source LLM community, making this release worth watching.

AllenAI Releases 3D Motion Prediction Models "MolmoMotion"

AllenAI has released two models from the MolmoMotion family (4B-H3-F30 and 4B-H1-F32) on Hugging Face. These are 4B-parameter vision-language models that predict point trajectories in 3D space based on natural language instructions.

Specifically, given a short history of RGB observations, user-specified 2D query points, and an action description, the models predict how a target object will move in the future, expressed as 3D coordinates. The F30 variant was trained with a 3-frame history and the F32 variant with a 1-frame history, accommodating different observation conditions.

They are expected to find use across a wide range of applications that require predicting the future positions of objects, including robotics, autonomous driving, and sports analytics.

Comprehensive Tutorial on Training LLMs from Scratch Gains Traction

A tutorial called "Train LLM from Scratch", which explains how to train large language models from the ground up, was featured on Hacker News. It comprehensively covers everything from tokenization and data preprocessing to architecture design and the training process, making it a valuable resource for developers who want a hands-on understanding of how LLMs work.

Questioning Transparency in AI Use — "Don't Say You Use AI for Writing"

A post titled "Don't Say You Use AI for Writing" was published on the blog of Satisfice, a company specializing in testing. It sounds a warning about the risks of not disclosing when content has been created with AI.

The post points out that presenting AI-generated content as if it were written by a human can undermine reader trust. It argues that, especially in domains where expertise is expected, being upfront about AI use leads to long-term trust building. As AI tools become pervasive, how to handle transparency around their use remains an ongoing topic of debate.

---