Edge AI makes waves: Qwen 2.5 Code Interpreter in your browser

Week 42 of Coding with Intelligence

Today’s demo is the most awesome Edge AI application I've seen, check it out. Python generated & executed fully in your browser.

📱 Demos

• Qwen-2.5-Coder 1.5B with access to an in-browser code interpreter

📰 News

• Mistral releases 2 Ministral edge focused models

  They release a 3B and 8B for mainly embedded use cases. It seems to compare favorably against Llama 3.2 3B on various benchmarks and against Llama 3.1 8B respectively.

📦 Repos

• OpenR: An Open Source Framework for Advanced Reasoning with Large Language Models

• Swarm by OpenAI

  A multi-agent orchestration framework.

• Answer.ai releases fastdata: a synthetic data generation library

  The repo: https://github.com/AnswerDotAI/fastdata

• Llama 3.1 Nemotron: finetune by NVIDIA

  The mention a slew of high benchmark scores: "This model reaches Arena Hard of 85.0, AlpacaEval 2 LC of 57.6 and GPT-4-Turbo MT-Bench of 8.98, which are known to be predictive of LMSys Chatbot Arena Elo" not sure how well that translates to real-word performance. Run your own evals 🙌 More interestingly, they detail the alignment procedure they used to create the model.

• Meta Lingua: a hackable training and inference library

  Looks very easy to modify and play with for trying out ideas.

• Janus-1.3B, a multimodal understand + generation model by DeepSeek

  "Janus is a novel autoregressive framework that unifies multimodal understanding and generation"

📄 Papers

• Meta releases Self-Taught Evaluators: iterative improvement through synthetic data

  Code is available on GitHub https://github.com/facebookresearch/RAM/tree/main/projects/self_taught_evaluator and weights on Hugging Face https://huggingface.co/facebook/Self-taught-evaluator-llama3.1-70B

• Fine-Tuning with Divergent Chains of Thought Boosts Reasoning Through Self-Correction in Language Models

  Divergent CoT (DCoT) in short: "further improving performance by requiring models to compare multiple reasoning chains before generating a solution in a single inference step". Interesting idea!

• What's the Magic Word? A Control Theory of LLM Prompting

  I like this exploration of a more principled approach to prompting.

• Thinking LLMs: General Instruction Following with Thought Generation

  "We propose a training method for equipping existing LLMs with such thinking abilities for general instruction following without use of additional human data. We achieve this by an iterative search and optimization procedure that explores the space of possible thought generations, allowing the model to learn how to think without direct supervision."

• Spirit LM: Interleaved Spoken and Written Language Model

  By Meta, a foundation multimodal language model that freely mixes text and speech.

• MEXMA: Token-level objectives improve sentence representations

  Improved multilingual embeddings by Meta.

• Artificial Kuramoto Oscillatory Neurons

  By Max Welling et al. "It replaces threshold units with generalized Kuramoto oscillators. It dynamically binds neurons, generates waves of activations, is adversarially robust, calibrated for uncertainty, and can reason. If you ask me: this is the next big thing!"

📚 Resources

• François Chollet keynote talk about ARC

• Chris Manning - Meaning and Intelligence in Language Models (COLM 2024)

  Christopher Manning is the director of the SAIL institute at Stanford.

---

Want more? Follow me on X! @ricklamers