New meta: using Agents for high quality Synthetic Data

Week 29 of Coding with Intelligence

📰 News

• LLaVA-NeXT-Interleave

  Text image text image text image text …

• OpenAI working on new reasoning technology under code name ‘Strawberry’ (Q*)

• Claude 3.5 Sonnet doubles max output tokens to 8192

  More tokens is more better :)

• Groq release of Tool Use Llama 3 finetuned models was well received

  Thanks y'all for the positive comments! 🫶

• Flash Attention v3

• Codestral Mamba by Mistral

📦 Repos

• RouteLLM: a framework for serving and evaluating LLM routers

• MambaVision: A Hybrid Mamba-Transformer Vision Backbone

📄 Papers

• AgentInstruct: Toward Generative Teaching with Agentic Flows
  This paper is what the title of the newsletter is referring to, AgentInstruct isn’t the only work describing this fundamental approach of using agents to generate synthetic data to then go train on with backprop, but I think it’s a good paper to get started on the topic.

• Associative Recurrent Memory Transformer

• Cradle: Empowering Foundation Agents Towards General Computer Control

• PaliGemma technical report

  by the Google DeepMind team.

• Teaching Transformers Causal Reasoning through Axiomatic Training

• Generative Space-Time Enhancement for Video Generation

  Demos look very impressive!

• MiniCache: KV Cache Compression in Depth Dimension for Large Language Models

• Empirical Analysis of Layer Dynamics in Pretrained Transformer Models

• Automatic Instruction Evolving for Large Language Models

• Prover-Verifier Games improve legibility of language model outputs

• The N+ Implementation Details of RLHF with PPO: A Case Study on TL;DR Summarization

  RLHF is finicky, don’t dive in without grokking this paper.

• SequenceMatch: Imitation Learning for Autoregressive Sequence Modelling with Backtracking

📱 Demos

• Groq Tool Use Demo on Hugging Face 🤗 Spaces

  In case you haven't played with the tool use models yet! Very nice deploy experience on HF Spaces, you should try it with Gradio.

📚 Resources

• What happened to BERT & T5? On Transformer Encoders, PrefixLM and Denoising Objectives

• Can ChatGPT do data science?

  I especially like their recommendation at the end of making it easier to include the right context for the LLM. I think many ChatGPT-esque products still make that hard today. Cursor comes to mind as a product that has done this quite well with @-tagging, but it is limited to the coding domain.

• Run CUDA, unmodified, on AMD GPUs

  By UK based HPC consultant

• Meet the AI Agent Engineer

  Interesting take. Still not a full solutions to the agent reliability problem, however.

• Prompting for Llama 3 by Meta

---

Want more? Follow me on X! @ricklamers