# Optimizing Instructions and Demonstrations for Multi-Stage Language Model Programs

- **Optimizing prompts** for Language Model Programs involves **crafting effective instructions** and demonstrations to **maximize downstream metrics** without direct access to module-level data.

# Getting 50% (SoTA) on Arc-AGI with GPT-4o

- Achieving a **50% accuracy** on the ARC-AGI test set, Redwood Research's method involves generating approximately **8,000 Python programs per problem** with GPT-4o, then selecting the most accurate implementations based on their performance on example inputs.

# Sharing new research, models, and datasets from Meta FAIR

- **Meta FAIR has released six new research artifacts** focusing on innovation, creativity, efficiency, and responsibility, including models for image-to-text and text-to-music generation, and a technique for detecting AI-generated speech.

# 
Accessing GPT-4 level Mathematical Olympiad Solutions via Monte Carlo Tree Self-refine with LLaMa-3 8B

- The **MCT Self-Refine (MCTSr) algorithm** innovatively combines **Large Language Models (LLMs) with Monte Carlo Tree Search (MCTS)** to tackle complex mathematical reasoning, enhancing LLMs' accuracy and reliability in strategic tasks.

# Every Way to Get Structured Output from LLMs

- **Frameworks like BAML, Instructor, and TypeChat offer various approaches to extracting structured JSON from LLMs**, with techniques ranging from parsing malformed JSON to constraining LLM token generation.

# 
Creativity Has Left the Chat: The Price of Debiasing Language Models

- **Debiasing LLMs through RLHF** significantly **reduces their creativity**, manifesting as lower entropy in token predictions and a tendency towards predictable outputs.

# YaFSDP: a sharded data parallelism framework, faster for pre-training LLMs

- **YaFSDP** is a **Sharded Data Parallelism framework** optimized for transformer-like neural networks, offering detailed insights through [Medium](https://medium.com/yandex/yafsdp-a-tool-for-faster-llm-training-and-optimized-gpu-utilization-is-no-632b7539f5b3) and [Habr](https://habr.com/ru/companies/yandex/articles/817509/) blog posts.

# Optical PCIe 7.0 connection hits 128 GT/s

- Cadence demonstrated a **working optical connectivity solution for PCIe 7.0**, achieving 128 GT/s over standard optical connectors, showcasing a significant technical advancement even before the PCIe 7.0 specification is finalized.

# Google DeepMind shifts from research lab to AI product factory

- **Google DeepMind** transitions from a **research-oriented entity** to a **production-centric AI product developer**, marking a significant shift in its operational focus.

# Generating audio for video

- **Google DeepMind's video-to-audio (V2A) technology** generates rich soundtracks for videos by combining video pixels with natural language text prompts, enhancing the realism of generated or silent films.

# 
AlphaMath Almost Zero: process Supervision without process

- **AlphaMath** introduces a **novel approach** that **eliminates the need for human or GPT-annotated process supervision** in enhancing large language models' (LLMs) mathematical reasoning, by leveraging the Monte Carlo Tree Search (MCTS) framework.

# 3D Gaussian Splatting as Markov Chain Monte Carlo

- The **3D Gaussian Splatting** method for neural rendering is enhanced by treating the set of 3D Gaussians as **Markov Chain Monte Carlo (MCMC) samples**, leading to higher quality scene reconstructions without the need for precise initial placements.

# LLM that can call multiple tool APIs with one request

- **Cohere introduces Command R+**, a model designed to automate complex business workflows by leveraging external tools, enhancing operations across various enterprise systems.

# Enhancing Code Completion for Rust in Cody

- **Early efforts to enhance code completion for Rust in Cody** have shown promising results, particularly in addressing the performance gap observed in languages not well-represented in training datasets.

# Will We Run Out of Data? Limits of LLM Scaling Based on Human-Generated Data

- The **total effective stock of human-generated public text data is estimated at 300 trillion tokens**, with language models expected to fully utilize this stock between **2026 and 2032**, depending on training intensity.