World Wire

Google has officially released Android Bench , a new leaderboard and evaluation framework designed to measure how Large Language Models (LLMs) perform specifically on Android development tasks. The dataset, methodology, and test harness have been made open-source and are publicly available on GitHub . Benchmark Methodology and Task Design General coding benchmarks often fail to capture the platform-specific dependencies and nuances of mobile development. Android Bench addresses this by curating a task set sourced directly from real-world, public GitHub Android repositories. Evaluated scenarios cover varying difficulty levels, including: Resolving breaking changes across Android releases. Domain-specific tasks, such as networking on Wear OS devices. Migrating code to the latest version of Jetpack Compose (Android’s modern toolkit for building native user interfaces). To ensure a model-agnostic evaluation, the framework prompts an LLM to fix a reported issue and then verifies t...

Liquid AI has released LFM2-24B-A2B , a model optimized for local, low-latency tool dispatch, alongside LocalCowork , an open-source desktop agent application available in their Liquid4All GitHub Cookbook . The release provides a deployable architecture for running enterprise workflows entirely on-device, eliminating API calls and data egress for privacy-sensitive environments. Architecture and Serving Configuration To achieve low-latency execution on consumer hardware, LFM2-24B-A2B utilizes a Sparse Mixture-of-Experts (MoE) architecture. While the model contains 24 billion parameters in total, it only activates approximately 2 billion parameters per token during inference. This structural design allows the model to maintain a broad knowledge base while significantly reducing the computational overhead required for each generation step. Liquid AI stress-tested the model using the following hardware and software stack: Hardware: Apple M4 Max, 36 GB unified memory, 32 GPU cores. S...

In this tutorial, we explore how we use Daft as a high-performance, Python-native data engine to build an end-to-end analytical pipeline. We start by loading a real-world MNIST dataset, then progressively transform it using UDFs, feature engineering, aggregations, joins, and lazy execution. Also, we demonstrate how to seamlessly combine structured data processing, numerical computation, and machine learning. By the end, we are not just manipulating data, we are building a complete model-ready pipeline powered by Daft’s scalable execution engine. Copy Code Copied Use a different Browser !pip -q install daft pyarrow pandas numpy scikit-learn import os os.environ["DO_NOT_TRACK"] = "true" import numpy as np import pandas as pd import daft from daft import col print("Daft version:", getattr(daft, "__version__", "unknown")) URL = "https://github.com/Eventual-Inc/mnist-json/raw/master/mnist_handwritten_test.json.gz...

Integrating Google Workspace APIs—such as Drive, Gmail, Calendar, and Sheets—into applications and data pipelines typically requires writing boilerplate code to handle REST endpoints, pagination, and OAuth 2.0 flows. Google AI team just released a CLI Tool (gws) for Google Workspace. The open-source googleworkspace/cli (invoked via the gws command) provides a unified, dynamic command-line interface to manage these services. Designed for both human developers and AI agents, gws eliminates the need for custom wrapper scripts by providing structured JSON outputs, native Model Context Protocol (MCP) support, and automated authentication workflows. Dynamic API Discovery Architecture Unlike traditional CLI tools that compile a static list of commands, gws builds its command surface dynamically at runtime. When executed, gws uses a two-phase parsing strategy: It reads the first argument to identify the target service (e.g., drive ). It fetches that service’s Google Discovery Docum...

OpenAI has released Symphony , an open-source framework designed to manage autonomous AI coding agents through structured ‘implementation runs.’ The project provides a system for automating software development tasks by connecting issue trackers to LLM-based agents. System Architecture: Elixir and the BEAM Symphony is built using Elixir and the Erlang/BEAM runtime. The choice of stack focuses on fault tolerance and concurrency. Since autonomous agents often perform long-running tasks that may fail or require retries, the BEAM’s supervision trees allow Symphony to manage hundreds of isolated implementation runs simultaneously. The system uses PostgreSQL (via Ecto) for state persistence and is designed to run as a persistent daemon. It operates by polling an issue tracker—currently defaulting to Linear —to identify tasks that are ready for an agent to address. The Implementation Run Lifecycle The core unit of work in Symphony is the implementation run . The lifecycle of a run fol...

In this tutorial, we build an advanced Tree-of-Thoughts (ToT) multi-branch reasoning agent from scratch. Instead of relying on linear chain-of-thought reasoning, we design a system that generates multiple reasoning branches, scores each branch using a heuristic evaluation function, prunes weak candidates, and continues expanding only the strongest paths. We combine an instruction-tuned transformer model with a custom tree structure and implement beam-search style selection with depth-limited search. By grounding the system in the 24-game domain, we create a clear, objective benchmark for reasoning where we can observe branch expansion, pruning, scoring, and goal detection in action. Copy Code Copied Use a different Browser !pip -q install -U transformers accelerate sentencepiece import re import math from dataclasses import dataclass, field from typing import List, Optional, Tuple, Dict, Any import torch from transformers import AutoTokenizer, AutoModelForSeq2SeqLM MOD...