One Useful Thing
·
3 months ago
● 2 sources
OpenAI released GPT-5.5, a new model family with improved coding and reasoning capabilities compared to previous versions. GPT-5.5 Pro completed a 3D harbor town simulation task in 20 minutes compared to 33 minutes for GPT-5.4 Pro, and demonstrated the ability to generate academic papers from raw data and create illustrated tabletop roleplaying games. The model advances across three areas—models, applications like desktop Codex, and tool integrations—enable more complex tasks, though limitations remain in long-form fiction generation.
Deep Learning Weekly
·
3 months ago
This Deep Learning Weekly issue covers releases of ChatGPT Images 2.0, Claude Opus 4.7, and Google's Deep Research agents, alongside MLOps tools like Comet's Ollie coding assistant for agent debugging. Research highlights include OpenThoughts' public reasoning datasets achieving 53% on AIME 2025 and Google's Simula framework for synthetic data generation via mechanism design. The updates reflect progress in multimodal AI, agent tooling, and open-source reasoning model training, with infrastructure innovations like Notion's vector search evolution and new memory management systems for agents.
OpenAI Blog
·
3 months ago
● 2 sources
OpenAI released GPT-5.5, a new model designed for complex tasks including coding, research, and data analysis. The company positioned it as faster and more capable than previous versions, with integration across multiple tools. Users working on technical projects now have access to improved performance for computationally demanding workflows.
OpenAI Blog
·
3 months ago
● 3 sources
ChatGPT Work enables users to automate routine tasks and generate deliverables by connecting across multiple tools, files, and workflows. The system integrates with external applications to process inputs and produce outputs without manual intervention across platforms. This reduces time spent on repetitive work and allows teams to focus on higher-level activities.
Allen Institute (AI2)
·
3 months ago
Researchers found that four architectural choices in language models—QK normalization, grouped-query attention, sliding window attention, and shorter pretraining context length—individually have modest negative effects on long context performance but combine to reduce benchmark scores by up to 47%. The study used OlmPool, a suite of 26 7-billion-parameter models trained on 140 billion tokens with identical data but different architectures, to isolate these effects. The findings show that Llama 3's long context capabilities come primarily from architecture rather than training data, meaning context extension recipes validated on Llama may not transfer to other model families without adjustment.
Allen Institute (AI2)
·
3 months ago
OlmoEarth Studio now allows users to compute and export embedding vectors from their open-source Earth observation foundation models for downstream analysis tasks. The platform supports three encoder variants ranging from Nano (1.4M parameters) to Base (89M parameters), with exports as Cloud-Optimized GeoTIFFs that can be customized by area, time range, resolution, and imagery source. Users can apply these embeddings to similarity search, few-shot segmentation, change detection, and unsupervised exploration with minimal labeled data, such as producing land-cover maps from just 60 labeled pixels.
Hugging Face Blog
·
3 months ago
Transformers.js was integrated into a Chrome extension using Manifest V3 by centralizing model inference and agent orchestration in a background service worker while keeping UI and page logic in separate side panel and content script runtimes. The extension loads two models—Gemma-4-E2B for text generation and MiniLM-L6-v2 for embeddings—entirely in the background, with all inference cached under the extension origin rather than per-website. The architecture separates conversation state (background memory), user settings (chrome.storage.local), and retrieval data (IndexedDB), enabling responsive UI, reusable models across tabs, and clear messaging contracts between runtimes.
