One Useful Thing
·
3 months ago
Anthropic released Claude Dispatch, a system that lets users control AI agents on their desktop via phone messaging, addressing how poor chatbot interfaces create cognitive load that undermines AI productivity gains. The tool combines Claude Cowork (which accesses local files and applications) with Dispatch messaging capability, allowing users to control AI agents through WhatsApp-like interfaces rather than traditional chat windows. As specialized interfaces improve—from coding agents to custom-generated visualizations—more workers will be able to access AI's actual capabilities, shifting from adapting to AI interfaces to AI adapting interfaces for different tasks and users.
Google Research
·
3 months ago
Researchers investigated the optimal trade-off between the number of items rated and the number of raters per item in AI benchmarking, using simulations across subjective datasets like toxicity detection and chatbot safety evaluation. The standard practice of using 1 to 5 raters per item often proves insufficient, with more than 10 raters per item typically needed to capture human disagreement and achieve reproducible results. The findings show that practitioners can achieve highly reproducible benchmarks with approximately 1,000 total annotations by optimizing the ratings-per-item ratio based on their chosen metric, rather than pursuing the traditional single-truth paradigm.
Sequoia
·
3 months ago
Block is redesigning its organization around AI-powered coordination instead of traditional hierarchy, replacing the information-routing function that managers have provided for 2,000 years by building company and customer world models from transaction data and machine-readable work artifacts. The system composes financial capabilities (payments, lending, card issuance) into personalized solutions for customers at specific moments, such as offering short-term loans to merchants before cash flow problems occur. This shift moves the organization from predetermined product roadmaps managed by humans to AI-driven proactive delivery of capabilities, fundamentally changing how large organizations coordinate work.
Hugging Face Blog
·
3 months ago
IBM released Granite 4.0 3B Vision, a compact vision-language model designed to extract information from enterprise documents, tables, charts, and forms. The model achieved 86.4% accuracy on chart-to-summary tasks and 92.1% on cropped table extraction benchmarks, outperforming larger competitors. It is available now on HuggingFace under the Apache 2.0 license and can operate standalone or integrated with Docling for multi-page document processing.
OpenAI Blog
·
3 months ago
OpenAI secured $122 billion in new funding to expand its AI capabilities and infrastructure globally. The company will use the capital to build next-generation computing infrastructure and support increased demand for its products including ChatGPT and enterprise AI services. OpenAI can now accelerate its development of frontier AI models and scale its operations across more markets.
Hugging Face Blog
·
3 months ago
OpenMed built an end-to-end protein engineering pipeline combining structure prediction, sequence design, and codon optimization across 25 species for $165 in compute costs. CodonRoBERTa-large-v2 achieved a perplexity of 4.10 and CAI correlation of 0.40, outperforming ModernBERT by 6x, with four production models trained in 55 GPU-hours. The system enables researchers to move from a protein concept to synthesis-ready DNA sequences by applying transformer models trained specifically on biological codon statistics rather than natural language patterns.
AI Act
·
3 months ago
The EU AI Act's enforcement provisions for general-purpose AI model providers come into two stages: substantive and procedural obligations took effect on 2 August 2025, while the Commission's supervision and enforcement powers activate on 2 August 2026. The Commission can impose fines up to 3% of annual worldwide turnover or EUR 15 million (whichever is higher) for violations including failure to provide documentation, non-compliance with requested measures, or denying access to models for evaluation. Multiple actors beyond the Commission—including national market surveillance authorities, downstream providers, and scientific panels—can trigger enforcement actions by requesting Commission intervention or lodging complaints against GPAI model providers.
Together AI
·
3 months ago
Aurora is an open-source reinforcement learning framework that continuously adapts speculative decoding draft models by learning directly from live inference traces without interrupting service. The system achieves a 1.25x additional speedup over static speculators on models like Qwen3 and Llama3, and demonstrates that online training from scratch can outperform carefully pretrained offline baselines. This approach shifts speculative decoding from a static one-time setup to a dynamic self-improving system that reduces infrastructure costs and adapts to changing traffic patterns.
Hugging Face Blog
·
3 months ago
TRL released v1.0, a post-training library that now implements over 75 methods and has shifted from research code to stable infrastructure powering production systems. The library is downloaded 3 million times per month and maintains a dual-contract stability model where stable components follow semantic versioning while experimental methods can evolve rapidly. The design deliberately avoids over-abstraction and accepts code duplication to remain adaptable as post-training methods continue to shift, with stable trainers for SFT, DPO, reward modeling, RLOO, and GRPO now in the core.
