Hugging Face Blog
·
2 months ago
PipelineRL's RL training system experienced discrepancies between vLLM V0 and V1 inference engines that affected rollout logprobs used in policy gradient computations, requiring systematic debugging before objective changes. The team fixed four specific issues: logprobs semantics (enabling processed_logprobs mode), runtime defaults (disabling prefix caching and async scheduling), inflight weight-update synchronization (using pause/resume with cache preservation), and fp32 precision for the language-model head projection. The final vLLM V1 configuration now produces training metrics matching the V0 reference, enabling future objective-level improvements on a correct backend foundation.
Amazon Science
·
2 months ago
Amazon developed computational tools combining optimization and machine learning to design its middle-mile logistics network that performs reliably under uncertainty rather than just in perfect conditions. The company created a graph attention network model that represents the network in two interconnected layers to capture spatial demand patterns and interdependencies between origin-destination pairs, enabling it to stress-test designs against hundreds of plausible scenarios. This approach allows Amazon to distinguish between network designs that appear efficient on average but are fragile under disruption and those that maintain stable performance during demand spikes, weather events, or facility outages.
Google DeepMind
·
2 months ago
Google's AlphaEvolve, a Gemini-powered coding agent that designs and optimizes algorithms, has expanded beyond mathematics and computer science into genomics, power grid optimization, quantum physics, and commercial applications across multiple industries. The system achieved a 30% reduction in DNA sequencing errors, increased electricity grid optimization feasibility from 14% to 88%, and helped Klarna double its model training speed while improving quality. AlphaEvolve is now available through Google Cloud as a commercial tool, shifting algorithm optimization from months-long manual efforts to automated discovery across scientific research and enterprise software.
OpenAI Blog
·
2 months ago
● 3 sources
ChatGPT uses privacy-preserving techniques during training to limit personal data exposure while allowing the model to learn from broad information sources. Users can opt out of having their conversations used to train future models through account settings, a choice OpenAI began offering in 2023. This approach lets individuals control their data participation while enabling continued model improvement from other available sources.
Hugging Face Blog
·
2 months ago
The Open ASR Leaderboard added private datasets from Appen Inc. and DataoceanAI covering scripted and conversational speech across multiple accents to prevent models from optimizing specifically for public benchmarks. The private datasets comprise 29.7 hours of audio across Australian, Canadian, Indian, American, and British English with gender-balanced speakers. The leaderboard now offers optional toggling between public and private datasets for evaluation, with the default ranking computed only on public data to preserve benchmark integrity.
OpenAI Blog
·
2 months ago
Uber has integrated OpenAI's technology to power AI assistants and voice features for both drivers and passengers. The system operates across Uber's global platform, which processes millions of rides daily in real time. Drivers can receive earnings optimization suggestions while riders gain faster booking capabilities through conversational AI.
OpenAI Blog
·
2 months ago
OpenAI announced a cohort of 26 students selected for its ChatGPT Futures Class of 2026 program. The program focuses on students using AI tools to build projects, conduct research, and create real-world applications. Participants gain access to OpenAI's resources to develop AI-driven solutions in their respective fields.
OpenAI Blog
·
2 months ago
Singular Bank created an internal AI assistant called Singularity using ChatGPT and Codex to automate routine banking tasks. The tool saves bankers between 60 and 90 minutes per day on meeting preparation, portfolio analysis, and follow-up work. Bankers can redirect this time toward client relationships and more complex financial decisions.
OpenAI Blog
·
2 months ago
OpenAI analyzed how leading enterprises are adopting AI at scale through their B2B Signals research, focusing on Codex-powered agentic workflows and competitive positioning. The research examines frontier firms' strategies for deepening AI integration across operations, though no specific metrics or adoption rates are disclosed in the available summary. Companies implementing these workflows gain measurable efficiency gains that reinforce their competitive position relative to peers not yet at scale.
