Daily Brief

Ollama, an open-source AI tool, has raised $65M in Series B funding, reaching nearly 9M users. It enables developers to run open-weight AI models on PCs and offers cloud-hosted models via subscription, tracking usage by GPU time.

ProMoE-FL introduces a Prototype-conditioned Mixture-of-Experts framework for multimodal federated learning, effectively addressing missing modalities. It outperforms existing methods on four chest X-ray datasets, demonstrating superior feature synthesis capabilities in both homogeneous and heterogeneous settings.

Meta's Muse Spark 1.1 API, priced at $1.25 per million input tokens, undercuts competitors like OpenAI and Anthropic, intensifying the AI price war. The model excels in multi-agent orchestration and coding tasks, leading benchmarks such as MCP Atlas and Humanity's Last Exam, while also promising significant cost efficiency for developers.

SpaceXAI has launched Grok 4.5, a new coding-focused model that is 3x larger than Grok 4.3, priced at $2 per million input tokens. Positioned as an Opus-class model, it aims for efficiency and speed, outperforming competitors like GPT-5.6 and Opus 4.8 in cost-effectiveness.

Google's AlloyDB AI functions now allow direct LLM calls within SQL, achieving 2,400x throughput improvements via smart batching and 23,000x with optimized proxy models. The proxy model architecture enables local inference, significantly reducing costs and latency for database queries.

The rise of Token billing in AI has transformed costs into operational expenses, with prices varying significantly due to factors like model efficiency, energy costs, and contract terms. As companies shift from GPU hours to Token-based billing, understanding the hidden complexities behind Token pricing becomes crucial for effective budgeting.

This study introduces a gradient-based alignment method applicable to any differentiable ASR model, including speech LLMs, without requiring training or model modifications. Evaluated across sixteen models, it shows promising alignment performance, particularly in scenarios where native alignments are weak, although it incurs a cost of one backward pass per token.

This study introduces a framework for predicting item parameters from text embeddings, achieving a predictive R squared of 0.53 for item difficulty in mathematics, while highlighting the limitations in predicting discrimination and pseudo guessing parameters. The findings emphasize the importance of repeated cross-validation to avoid inflated accuracy in calibration applications.