読書

Microsoft Research Asia open-sourced Agent Lightning v1.0 on October 7, a lightweight framework of approximately 3,500 lines of code designed to allow the same agent harness used in deployment to participate directly in reinforcement learning.

Traditional agentic RL often requires reimplementing agent logic within the training framework, leading to discrepancies between training and production environments and high costs. Agent Lightning inserts an LLM proxy between the agent and the model, allowing existing harness code to remain unchanged while connecting to RL training.

The framework natively supports running agents as Kubernetes jobs, eliminating dependency on paid commercial sandbox services. In coding agent experiments, using about 6,000 training samples, it improved Qwen3.5-9B's Pass@1 on SWE-bench Verified from 41.8% to 56.4%, an absolute gain of 14.6 percentage points. These results are self-reported by Microsoft and await independent reproduction.

Microsoft has opened direct preorders for its Surface RTX Spark Dev Box, which is slated to ship in November for $5,999.

Built on Nvidia’s Arm-based RTX Spark platform with 128GB of unified memory and Tensor cores, the device is optimized for developers to run local AI models exceeding 120 billion parameters and prototype agents.

As part of the Project Zenith lineup, it ships with a developer-optimized Windows 11 Pro setup including VS Code and GitHub Copilot. The price point is higher than the previously launched Nvidia DGX Spark, reflecting ongoing shortages in RAM and other PC components.

Microsoft introduced an SDK that sandboxes AI agent code execution on Windows.

The MXC (Microsoft Execution Containers) SDK was unveiled at the Windows and Surface event, positioned as a security isolation layer that lets agents execute code and call tools locally while minimizing access to sensitive user data. Per The Verge's report, Microsoft described it as a key component of its Windows AI strategy.

Microsoft has not disclosed full technical specifications, developer availability, or a launch timeline.

Healthleap announced it has raised $38 million across two rounds: an $8 million seed co-led by Sequoia Capital and First Round Capital, and a $30 million Series A led by Hummingbird Ventures.

Founded in 2022, the startup uses large language models to analyze unstructured text in electronic health records, such as clinician notes, to identify patients at risk for undiagnosed conditions like malnutrition and delirium. The platform is currently deployed in over 50 hospitals, including Penn Medicine and Cedars-Sinai.

The company did not disclose its valuation. CEO Josiah Meyer stated that revenue has grown more than 10x over the past three years and that all customers have achieved at least 5x hard ROI, though these figures are self-reported and not independently audited. The new capital will support engineering, sales, and expansion into outpatient and home care settings.

Anthropic launched Claude Haiku 5.5 on October 7, targeting high-volume and latency-sensitive workloads.

The model features a 1 million token context window, 128k max output tokens, and enables 'adaptive thinking' by default. Requests that previously used manual budget_tokens for extended thinking now return a 400 error, and responses may begin with thinking blocks, altering parsing logic.

Production environments currently running Claude Haiku 4.5 face potential breakage if switched directly. Developers must consult the official migration guide to adjust request parameters for the new defaults.

a16z announced an investment in Preference Model, a company building reinforcement learning training environments for AI labs. The company this week open-sourced Karotte on GitHub, a framework designed to prevent models from exploiting shortcuts such as reading answer keys, modifying tests, or crashing graders during RL training.

Karotte's defenses include terminating stray processes before grading and rejecting malicious files. a16z states the framework has been hardened through over a million evaluation runs and red-teaming, though this claim comes from the investor's own announcement and has not been independently reproduced.

No investment amount, round stage, or valuation was disclosed.

Penguin Solutions reported fourth-quarter results showing nearly 68% year-over-year revenue growth, beating consensus estimates by more than $45 million.

The company raised its full-year net revenue guidance to $2.43 billion, approximately 500 basis points above consensus forecasts. This adjustment reflects sustained strength in AI memory and computing demand, with the Integrated Memory segment growing 157%.

Additionally, the company appointed Stephen Cumming, a semiconductor industry veteran with 21 years of experience, as the new Chief Financial Officer, aiming to eliminate uncertainty following the previous CFO's departure.

ALBERT is an approach to saving parameters in text models: it reduces the size of word representations and shares parameters across layers, enabling faster learning of sentence meanings with fewer parameters. When you see it achieving high scores on text tasks with fewer parameters, first ask where those saved parameters come from.

The mechanism it provides is this: parameters are saved by reducing the size of word representations and sharing them across layers, while using sentence-order prediction to retain multi-sentence task capabilities. Therefore, when reading about low-parameter text models, ask whether these two aspects still fit your specific task.

If dealing with non-English languages, extremely small datasets, or if the width per layer is too large, do not use it to judge general optimality. In some single-sentence tasks, removing the sentence-order objective has limited impact.

"ALBERT" (2019) | Next review date: 2027-09-20