Anthropic doubles development speed as startups target edge AI

San Francisco artificial intelligence laboratory Anthropic is developing its artificial intelligence systems around twice as fast as a year ago, according to an independent assessment reported by NBC Bay Area. At the same time, startup builders are pushing synthetic physics simulation and local device execution forward, with Physicl raising $10 million from Inovia Capital for robot training environments and Underdog releasing a 32MB speech generation model with 31 million parameters.

METR Evaluates 11 Weeks of Anthropic Internal Data

Anthropic invited researchers from the Berkeley, California-based nonprofit Model Evaluation and Threat Research (METR) to review its internal data and interview staff over a period of about 11 weeks, from mid-August through late October, as the San Francisco company prepares to go public later this month. METR estimated that Anthropic is improving the capabilities of its models to perform well-specified tasks around twice as fast as a year earlier, attributing much of the faster development to the assistance of AI itself, including Anthropic's Claude chatbot helping develop newer versions of itself.

During the assessment period, Anthropic CEO Dario Amodei published a Sept. 12 essay urging the industry to slow the pace at which capabilities improve so safety measures can catch up, pointing to recursive self-improvement. METR also reviewed both a public risk report Anthropic published in August and a private version, stating that the public report overall understates the level of acceleration that was occurring. METR researchers ruled out computing power growth as the cause of the speedup and noted that no bottlenecks in compute or data were holding back progress. Anthropic co-founder and responsible scaling officer Sam McCandlish wrote on X that many of METR's conclusions matched his own, noting that AI is speeding up research while explaining that redactions in the report protected commercially sensitive details.

Physicl Raises $10 Million for High-Fidelity Robot Simulations

While frontier labs accelerate software capabilities, founder and CEO Alex de Vigan — who previously spent a decade building 3D e-commerce rendering company nfinite — has raised $10 million in seed funding from Inovia Capital for his new startup, Physicl, as reported by TechCrunch. Physicl transforms unstructured real-world information into high-fidelity 3D physical representations with accurate geometry, materials, and physics to help world model designers and robot builders post-train models on difficult edge cases.

Physicl is releasing a product called Studio featuring 10,000 3D environments and 100,000 assets representing diverse objects, enabling users to generate synthetic data for tasks such as discerning a cup on a cluttered countertop or in shadow. World model builder SpAItial, led by CEO and founder Matthias Niessner — also a professor of robotics at the Technical University of Munich — collaborated with Physicl as an early customer to train its Echo-3 tool, which generates 3D simulations from 2D imagery.

Underdog Demos 32MB Woof Speech Model for Offline Devices

On the local device frontier, startup Underdog, founded by Sigil Wen, released a browser-based retro room game demo featuring a typewriter, a CRT monitor, a desk lamp, and a music tuner powered by a 32MB voice generation model called Woof, according to TechCrunch. Woof has 31 million parameters, is open for download, and can run on most devices without an internet connection.

The release follows Underdog's launch weeks earlier of Bark, a 167MB speech-to-text model designed for edge transcription and dictation. Underdog has also released a 27-billion-parameter model capable of running on laptops alongside a waitlisted locally hosted application positioned as an alternative to cloud notetakers and assistants such as Granola, Muse, and Instinct.