Company@X

Signal of the Day

Anthropic unveiled a major interpretability breakthrough dubbed “J-space,” an internal mathematical workspace within Claude that allows the model to perform silent, multi-step reasoning steps independent of its visible text output. This discovery gives researchers the ability to audit a model’s private thoughts, exposing hidden goals and revealing the model’s situational awareness—such as privately noticing when a test evaluation is artificially staged.

Key Announcements

Anthropic · Link In addition to the fundamental J-space research, Anthropic demonstrated that deleting this internal workspace removes Claude’s multi-step reasoning capabilities while leaving basic factual recall intact. The company also partnered with Neuronpedia to release an interactive demo of these interpretability methods on open-weights models.

Tencent · Link Tencent launched Hy3, a 295-billion parameter Mixture-of-Experts (MoE) model under a commercial-friendly Apache 2.0 license. Hosted on Hugging Face and available via a free API tier on OpenRouter, the model aims to provide affordable agentic capabilities that rival trillion-scale flagship models.

Google DeepMind · Link DeepMind announced a research partnership with Apptronik to help train and advance Gemini Robotics. DeepMind will leverage real-world data collected by Apptronik’s new Apollo 2 humanoid platform as the hardware company expands its Robot Park facility.

Nvidia · Link Nvidia reported that its open-weights Nemotron family of models has officially surpassed 100 million downloads. This signals massive adoption of Nvidia’s open ecosystem assets among developers building enterprise AI.

OpenClaw · Link The local agent framework launched integration with Hugging Face local apps, enabling users to run cloud-free, tool-calling agents using any GGUF or MLX model. Concurrently, OpenClaw deployed significant iOS and Android updates featuring local gateways and a redesigned UI.

Google Cloud · Link Google AI Studio introduced one-click integrations with Cloud SQL and Firestore. Google positioned this capability as a way to turn “vibe coding” into an on-ramp for enterprise infrastructure, solving Day 2 deployment challenges.

Also Noted

• FactoryAI (Link): Partnered with Hugging Face to upload Droid agent traces to datasets, establishing a pipeline to fuel the training of next-generation open coding models.
• Wafer AI (Link): Demonstrated GLM 5.2 running on AMD MI355X chips at over 2x lower cost than Nvidia Blackwell GPUs, capturing approximately 80% of B200 throughput.
• Zapier (Link): Launched a Model Context Protocol (MCP) integration, securely giving AI agents direct operational access to over 9,000 applications.
• fal (Link): Announced a 72-hour AI video hackathon with a $150k prize pool, supported by Sequoia Capital, Google DeepMind, xAI, and Kling.
• Hyperbrowser (Link): Released Agentmap, an open-source tool that crawls and structures websites so that coding agents like Claude Code and Cursor can bypass blind exploration and map sites in one shot.