Tech • IA • Crypto
Claude “J-space” reveals hidden reasoning; DoorDash scales AI coding
Briefing
Claude J-space exposes hidden cognition
Researchers analyzing Anthropic’s Claude identified a structured internal representation dubbed “J-space” using Jacobian-based methods. This space maps latent activity patterns to specific conceptual tokens the model appears to be “thinking.” The finding offers a rare window into otherwise opaque neural processes. It suggests large language models maintain an internal workspace rather than producing outputs purely reactively.
Internal steps emerge without output
Experiments show Claude reaches correct answers without displaying intermediate reasoning steps. However, J-space analysis reveals sequential internal states corresponding to values like “21,” “42,” and “49.” This indicates the model performs structured, multi-step reasoning internally. The results challenge assumptions that hidden-chain reasoning is absent when not explicitly shown.
Global workspace theory echoed in AI
The structure of J-space closely mirrors the global workspace theory from cognitive science. That theory posits conscious thought arises when information is broadcast across a central neural workspace. Similarly, Claude appears to coordinate reasoning through a shared internal representation. This alignment strengthens arguments that advanced AI may parallel aspects of human cognition.
Researchers steer model thought patterns
The study indicates partial ability to influence Claude’s internal workspace directly. By manipulating inputs, researchers could nudge activity within J-space toward desired conceptual states. This opens the possibility of guiding reasoning paths without altering final outputs explicitly. It also raises new questions about control, interpretability, and alignment.
DoorDash scales AI coding company-wide
DoorDash has rolled out AI-assisted coding tools across much of its workforce, including non-engineers. Leadership reports “massive increases in throughput” as employees integrate AI into daily workflows. Tools are embedded across environments like Slack, Gmail, and internal systems. The push aims to raise baseline technical capability across the organization.
2026 marks coding capability inflection
Executives point to a sharp leap between 2025 and early 2026 in AI coding reliability. Earlier tools struggled with real-world deployment, but newer systems handle multi-language production code effectively. This shift has made AI viable for end-to-end feature development. It represents a transition from experimental to operational use.
Engineers supervise autonomous agents
Developers at DoorDash increasingly act as supervisors of AI coding agents rather than primary code authors. Teams run multiple parallel agent sessions across repositories, often using worktrees to manage conflicts. Some engineers aim to minimize manual coding entirely. The workflow emphasizes orchestration, validation, and iteration over direct implementation.
Workflows and teams rapidly restructure
The adoption of AI coding is reshaping software development processes and team structures. Faster iteration cycles are changing expectations around delivery speed and ownership. Non-engineers are contributing to technical workflows, blurring traditional role boundaries. Companies are beginning to reorganize around AI-augmented productivity rather than human-only output.
Videos covered
What’s at the center of Claude’s mind?
- •Discovery of a “J-space” in AI
- •Parallel to human conscious thought
- •Evidence of step-by-step internal reasoning
DoorDash gave every employee Claude Code
- •Company-wide AI adoption accelerates productivity
- •Inflection point in AI coding capabilities
- •Engineers shift from writing code to supervising agents