This study investigated how Large Language Models (LLMs) align with the neural representation and computation of human language, using representational geometry as a mechanistic lens.

By tracking entropy, curvature, and fMRI encoding scores throughout the training of Pythia (70M-1B), a geometric modularization was identified where layers self-organize into stable low- and high-complexity clusters.

The low-complexity module, characterized by reduced entropy and curvature, consistently better predicted human language network activity. This alignment followed heterogeneous spatial-temporal trajectories: rapid and stable in temporal regions (AntTemp, PostTemp), but delayed and dynamic in frontal areas (IFG, IFGorb).

Crucially, reduced curvature remained a robust predictor of model-brain alignment even after controlling for training progress, an effect that strengthened with model scale. These results link training-driven geometric reorganization to temporal-frontal functional specialization, suggesting that representational smoothing facilitates neural-like linguistic processing.

👉 Read the full article on arXiv

This paper advocates for a genuinely life-inspired approach to machine intelligence, drawing on the adaptive and goal-directed behavioral strategies found in biological systems, which remain a central challenge in current AI research.

It is argued that biological evolution has discovered a scalable recipe for intelligence that enables robustness, autonomy, and open-ended problem-solving across diverse scales. This recipe is based on five design principles: multiscale autonomy, growth through self-assemblage of active components, continuous reconstruction of capabilities, exploitation of physical and embodied constraints, and pervasive signaling enabling self-organization and top-down control from goals.

These principles contrast with current AI paradigms and outline pathways for integrating them into future autonomous, embodied, and resilient artificial systems. Intelligence is framed as flexible problem-solving, and the concept of "cognitive light cones" is used to characterize the continuum of intelligence in living systems and machines.

👉 Read the full article on arXiv

The open-source AI agent platform OpenClaw (formerly Clawdbot and Moltbot) presents significant security risks despite its convenience. Over 135,000 internet-exposed OpenClaw instances have been discovered, many accessible without authentication due to default settings.

The OpenClaw "skill store" (ClawHub) is riddled with malicious extensions capable of stealing sensitive data such as API keys, personal information, and credit card details. Multiple vulnerabilities (CVEs) related to OpenClaw have been reported, with some malicious skills downloaded thousands of times.

OpenClaw can execute shell commands, read/write files, and run scripts, granting it extensive privileges on user systems. This capability makes it susceptible to severe security incidents if misconfigured or if malicious skills are installed. In response, OpenClaw has partnered with VirusTotal to enhance the scanning of extensions uploaded to its skill marketplace.

👉 Read the full article on The Register / SecurityScorecard