The case for keeping an AI model you can run yourself

Hugging Face says hosted models blocked requests containing real attack commands during its investigation. The team instead analyzed the logs with the open-weight GLM-5.2 model running on its own infrastructure.

The same week, Thinking Machines released Inkling’s full weights for customization and fine-tuning.

Hugging Face disclosed an intrusion on July 16 that it attributed to an autonomous AI agent system. The company described an attack that began with malicious dataset processing, gained greater system permissions, and moved between internal computing clusters.

Hugging Face said it closed the entry paths and rebuilt affected infrastructure. Its defensive analysis covered more than 17,000 recorded events, but the impact assessment was still underway, and the company did not know which model powered the attacker. It reported no evidence of tampering with public models, datasets, or Spaces and said its published software supply chain was clean.

Top Stories#

Thinking Machines released Inkling's full weights for customization. The company also made its multimodal model available for fine-tuning, with an emphasis on coding and tool use. Thinking Machines reports 975 billion total parameters, with 41 billion active, and pretraining on 45 trillion tokens spanning text, images, audio, and video.

The announcement describes a maximum context of one million tokens; Tinker initially offers 64K and 256K options. In one demonstration, Inkling generates data, runs fine-tuning, evaluates the result, and switches weights to avoid a particular letter.

OpenAI claimed a solution to a graph-theory problem pursued for more than half a century. Scientific American reported the claim about the cycle double cover conjecture.

A graph has vertices joined by edges; a cycle is a loop returning to its starting point. A cycle double cover is a collection of loops containing every edge exactly twice.

The conjecture says every bridgeless graph has such a cover: removing any one edge must not split a connected part into two. Earlier work had established special cases.

A nine-page exposition appeared on July 17, presenting the announced proof with changes intended to make it more accessible. The first version is a preprint and invites comments.

A biomedical research agent reproduced known genetic associations after human-AI review corrected discrepancies. Researchers describe a hypertension study using linked genetic and electronic-health-record data from 286,422 people under an aggregate-only data policy. Protected data stayed within institutional boundaries, with human oversight retained.

The agent planned cohort extraction, ran workflows, and generated quality-control and publication materials. Human-AI review uncovered differences in how hypertension was defined. After reconciliation, the workflow reproduced established genetic associations.

Top Links#

• Kimi Code's agent updates (kimi.com): The July 16 CLI entry adds background tasks, task lists, planning, skills, and nested agents to coding subagents.
• Performance per watt (blogs.nvidia.com): NVIDIA's infrastructure argument considers latency-throughput tradeoffs while comparing an eight-GPU Hopper domain with a 72-GPU Blackwell domain.
• The Tower Keeps Rising (lucumr.pocoo.org): An engineering essay argues that agents can accelerate individually reasonable code changes without the communication needed to keep a team's understanding of its software current.