Fabricated Cases in KPMG's AI Adoption Report — The Danger of "Second-Order Hallucination"

KPMG published a report encouraging enterprises to adopt AI, only for it to emerge that case studies featuring UBS, the NHS, and others did not match the facts. Edward Tian, CEO of GPTZero, discovered the issue, and KPMG has since withdrawn the report.

Tian calls the phenomenon of a trusted consulting firm publishing AI-generated misinformation as-is "second-order hallucination," pointing to the danger of plausible AI-generated errors spreading through authoritative organizations. It's an ironic case of the very parties pushing AI adoption being tripped up by AI's own hallucinations.

Foreign-National Usage Restrictions on Anthropic's Latest Models, "Fable 5" and "Mythos 5" — Amazon CEO Also Voices Concern

The Trump administration has ordered Anthropic to restrict foreign nationals from using its latest AI models, Fable 5 and Mythos 5, both inside and outside the United States. According to Reuters, Amazon CEO Andy Jassy had conveyed concerns to senior administration officials about the security risks of these models.

This means Amazon, a major investor in Anthropic, held national security concerns about its partner company's frontier models. As improvements in frontier model capabilities become directly tied to national security debates, the three-way relationship among AI companies, investors, and governments is growing increasingly complex.

Running DeepSeek V4 Flash on a Dual DGX Spark Setup at 40 tok/s — The Frontier of Local Inference

The LocalLLaMA community on Reddit has shared benchmark results for running DeepSeek V4 Flash (a large-scale MoE model) on two NVIDIA DGX Spark units. The setup reportedly achieved around 40 tok/s at FP8 precision, with aggregate throughput reaching 350 tok/s when handling 32 concurrent requests.

For comparison, figures of about 46 tok/s on an RTX Pro 6000 (Q2 precision) and about 29 tok/s on a Mac M2 Ultra 192GB (Q2 precision) have also been reported. The dual DGX Spark configuration runs fast while maintaining higher precision at FP8, and reportedly holds an advantage in parallel processing as well. The connection requires a 200G/s ConnectX-7 cable (around $180), but being able to run a frontier-class model at practical speeds locally looks like a major milestone for the community.

"Home Opus" — A White Paper on Running Frontier Models at Home in a Post-Regulation World

A white paper titled "Home Opus" has been drawing attention on Hacker News. It is a technical document outlining methods for deploying frontier AI model weights in a local environment, prompted by the usage restrictions on Fable 5.

The full details are available on GitHub, and discussion has already begun among developer communities interested in self-hosting LLMs. As tighter regulation and advances in local deployment technology progress side by side, the importance of open-weight models is being thrown into sharp relief once again.

Repo-Slopscore — Detecting AI-Generated Commits in Git Repositories

A tool called "Repo-Slopscore" has appeared on Hacker News. It analyzes the commit history of Git repositories to detect AI-generated contributions.

As AI-generated code becomes an everyday reality, an attempt to visualize a project's overall "AI dependency" deserves attention from the perspectives of code quality management and security auditing. As a health metric for open-source projects, it is expected to see broader adoption going forward.

---