UK Foreign Secretary Warns AI Is the 'Biggest Security Risk of the Next Decade'

UK Foreign Secretary Yvette Cooper is set to describe AI as 'the biggest security challenge of the next decade.' Cooper plans to call on governments around the world to accelerate the construction of international guardrails to contain the risks AI poses.

The remarks come amid mounting concern over the security risks AI technology could unleash, such as disinformation spread via deepfakes and the automation of cyberattacks. The UK government has already led international debate by hosting the AI Safety Summit, but this statement is being positioned as a more direct warning.

Palantir's 'Secret Invoices' Problem Haunts Canada's AI Strategy

Canada's AI strategy is drawing criticism over the opaque contractual relationship between the government and data analytics giant Palantir. According to journalist Al Vigier, the Canadian government is paying substantial sums for Palantir's services while the details remain undisclosed.

The article drew 40 points on Hacker News, where the comment section saw lively debate over the transparency of government AI procurement processes and the risks of depending on a single vendor. The prevailing view was that with taxpayer money on the line, AI-related government contracts should be held to a higher standard of transparency.

Community Effort Makes DeepSeek V4 Flash 8-bit 3x Faster on a Mac Studio

A report shared on the LocalLLaMA community describes significant performance gains from running the 8-bit MLX build (302GB) of DeepSeek V4 Flash on a Mac Studio M3 Ultra (512GB).

Using Codex, the reporter optimized oMLX (an LLM inference engine for Apple Silicon), achieving roughly a 1.6x speedup in prefill (~300 tok/s → ~533 tok/s) and about 3x in decode (~7.3 tok/s → ~20 tok/s). The core of the optimization was enabling native Metal kernels specialized for DeepSeek's MoE (Mixture of Experts) architecture and streamlining the routing process.

Current real-world figures hold at about 19–20 tok/s even in 80K–120K long-context settings, putting the 302GB 8-bit model at a practical level. The reporter is asking the community to review the accuracy validation.

Fix for a llama-server KV Cache Restore Bug Delivers a 720x Speedup

A bug has been discovered in llama.cpp's server mode (llama-server) that effectively broke KV cache save and restore, and a fix from a community member delivered a roughly 720x speedup.

The problem: even after restoring session state following a process restart, the very next query discarded the entire cache and triggered a full recomputation (full prefill) from scratch. The root cause was that checkpoint metadata existed only in process memory and was never persisted to disk.

The fix adds just 117 lines, persisting the checkpoint list to a sidecar file. In tests with a 100K context, a delta query after restoration dropped from about 720 seconds (12 minutes) to about 1 second. The fix is published as a PR on the llama-cpp-turboquant fork, and the same bug has been confirmed to exist in upstream llama.cpp.

ASICS Unveils AI-Powered Shoe Design Technology

ASICS, working together with RebuilderAI, has announced a next-generation shoe design and manufacturing simulation technology powered by AI. Designers input 2D sketches or concepts, the AI generates 3D data, and the results feed directly into structural verification via CAE (computer-aided engineering) / FEA (finite element analysis).

Traditionally, getting from design to verification required multiple steps across specialized tools, but this technology is expected to shorten the design-to-verification timeline. In apparel and sporting goods, AI adoption is steadily expanding in the practical domain of design-process efficiency.

---