Apple Lobbies US Government to Source Chips From Blacklisted Firm CXMT

It has emerged that Apple Inc. is seeking White House approval to procure memory chips from ChangXin Memory Technologies (CXMT), a company listed on the U.S. Department of Defense's "1260H List" of firms deemed linked to the Chinese military. The Financial Times reported the story, citing six people familiar with the matter.

The company has been lobbying Trump administration officials, including at the Department of Commerce, reportedly with the aim of keeping chip costs down. For a tech giant like Apple, securing memory chip suppliers bears directly on its competitiveness in AI and devices, and the move is now testing how that stacks up against national security concerns.

When Should an LLM "Think Deep"? — The IDPR Framework

Reasoning-optimized LLMs are powerful, but running deep reasoning on every input drives computational costs sky-high. To tackle this problem, researchers have proposed a framework called **IDPR (Inhibitory Deliberative Problem Reasoning)**, drawing inspiration from dual-process theory in cognitive science and the concept of response inhibition.

IDPR first generates a concise intuitive answer, then uses an inhibitory controller that reads the confidence signal of that answer to decide whether the problem is "worth thinking deeply about." On a 5,000-question math reasoning test, it invoked slow reasoning on just 8.20% of samples while lifting accuracy from 47.90% to 48.92%. Random routing under the same latency budget actually degraded performance by 1.14%, demonstrating the value of smart routing.

As a practical solution to the cost-accuracy trade-off in LLM reasoning, the approach holds promise for edge devices and real-time applications.

vLLM Sparse Attention Cuts TTFT for Long-Context RAG by Up to 9x

An implementation guide on leveraging vLLM's sparse attention has been published, targeting the latency problem in long-context inference.

Standard attention scales quadratically with sequence length, while sparse attention cuts this drastically by computing only the relevant portions. The guide compares three frameworks — FlexPrefill, LServe, and SparseServe — and in a proof of concept on an internal knowledge-search RAG system (32K–128K tokens), TTFT (Time To First Token) was reduced by 60–90%.

As a practical optimization technique for running long-context RAG in production, it is attracting attention across the engineering community.

Why Codex CLI's Web Search Doesn't Fetch Live Pages

Bolting web search onto a coding agent trades convenience for security risk. OpenAI's Codex CLI has taken a conservative design choice: even with web search enabled, it **does not fetch live pages by default**.

The reason is defense against prompt injection. Because an agent processes "instructions from the system" and "loaded text" in the same context, the content of a malicious web page can act directly as an instruction for "what to do next." Codex uses only cached search results and makes live page access an explicit opt-in.

It offers an interesting, production-level reference point for how to design the balance between agent autonomy and safety.

Doing Chemistry with LLMs: How You "Hand Over" a Molecule Changes Performance

Simply dropping SMILES (a format that expresses molecular structures as strings) into an LLM prompt makes it hard to get the model to reason correctly about molecular structure. An engineer who worked through a recent arXiv paper distilled three approaches to this problem and published a case study implementing them in a Rust library.

The proposed approaches are threefold: (1) use molecular representation formats whose structure is more explicit than SMILES, (2) also pass physical property data and similar molecules as context, and (3) specialize the LLM's role in reasoning rather than structural understanding.

The SMILES for aspirin, "CC(=O)Oc1ccccc1C(=O)O", is readable to a chemist, but an LLM has to parse the bonding relationships between atoms straight from the string — a perfect emblem of the challenge — making this an insightful piece for prompt engineering with domain-specific LLMs.

---