US Grants GlobalFoundries $300 Million — Silicon Photonics to Speed Up Chip-to-Chip AI Communication
The US government has announced roughly $300 million in funding for semiconductor manufacturer GlobalFoundries. The grant targets "silicon photonics" technology, with the aim of accelerating the development of links that connect AI chips to one another at higher speeds (Reuters).
In AI training and inference, the volume of data exchanged between chips has grown enormous, and conventional electrical wiring tends to become a bottleneck. Silicon photonics, which uses light as the medium to speed up communication, has long been regarded as a promising technology for improving data center performance.
Publicly available information at this point does not confirm which production lines or timelines the funding will specifically cover. That said, it can be read as part of the ongoing US push to strengthen the semiconductor supply chain and keep investing in AI infrastructure.
Australian Regulator Takes Telegram to Court — Cites Failure to Remove Terror-Related Material
Australia's online safety regulator, eSafety, has initiated civil penalty proceedings in federal court against the provider of the messaging app Telegram. The regulator alleges the company failed to meet its duty to detect and remove terrorist content, pointing to videos depicting executions and mass shootings that were left unaddressed (Bloomberg).
The case follows a year-long investigation by eSafety, and signals the regulator's willingness to hold platforms clearly accountable for handling illegal and harmful content. The platform's side of the story and the upcoming court schedule could not be confirmed in detail at this time.
This is not AI news per se, but the mounting regulatory pressure around major platforms' content moderation responsibilities is a development that could also shape debates over the circulation and monitoring of generated content.
Japan HP Ships PCs Bundled with Rakuten's Local AI "Rakuten AI 7B"
Japan HP and Rakuten have started prebundling the AI app "Rakuten AI for Desktop" on HP PCs. The bundled model is "Rakuten AI 7B," a 7-billion-parameter Japanese LLM that can reportedly run summarization and translation on the PC itself even without an internet connection (ITmedia AI+).
Running AI locally (on-device) offers privacy reassurance since no data is sent to the cloud, and is also considered advantageous in terms of communication costs and latency. Japan HP President Okado has positioned the move as "an important milestone for hybrid AI."
The pace of adoption, the range of supported models, and real-world accuracy all remain to be seen, but it is notable that a domestic LLM is teaming up with a PC maker to reach end users.
Research: The ISNAD Framework for Verifying the Sources of Claims in Multi-Agent Systems
A study featured on Hugging Face's daily papers proposes "ISNAD," a framework for improving the reliability of multi-agent (multi-AI collaboration) systems.
The authors point out that in multi-agent pipelines, even when intermediate agents are unreliable, the final answer is still presented fluently and confidently, so problems tend to worsen "silently." Existing research, they argue, has focused heavily on verifying agents' identities and permissions, while verifying whether individual "claims" are correct and well-grounded has received far less attention.
ISNAD applies ideas from the classical Islamic tradition of hadith verification (isnad and rijal). It reportedly combines a "weakest-link" approach that evaluates the chain of transmitters step by step and caps overall trust at the weakest link, cross-corroboration scores across multiple transmission paths, and a mechanism that separates source verification from content critique. The code and evaluations are publicly available, and the paper is said to candidly document which components have been verified and which remain unverified.
As agent orchestration spreads, this study is a useful reference for anyone confronting head-on the question of how to guarantee the provenance of final answers.
From the Field in Japan: Agent Memory Management, Coding Practice, and Test-Based Judgment
Japanese tech media saw a wave of shared know-how on running AI agents in production.
Managing memory with git, partitioned by "decay rate"
One developer shared their setup for managing an internal AI agent's "memory (brain)" with git. The agent's job, they write, is not to answer questions but to trace how the task at hand connects back to "which initiative and which objective." They argue that teams should stop consolidating everything into a single canonical source and instead split management according to each piece of information's "decay rate" (Qiita).
Agentic coding in practice
Another article observes that our relationship with AI has shifted from "completion (autocomplete)" to "delegation (agents)," and lays out a practical approach to "agentic coding" — handing implementation tasks to AI agents — grounded in the state of the field as of 2026 (Qiita).
Why they didn't ship even with all tests Green
Yet another article looks back at a case where the release decision was withheld even though every test was Green (passing). The more Greens piled up, the louder the question became: "Does this Green guarantee things through tomorrow morning's 06:30 operations?" — prompting a rethink of the quality bar a change must clear (Qiita).
All of these reflect a shared concern — how to trust and verify the results of delegating to AI — suggesting that agent operations are entering a phase of accumulating hard-won experience.
Community Watch: Does the Pruned Kimi K3 (342GB) Actually Run?
Over in the LocalLLaMA community, a post asked for hands-on impressions of running a pruned and quantized build of the large-scale model Kimi K3, labeled "IQ1_M / 342GB." Community verification of whether the compressed, more manageable version holds up in practice appears to be just getting started (Reddit). Kimi K3 itself is already a widely covered model; this time, the interest centers on a form that can actually run locally.