Fujitsu Develops PHOTON, a New Architecture That Boosts LLM GPU Efficiency by Up to 475x

On June 24, Fujitsu announced the development of PHOTON, a new architecture that runs large language models (LLMs) efficiently on fewer GPUs. Per-GPU processing performance (throughput) is claimed to reach up to 475 times that of today's mainstream Transformer architecture.

PHOTON's biggest feature is a design that makes far better use of GPU compute resources, reportedly slashing the number of GPUs needed to operate LLMs. According to Fujitsu, this is expected to deliver cost reductions and a smaller environmental footprint at the same time.

LLM inference costs are a challenge for the entire industry, and architecture-level efficiency gains could have wide-ranging implications, from cloud providers to edge devices. The 475x figure is a theoretical maximum, and its real-world impact remains to be seen, but an approach driven by architectural innovation is worth paying attention to.

Mistral AI Releases High-Precision OCR Model "OCR 4"

Mistral AI has released OCR 4, a new OCR model that extracts text from PDFs, Word files, PowerPoint presentations, and other formats. According to the company, it outperformed competing models in 72% of blind test cases.

Document digitization is often the first step in enterprise AI adoption, and better OCR accuracy directly translates into higher-quality downstream RAG (retrieval-augmented generation) and document analysis. Mistral has built a strong track record with its text models, and its full-fledged entry into the OCR space looks set to add a new contender to the document processing race.

Language Models Keep Making "the Same Arguments" — Pangram CEO Proposes a New AI Detection Method

Max Spero, CEO of AI-generated text detection firm Pangram, argues that language models give themselves away through a "reasoning bias" when it comes to text detection.

According to Spero, when a language model is asked to write 100 arguments on the same topic, they converge on surprisingly similar points, whereas human reasoning is far more diverse and varied. This "argument clustering" property, he says, can be exploited to detect AI-generated text.

Detecting AI-generated text is a critical challenge for academia, journalism, and platform operators. As models evolve, however, this bias is likely to shrink, and the cat-and-mouse game between detection methods and generative models is expected to continue.

"Forgetting" May Improve AI Language Learning, MPI Study Suggests

A new study from the Max Planck Institute (MPI) suggests that giving AI human-like memory limitations (3–7 words) can improve language learning.

The hypothesis: human working memory is limited, but that very "forgetfulness" is what makes efficient language acquisition possible. AI models can hold vast amounts of context, yet remembering everything does not necessarily lead to optimal learning.

The study is a fascinating demonstration of the "cognition-inspired AI" direction — drawing insights for AI design from human cognitive mechanisms. That said, applying these findings to practical large-scale models will require further validation.

llama.cpp Gains Support for IBM Speech Model and Vulkan Improvements

llama.cpp, the leading engine for running LLMs locally, has received a batch of updates. The main changes are as follows:

Support for **IBM granite-speech-4.1-2b-plus** (PR #24818)

Support for **LiquidAI LFM2.5-ColBERT-350M / Embedding-350M** (PR #24913)

Major Vulkan backend improvements: CONV_3D support, overflow prevention by applying bias before softmax, and more

The Vulkan improvements could boost inference performance on AMD and mobile GPUs. The addition of speech model support also shows llama.cpp expanding its reach beyond text and into multimodal territory.

GMO Internet Group to Trial Quadruped Security Robot at JGSDF Camp

Four GMO Internet Group companies, working with Japanese robot developer Mirai Robot, have developed a quadruped security robot and announced that deployment trials will begin at a Japan Ground Self-Defense Force camp.

The robot aims to enable round-the-clock security operations and reduce staffing requirements for guarding. Quadruped robots excel at moving over uneven terrain and stairs, making them well suited for patrolling areas that fixed cameras struggle to cover.

The practical deployment of robotic AI in defense and security is progressing around the world, and it seems that homegrown efforts in Japan are now getting serious as well.

---