Introduction

Here are five key topics from the international AI news collected on July 12. It was a day packed with wide-ranging developments: regulation, model downsizing, hardware co-design, local LLMs, and how researchers are actually using AI.

---

1. EU AI Act Penalties Take Effect on August 2 — Fines of Up to 7% of Revenue

The remaining provisions of the EU's AI regulation, the EU AI Act, take effect on August 2, 2026, at which point the European Commission will gain the authority to impose sanctions (fines) on companies that violate it. Fines could reach up to 7% of total global revenue, making for a very significant impact.

That said, the core regulations for "high-risk AI systems" — those deemed to carry elevated risk — are reportedly deferred until 2027. In other words, not everything becomes strict overnight on August 2; enforcement will be phased in. Even so, for companies offering AI-related services in the EU, getting compliance preparations in order has become an urgent priority.

For Japanese companies as well, this regulatory trend is a critical factor to watch when considering entering or expanding in the EU market.

2. The Shift Toward "Small and Beautiful" AI Models Gains Momentum

According to a report by The Register, a movement to reassess "small models" is spreading among enterprise AI customers. Interest is shifting away from models with massive parameter counts toward smaller, more efficient ones.

This appears to be driven not only by cost considerations, but also by a growing recognition of practical operational benefits such as lower latency, stronger privacy, and the ability to run on edge devices. The increasing availability of models that are "small yet good enough" is reinforcing this trend.

3. NVIDIA's Case for Co-Designing AI Models and Hardware

NVIDIA's developer blog recently published an interesting technical article titled "AI Model Co-Design." It introduces an approach to "hardware-friendly LLM design" that takes hardware characteristics into account from the model design stage onward.

The idea is that by optimizing a model's architecture to match the structure of the hardware it runs on (such as GPUs), inference efficiency can be dramatically improved. This relates to the "small model" trend in the previous item: rather than simply making models smaller, the industry is exploring ways to achieve both performance and efficiency by improving affinity with the underlying hardware.

4. Running a 75B-Parameter Model on a 64GB Mac — A Community Challenge

A report on Reddit's LocalLLaMA community has been making waves, claiming that NVIDIA's "Nemotron Puzzle 75B" model runs smoothly on a 64GB M2 Max MacBook.

The poster added native support to mlx-lm, achieving an average generation speed of 14.27 tok/s with a 4-bit quantized checkpoint of about 42GiB and peak memory usage of around 50GB. Interestingly, the higher-precision 5-bit quantization pushed close to the 64GB memory limit and actually resulted in worse performance.

The post also shares a practical insight: the language modeling head in the output layer must be kept in BF16 (quantizing it to 4-bit produces repetitive garbage output). It's clear that the technical barriers to running large models locally are steadily coming down.

5. Elsevier Survey: Fewer Than Half of Researchers Actually Use AI

A survey of roughly 3,000 researchers worldwide, conducted by the academic publishing giant Elsevier, revealed that fewer than half of researchers are actually using AI tools.

This result highlights the reality that, despite the high expectations placed on AI in research, actual adoption remains limited. We could not confirm the detailed breakdown of the survey in the article, but it suggests that AI adoption in research faces its own unique barriers — not just technical hurdles, but also concerns about reliability and academic validity.

---

Wrap-Up

What today's news reveals is a picture of "technology evolving toward more practical and efficient approaches just as AI regulation gets serious." The activation of EU penalties will have a major impact on companies, but at the same time, progress in small models and hardware co-design is arguably opening the way toward more sustainable AI use.

---