1. "AI Washing" — The Reality of Companies Rushing to Brand Themselves with AI

The "AI washing" issue reported by The Guardian on May 24 has drawn significant attention on Hacker News, picking up 46 points and 35 comments. Modeled on "greenwashing," the term refers to the phenomenon of companies rebranding themselves by emphasizing AI far beyond what their actual substance warrants.

Behind the rush of PR firms to relabel themselves as "technology companies" lies the intense interest in AI among investors and consumers. How much AI is actually built into their products and services, however, varies case by case, and the growing gap between the label and the reality is starting to raise concerns.

In the Hacker News comment section, remarks such as "the word AI is losing its meaning" and "a trend of prioritizing marketing over practicality" have been piling up, and the discussion has developed into a debate over the credibility of the AI industry as a whole.

2. BitCPM-CANN: Native 1.58-bit LLM Training on Huawei Ascend NPUs

The OpenBMB team has announced "BitCPM-CANN," a system that natively performs 1.58-bit (ternary) quantized training on Huawei Ascend NPUs. It has become a topic of discussion in the LocalLLaMA community on Reddit.

They ran experiments across four model sizes — 0.5B/1B/3B/8B — and reportedly retained 95.7% to 97.2% of full-precision performance for models of 1B and above. The 3B model achieved results on par with full precision on the BBH benchmark, and the 3B/8B models recovered nearly all of their GSM8K scores.

Notably, integrating quantized training costs only 4.5% in throughput (148 vs 155 TFLOP/s per NPU). Memory reduction at inference time reaches up to 8x (about 6x even after including the scaling factor), paving the way for practical low-bit training.

Also interesting from a data-efficiency standpoint: MiniCPM4 8B matches the performance of Qwen3-8B (trained on 36 trillion tokens) using only 8 trillion tokens.

3. Apple Apparently Preparing a Dedicated Gen AI Site Ahead of WWDC

MacRumors has reported that Apple is preparing a new "Gen AI" subdomain. The move appears to be aimed at AI-related announcements at the upcoming WWDC.

Over the past few years, Apple has focused on strengthening on-device AI, progressively integrating AI features into iOS and macOS. The launch of a dedicated site is read as a sign that the company intends to position Gen AI not merely as one feature among many, but as an independent product area.

Details remain unknown, and we will have to wait for the official announcement at WWDC.

4. Memory's Share of AI Chip Costs Surges to About Two-Thirds

According to data analysis by Epoch AI, memory now accounts for roughly two-thirds of the component costs of AI chips. The figure reflects the surge in memory demand driven by ever-larger AI models.

As LLM parameter counts keep growing, fast, high-capacity memory (particularly HBM) has become both a bottleneck and a primary cost driver. The trend signals that memory bandwidth and capacity, rather than raw compute, are increasingly the constraining factors in AI inference and training — with major implications for hardware design.

5. GitHub Copilot App Arrives as a Standalone Application

Making the rounds in a Qiita article is the "GitHub Copilot App," a desktop application released by GitHub. Unlike Copilot as a VSCode extension, it works as a standalone app independent of any IDE.

Released as a public preview in 2026, the app has drawn growing interest in the Japanese developer community over how it differs from and complements the VSCode version. Being able to use an AI assistant regardless of your editor marks a significant change.

---

**Also worth noting**: Hacker News recently featured "Strudel," a CLI tool that auto-generates commit messages using Apple's on-device LLM. Written in the Odin language, it's an interesting project that leverages a local LLM on Apple Silicon.