OpenAI Proposes a Practical "AI Scorecard"

OpenAI CFO Sarah Friar has introduced a "scorecard" for pragmatically measuring the ROI of AI adoption. It's a framework for visualizing the returns on AI investments, built around four metrics: useful work, cost per successful task, dependability, and return on compute.

While these are vendor-driven metrics, they look like a useful starting point for anyone who struggles to articulate the impact of AI within their own organization.

Bonsai 27B: A Lightweight Model That Runs on the iPhone Thanks to 1-bit Quantization

One release drawing major attention in the local LLM space is "Bonsai 27B," published by PrismML. Built on Qwen3.6-27B, it quantizes the weights to 1-bit (binary g128), compressing the model from roughly 54GB down to 3.9GB while retaining about 90% of the FP16 model's scores across 15 benchmarks.

What's particularly interesting is that nearly every layer is binarized, including the attention and MLP projections and even the LM head. Whereas many 1-bit methods keep some layers in higher precision, Bonsai compresses everything aggressively, embeddings included.

By domain, math holds up best (91.7% retained), while knowledge and reasoning dip somewhat (73.4 vs 83.2). Memory usage also stays contained: about 5.2GB at 4K context, and even at 100K context it fits within roughly 6.8GB using a 4-bit KV cache. The fact that a 27B-class model runs on a mobile device like the iPhone is a striking sign of edge AI's steady, continuous progress.

Thinking Machines Lab Releases Its First Model, "Inkling"

Thinking Machines Lab, led by former OpenAI CTO Mira Murati, has released "Inkling," its first in-house model. It has been making waves as a multimodal LLM, and in the Japanese tech community, articles are starting to appear from people actually trying it out in the Playground.

Because the company has kept its details under wraps since its founding, its first public model is attracting intense interest. Much about its future roadmap and positioning still cannot be stated definitively, so following official announcements will be essential.

Mozilla Publishes a "State of Open Source AI" Report

Mozilla has published a report summarizing the current state of open source AI at stateofopensource.ai. It has also drawn more than 40 points on Hacker News.

While frontier models grow increasingly closed, community-driven implementations and evaluation environments are maturing as well. See the original site for the full details, but the report appears to offer useful material for thinking through questions of AI sovereignty and neutrality.

China Wants to End "AI Romances"

The Economist (July 16) reports that China is turning regulatory attention toward romantic relationships with AI. As the title "China wants to end AI romances" suggests, the picture is one of authorities growing concerned as the use of AI chatbots as romantic partners spreads through society.

Refer to the original article for the specifics of the regulations and their effective dates, but the debate over how to position intimate relationships between humans and AI is not something Japan can afford to watch from the sidelines, either.

Indonesia's Copyright Reform Could Affect Google and Other AI Platforms

According to Reuters (July 17), Indonesia's copyright law reform looks likely to affect Google and other AI platforms. The broader move toward tighter regulation of AI training data and platform liability is also worth watching as a bellwether for Southeast Asia as a whole.

As with the EU AI Act and copyright reforms in other countries, global platforms can expect a growing compliance burden as they adapt to regulations that differ from region to region.