OpenAI's Chief Futurist Joshua Achiam Is Leaving the Company

According to WIRED, Joshua Achiam, who served as Chief Futurist at OpenAI, is leaving the company. Achiam spent roughly nine years at OpenAI working on AI safety research, and he also delivered memorable testimony in the Musk v. Altman trial.

His departure is being viewed in the context of a continuing exodus of key talent from OpenAI. Losing a researcher central to AI safety work could have an impact on the company's research organization.

Meta's "Muse Image" Release Puts Instagram Photos to Work for AI — Why You Need to Opt Out

Meta has released its new image generation model, Muse Image (which we covered in our previous digest), and WIRED has now reported on how the model's rollout means photos from public Instagram accounts are being used for AI image generation.

Users with public accounts must explicitly opt out if they don't want their photos used for AI generation. Because use is permitted by default, many users are unknowingly contributing their photos to AI training.

With the line between privacy and AI once again in the spotlight, criticism is mounting of a design built on "opt-out" rather than "opt-in" as the default.

GGUF Version of DeepSeek-V4-Flash Released, Local LLM Community Buzzing

As discussed in Reddit's LocalLLaMA community, Unsloth has published GGUF files of DeepSeek-V4-Flash in multiple sizes on Hugging Face. With the release in GGUF format, everyday users now have a path to running DeepSeek's latest model in their own local environments.

As costs climb and performance leaps forward among frontier models, running open-weight models locally is becoming increasingly practical from privacy and latency perspectives as well.

New Anthropic Research: How a "Global Workspace" Forms Inside Language Models

Anthropic's research team has published a paper titled "Verbalizable Representations Form a Global Workspace in Language Models." The study, which has also drawn attention on Reddit's LocalLLaMA, suggests that the internal representations of language models have a structure similar to what cognitive science calls a "global workspace."

The paper investigates how "verbalizable representations" form inside the model and are shared across multiple processing modules, and it is regarded as a significant step forward in AI interpretability research. Understanding a model's internal structure bears directly on improving its safety and reliability.

NTT Docomo Business Launches Distributed GPU Environment Leveraging IOWN

According to ITmedia, NTT Docomo Business has begun offering a proof-of-concept environment that makes integrated use of GPUs distributed across eight sites nationwide via the next-generation "IOWN APN" network. The setup can reportedly transfer 25GB of data in two seconds.

For Japan's AI infrastructure, which faces power constraints and data sovereignty challenges, this approach of integrating physically distributed GPU resources over a low-latency network presents an alternative to the cloud-concentrated model.

The "Three Pitfalls" of Local Government Operations Lurking in the Government AI "Gennai (GENAI)"

Also on ITmedia, an analysis piece examines the Digital Agency's government AI, "Gennai." According to the author, who works on local government DX as an assistant CIO, Gennai is groundbreaking as a government-released AI platform, but when it comes to real-world deployment by local governments, three challenges arise from its "cloud dependency."

Initiatives providing AI infrastructure at the national government level are rare worldwide, but the article argues that carefully designing dependencies, operating costs, and data management practices will be essential to lowering the adoption barrier on the ground.

---