This midday catch-up for September 13 brought two weighty themes side by side: personal data and AI risk. Apple published an article introducing its third-generation Foundation Model, and on Hacker News, the question of training AI on individuals' private data generated the biggest buzz of the time slot. Meanwhile, Anthropic's Dario Amodei continued his media push with a long-form piece in The Atlantic and a CNN interview. Behind the scenes, the open community was visibly busy with self-defense and infrastructure-building, from a mirror site launched out of concern over Hugging Face censorship and access restrictions to serverless hosting for LoRA adapters.

Apple — The Third-Generation Foundation Model and the "Trained on Private Data" Controversy

Apple's machine learning research team published an article introducing "the third generation of Apple foundation models." It drew the most attention of the period on Hacker News with 24 points and 16 comments, and the submitted headline pulled no punches: "Apple wants to train AI on your personal private data."

For the details of the third-generation model's training methods, refer to the original article for now — the real focus of the debate is the reaction. How far will Apple, a company that has long flown the "privacy" banner, go in using personal data to improve its models? This is the point where the practical benefits of on-device AI collide head-on with anxiety about data leaving the device. The structure at play — the more personalized the AI, the more personal context it needs, and that context is precisely the most sensitive data — is a challenge shared by every player, not just Apple.

Amodei's "Code Red" — The Atlantic Long-Form Piece and a Pointed CNN Interview (Updated from Last Time)

The series of arguments from Amodei — which this morning's roundup reported as a proposal for "speed limits" and auditors on AI development — has spread to even bigger outlets. The Atlantic published a long-form article titled "AI's Code-Red Moment." As the slug in the article's URL — "dario-amodei-slow-down-ai-save-humanity" — suggests, the piece apparently argues that the pace of AI development should be deliberately slowed.

On top of that, CNN released an interview video in which correspondent Anderson Cooper put the question to Amodei directly: "Do you seriously believe AI could kill all of humanity?" The unusual spectacle continues of a CEO touring major media outlets while voicing catastrophic risks himself. As I wrote last time, the weight of safety advocacy coming from the top of the company rather than the research sidelines is considerable, and its influence on the regulatory debate will be worth watching.

"The Hugging Bay" — Insurance Against Hugging Face Censorship

Over on Reddit's LocalLLaMA community, a new model download site called "The Hugging Bay" was introduced, built as a hedge in case Hugging Face starts censoring content or restricting access. The poster described its aim as "insurance in case HF starts censoring or restricting access."

As this blog has been tracking, the environment around open-weight models has shifted dramatically over the past few months. Licenses are diverging between East and West, conflict over unauthorized training via APIs has surfaced, and the recognition is spreading through the community that centralized model hubs are becoming single points of failure. The practicality and legality of mirrors deserve their own careful discussion, but the very fact that "download while you still can" has become shared sentiment speaks volumes about where the open-model ecosystem stands today.

"The Last AI Built by Humans" — Recursive Self-Improvement Research Hits arXiv

A paper titled "The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement" was posted to arXiv. When "recursive self-improvement" — AI improving AI itself — actually works, when and through what process will humans hand over the reins of model-building? The paper appears to confront that turning point head-on.

Self-improving AI is a familiar theme in science fiction, but real-world R&D has been steadily moving in that direction in recent years, notably through training automation with synthetic data. The "last" in the title is no exaggeration: it points straight at a central question of AI safety research — how to design and guarantee control once the improvement loop closes. For the details of the argument, refer to the paper itself for now.

Lorivo — A Serverless Platform for Sharing LoRA Adapters on "One GPU Server"

A post appeared on LocalLLaMA about building Lorivo, a serverless hosting platform dedicated to LoRA adapters. The developer's motivation is clear-cut: rank 8–32 LoRA/QLoRA adapters share more than 99% of their weights with the base model, yet operations demand an entire GPU instance for each one. To fix this inefficiency, they implemented a setup that stands up a single GPU server per base model and shares it across multiple adapters, using vLLM features such as batching, memory management, and fused LoRA kernels.

The flow is simple: upload an adapter, the base model is auto-detected, it is loaded onto the corresponding GPU server, and an OpenAI-compatible endpoint is issued. There is also a CLI that deploys with a single command like `lorivo deploy ./my-lora`, and at the moment Qwen 3.5 4B is always free to use. The privacy stance — chats and inference requests are not stored — is stated explicitly as well. As an attempt to fill in the "last mile" from fine-tuning to publication, infrastructure built by individual developers is steadily maturing.

RoastMyHarness — Benchmarking Your Homemade Agent Tooling Against the "Vanilla" Setup

Also making the rounds on LocalLLaMA is "RoastMyHarness." It is a small engine that runs the harness, extensions, skills, AGENTS.md, and other artifacts you built for the coding agent "Pi" against the same DeepSWE benchmark tasks as an unmodified, vanilla Pi (the control), then visualizes "what got better, what broke, and what it cost."

The poster also wrote that "beating vanilla Pi on task quality and token efficiency is surprisingly hard" — an interesting sober prescription for the agent-extension boom. If you build your own tooling, verify it with benchmarks, not conviction: that culture is taking hold around coding agents too.

Japan: Choosing Between Usage-Based and Per-Seat Pricing for Agentforce

Qiita published an interesting hands-on article about Salesforce's Agentforce. It points out that the real-world cost of operation can't be determined just by lining up monthly license fees against per-action pricing, then walks through a method for estimating the cost per completed task — noting that the same operation may or may not be usage-billed depending on the executing user and permissions, and bridging the gap between "the number of successfully billed actions" and "the number of customer requests actually completed."

The success of agent adoption depends heavily on billing design and unit-cost management — arguably more than on model performance. Case studies in Japan are still scarce, and the accumulation of this kind of practical, ground-level verification looks set to pave the way for agent adoption at Japanese companies.

Wrap-Up

The tug-of-war between privacy and personalization at Apple, and the continuing media rollout of Amodei's "code red" warnings on one side; on the other, self-defense and institution-building from the open community — a mirror site born of censorship fears around the hub, infrastructure for publishing LoRAs with ease, and a culture of putting homemade tools through benchmarks. We will keep following the process by which the very "premises of AI" are being reassembled on both the giant-platform side and the community side.