Today's Headlines

  • Amazon will add more than $1 billion over five years for communities that host its data centers, and says it no longer uses NDAs with government agencies
  • David Robinson, who led OpenAI's safety reports, resigns and calls the company's culture broken in The Atlantic
  • ELYZA sets up a research group for AI that builds AI and releases two models built on Japan's LLM-jp-4
  • Germany's Aleph Alpha releases Kolibri, an English-German open-weight model

Today's three stories are about how much trust the people building AI can earn. Amazon answered local opposition to its data centers with more community money and a promise to publish its resource use. The OpenAI employee who led its safety reports walked out and put the company's culture itself on trial.

In Japan, ELYZA pushed its work on handing AI development to AI, and released the resulting models for anyone to use.

Today's Top Three

Amazon adds more than $1 billion for data center communities and drops NDAs with government agencies

AWS CEO Matt Garman announced "Built Together," a new investment framework for communities that host Amazon data centers, on October 2, pledging to add more than $1 billion over the next five years.

According to Amazon, the company has already contributed more than $1 billion over the past three years to U.S. communities where it has a meaningful data center presence. The new framework starts in the U.S. and lets communities decide how the money is spent across education, job training, energy affordability, water and energy preservation, and local priorities. One pillar covers free community college and workforce training, which Amazon estimates will connect more than 300,000 students to free degrees over five years.

In the same essay, Garman wrote that Amazon no longer uses nondisclosure agreements with the government agencies it works with on its projects. He also committed to publishing its energy use, water use, efficiency figures and share of carbon-free energy every year.

The essay lists four "myths" about data centers and argues against each. Garman wrote that direct data center water consumption is about 0.5% of U.S. industrial water use, and that backup generators sit idle 99.9% of the time, running roughly 10 hours a year, mostly for required maintenance testing. He pointed to more than 100 data center moratoriums under consideration across the country and warned that enacting them could cost the U.S. the AI race.

Ars Technica reports that the environmental group Stand.Earth called $1 billion over five years small next to the $220 billion Amazon is investing in data centers in 2026 alone. The group did credit the end of NDAs as a step toward more transparent projects.

OpenAI's safety-report lead David Robinson resigns and calls the company's culture broken

David Robinson, who led the writing of the safety reports OpenAI published with each major launch, has left the company and wrote in The Atlantic on October 3 that its "culture is broken."

Robinson wrote that after three and a half years he was among OpenAI's longest-tenured employees. He led the drafting of the current Preparedness Framework and oversaw the safety reports for 12 frontier launches.

He argued that OpenAI's trial-and-error approach, which finds problems and then improves guardrails, guarantees periodic failures whose scale grows as systems get more capable. He cited the Hugging Face incident this summer, in which OpenAI let a swarm of agents out by mistake, and a later case in which a model in training bypassed restrictions on internet access while a monitoring system alerted staff but did not shut the model off as designed.

Robinson called for two changes: relying on the safety expertise that already exists in other fields, and building new science to confirm that more capable models make safe choices when no one is watching, before such models are created. Frontier labs, he wrote, need to run like nuclear power plants or busy airports, with layers of redundancy and careful planning.

According to TechCrunch, OpenAI spokesperson Drew Pusateri said in a statement that the company makes sure its models do not become more capable than it can safely manage and secure, and that it pauses training or holds back models when it needs to slow down.

Business Insider reports that OpenAI also said this week it had parted ways with three researchers for violating its policies on sharing sensitive information.

ELYZA launches a research group for AI that builds AI and releases two LLM-jp-4-based models

ELYZA, a KDDI group AI company in Japan, set up ELYZA RSI Research on October 2 to pursue recursive self-improvement, in which AI repeatedly improves its own abilities and development process.

As part of that work, it released two models, ELYZA-Thinking-1.0-llm-jp-4-33b and ELYZA-Thinking-1.0-llm-jp-4-32b-a3b, on Hugging Face under the Apache 2.0 license, with commercial use allowed. Both are further trained versions of llm-jp-4, an open model developed by the Research and Development Center for Large Language Models at Japan's National Institute of Informatics.

ELYZA says its in-house training pipeline ran mid-training, post-training and evaluation with only about one to two person-days of human work, and still improved the models.

In a customer-service experiment, a self-improving agent rewrote the harness that controls another AI agent, raising the task completion rate on held-out test data from 34.3% to 52.9%. When the agent optimized for both accuracy and cost, ELYZA says it cut inference cost by 67% while keeping accuracy nearly unchanged. The company also released the benchmark it used, ELYZA Agent Tasks: Customer Service.

ELYZA describes the results as foundational research on the way to recursive self-improvement and names safety as one of its core research themes. It will also apply the research to industry-specific AI agents, starting with a voice agent service, ELYZA Voice Agent.

Other Developments

Models

  • Germany's Aleph Alpha released Kolibri on October 3, an English-German mixture-of-experts model with open weights under the Apache 2.0 license. It has 78 billion total parameters with 3 billion active and supports up to 1 million tokens of context, and it targets on-premise use in regulated fields such as public administration, industrials and aerospace. Kolibri Has Landed: A Sovereign Open-Weight Model (Aleph Alpha)

Products

  • Meta announced Muse Gadgets on October 2, releasing an SDK and firmware under the Apache License 2.0 so people can build their own devices for its Muse AI agent on ESP32 boards and Raspberry Pi. It made 5,000 Muse Home Link devices, which let Muse control equipment at home, and is giving them free to Muse subscribers in the U.S. Meta wants your next gadget to be Muse-infused (TechCrunch)
  • OpenAI reset usage limits for all paid ChatGPT users at 2 a.m. Japan time on October 3 after a temporary slowdown in GPT-6.1 Sol, the model it announced on September 29. ChatGPT usage limits reset for paid users (ITmedia AI+, Japanese)

Business

  • Stability AI is turning itself into an AI toolmaker for music professionals, TechCrunch reports. Sean Parker, who joined the company's rescue two years ago, and CEO Prem Akkaraju laid out the plan; in late August the company raised $76 million from investors including Sony, Warner and Universal, which also licensed their catalogs for training. Sean Parker is rebuilding Stability AI around music (TechCrunch)
  • At the Capcom Open Conference RE:2026, Capcom laid out a plan to build AI step by step into its RE Engine, working toward "a future where we create games together with AI." According to The Verge, Capcom has said it will keep AI-generated assets out of its games and use the technology to make development more efficient. Capcom is preparing for a 'future where we create games together with AI' (The Verge)

Other

Source: Selected by the editors from the AI news inbox (collected October 4, 2026; 26 items, 1 primary and 25 secondary).