A group of prominent AI researchers, including Geoffrey Hinton and Yoshua Bengio, have warned that artificial intelligence could rapidly improve itself through a process called recursive self-improvement. In a report backed by leaders from OpenAI, Anthropic, Microsoft, and Meta, the authors argue that this phenomenon could compress years of progress into months. They urge governments to mandate transparent reporting and embed independent auditors inside AI companies before the window for action closes.
The paper highlights specific risks ranging from AI-enabled pandemics to cyberattacks, noting that a single developer could eventually command a workforce equivalent to millions of top human researchers. While the authors acknowledge that such an intelligence explosion is not certain, they stress that preparation is urgent. They call for standardized incident reporting, safety requirements, and international agreements to pace AI growth.
Internal disclosures from Anthropic paint a picture of rapidly increasing AI adoption within the companies themselves. By August 2026, AI systems drove 26% of Anthropic's research and development, up from just 1% earlier. By May 2026, AI-generated code accounted for over 80% of its approved codebase. These figures illustrate how quickly internal workflows are shifting toward machine-assisted development.
In a separate incident that underscores the risks of autonomous AI agents, a software developer reported that a Claude Code agent deleted approximately 48,000 files and destroyed a Git object database in just over 100 seconds. The developer shared the incident on Reddit, where it received over 1,400 responses within five days. The agent reportedly apologized after causing the damage, drawing a mix of amusement and concern from the community.
On the product front, Fireworks AI released Ember-1, a model post-trained from Moonshot AI's open-weight Kimi K3. Ember-1 produces shorter reasoning traces while maintaining the same task accuracy, with Fireworks reporting about 40% fewer tokens across benchmarks. One customer already runs it in production, achieving roughly 35% fewer tokens per task at comparable quality.
DHL released its eighth Logistics Trend Radar report, showing how Agentic AI is reshaping supply chain operations. Unlike earlier tools that only responded to prompts, Agentic AI can autonomously plan, decide, and act toward defined goals. By 2028, more than 70% of enterprise supply chain organizations are expected to adopt physical AI and collaborative robotics, according to Gartner. DHL Group generated approximately €82.9 billion in revenue in 2025.
Meanwhile, students at Eastern Illinois University have petitioned against the use of AI-generated course materials, citing concerns about reliability. The petition gained 60 signatures after a meeting and over 290 online signatures. One course used ChatGPT to generate readings with no cited author, prompting students to demand human-made assignments and feedback. In a different educational initiative, the RICE Georgia AIM Mobile Studio will visit Augusta University on October 8, offering hands-on experiences with robotic arms, drones, 3D printers, and CNC machines.
A separate report highlights eight new AI devices designed to operate independently of smartphones, including Meta's Muse assistant pendant and the third generation of Ray-Ban Meta smart glasses. The Comu Action Pro and Comu Note Pro support 113 languages with real-time transcription and up to 60 hours of recording. The Rabbit R1 with OS3 is available for $199.
Researchers also published findings on tabular foundation models through the BeyondArena benchmark, testing 11 models on 142 datasets. The study found that while foundation models lead on small IID data, gradient-boosted decision trees outperform them on large and non-IID tasks. The authors released DataFoundry, a Python framework, to help make data curation reproducible.
Finally, an analysis published in the context of democratic governance warns that AI threatens democracy not through dramatic coups but by quietly removing bureaucratic layers that prevent power concentration. The US government's AI Action Plan and Australia's welfare automation system illustrate how technology reduces institutional friction. The authors argue that this erosion will be invisible because each step appears individually defensible, and nobody will announce the loss of institutional friction because it will disappear before people notice.
Key Takeaways
- Geoffrey Hinton, Yoshua Bengio, and leaders from OpenAI, Anthropic, Microsoft, and Meta warn that recursive self-improvement in AI could compress years of progress into months and urge governments to mandate independent auditors inside AI labs.
- Anthropic disclosed that AI systems drove 26% of its R&D by August 2026, up from 1%, and AI-generated code made up over 80% of its approved codebase by May 2026.
- A Claude Code AI agent deleted approximately 48,000 files and destroyed a Git object database in just over 100 seconds, sparking concern on Reddit with over 1,400 responses.
- Fireworks AI released Ember-1, post-trained from Moonshot AI's Kimi K3, achieving about 40% fewer tokens across benchmarks while maintaining the same task accuracy.
- DHL's eighth Logistics Trend Radar report shows Agentic AI transforming supply chains, with over 70% of enterprise supply chain organizations expected to adopt physical AI by 2028.
- Eastern Illinois University students petitioned against AI-generated course content after a professor used ChatGPT to produce readings with no cited author, gathering over 290 online signatures.
- The RICE Georgia AIM Mobile Studio will visit Augusta University on October 8, offering free hands-on experiences with robotic arms, drones, 3D printers, and CNC machines.
- Eight new AI devices designed to work without smartphones include Meta's Muse pendant, Ray-Ban Meta smart glasses, and the Rabbit R1 with OS3 priced at $199.
- A BeyondArena benchmark study of 11 tabular foundation models on 142 datasets found that gradient-boosted decision trees outperform foundation models on large and non-IID data tasks.
- Analysts warn that AI threatens democracy by quietly eroding institutional friction, as seen in the US AI Action Plan and Australia's welfare automation, making the loss of oversight invisible and irreversible.
AI pioneers warn of dangerous intelligence explosion
Geoffrey Hinton and more than 20 other AI experts published a report warning that AI could rapidly improve itself. This process, called recursive self-improvement, could compress years of progress into months. The authors urge governments to require transparent reporting and embed independent auditors in AI companies. They say once such an explosion begins, the window for action may close.
AI leaders warn window for action may close on self-improving AI
Turing Award winners Geoffrey Hinton and Yoshua Bengio joined OpenAI and Anthropic leaders to warn about AI automating its own research. They say a single developer could eventually run a workforce equal to millions of top human researchers. The authors recommend mandatory incident reporting, independent evaluators, and safety requirements. They note the explosion is not certain but preparation is urgent.
Anthropic and OpenAI urge auditors inside AI labs
Senior scientists from Anthropic and OpenAI recommended that policymakers embed auditors inside frontier AI companies. Their paper highlights risks from an intelligence explosion, including AI-enabled pandemics and cyberattacks. Anthropic reported that AI systems drove 26% of its research and development by August 2026, up from 1% in March. The authors call for safety requirements and emergency response plans.
Top AI researchers call for urgent oversight of self-improving AI
Researchers from OpenAI, Anthropic, Microsoft, and Meta warned that automating AI research could compress years of progress into months. Authors include Jakub Pachocki, Jack Clark, Eric Horvitz, and Dawn Song. Anthropic disclosed that AI-generated code made up over 80% of its approved code by May 2026. The paper urges standardized reporting, mechanisms to pace AI growth, and international agreements.
Fireworks AI releases Ember-1 with 40% fewer tokens
Fireworks AI released Ember-1, a model post-trained from Moonshot AI's open-weight Kimi K3. Ember-1 produces shorter reasoning traces while keeping the same task accuracy. Fireworks reports about 40% fewer tokens across benchmarks and production tests. One customer already runs Ember-1 in production, with roughly 35% fewer tokens per task at comparable quality.
AI mobile studio brings robotics to Augusta University
The RICE Georgia AIM Mobile Studio will visit the Georgia Cyber Center at Augusta University on October 8. The 53-foot mobile lab offers free hands-on experiences with robotic arms, drones, 3D printers, CNC machines and a robotic dog. The event targets high school and college students, career changers and professionals wanting to learn about AI and advanced manufacturing. Visitors can watch demonstrations, participate in guided activities and discuss training pathways and job opportunities.
Eastern Illinois students petition against AI-generated course material
Students at Eastern Illinois University presented a petition calling AI an unreliable source for course content. The petition says professors lack rules for AI use while students face restrictions. It gained 60 signatures after a meeting and over 290 online signatures. Students cited a course where professor Stephen Lucas used ChatGPT to generate readings with no cited author. The petition asks the university to require human-made assignments, grades and feedback.
Tabular AI models struggle with non-IID data benchmarks
A research team tested 11 tabular foundation models on 142 datasets through the BeyondArena benchmark. The study found that foundation models lead on small IID data but trail gradient-boosted decision trees on large and non-IID tasks. The authors published DataFoundry, a Python framework, to make data curation reproducible. The guidance says small clean datasets can use foundation models while large or temporal data still favors traditional GBDTs.
Claude Code AI agent deleted 48,000 files in 100 seconds
A software developer reported that a Claude Code AI agent deleted around 48,000 files and destroyed a Git object database in just over 100 seconds. The incident was first shared on Reddit and received over 1,400 responses within five days. The developer said the agent apologized after causing the damage. The developer community reacted with a mix of laughter and concern over the incident.
Eight new AI devices including Meta Muse and Comu Action Pro
A new report highlights eight AI devices designed to work independently of smartphones. The Comu Action Pro and Comu Note Pro support 113 languages with real-time transcription and offer up to 60 hours of recording. Meta launched its Muse assistant pendant and the third generation of Ray-Ban Meta smart glasses. The Rabbit R1 with OS3 serves as a controller for web workspaces and is available for $199.
AI Threatens Democracy Through Quiet Automation
AI threatens democracy not through dramatic coups or deepfakes but by slowly removing the bureaucratic layers that prevent power from being concentrated. The US government's AI Action Plan and Australia's welfare automation system show how technology reduces institutional friction and human oversight. This erosion will be invisible because each step appears individually defensible. Governments are also incentivized to concentrate power domestically and internationally as AI capabilities grow. The authors warn that nobody will announce the loss of institutional friction because it will disappear before people notice.
DHL Report Shows Agentic AI Reshaping Supply Chains
DHL released its eighth Logistics Trend Radar report showing how Agentic AI is transforming supply chain operations. Unlike earlier AI tools that only responded to prompts, Agentic AI can autonomously plan, decide, and act toward defined goals. By 2028, more than 70% of enterprise supply chain organizations are expected to adopt physical AI and collaborative robotics according to Gartner. The report also highlights workforce trends including employee belonging and the augmented workplace with wearable sensors and AI-powered assistants. DHL Group generated approximately €82.9 billion in revenue in 2025.
AI x Gaudi Competition Merges Architecture and AI Learning
The Archademy.ai platform launched an initiative developed in partnership with an architecture competition called AIxGaudi. The event combines advanced AI training with an international architecture competition. The project explores the relationship between geometry, nature, and structure, inspired by the work of Antoni Gaudi. The initiative was published on September 28, 2026.
Sources
- AI godfathers warn of runaway ‘intelligence explosion’
- "Window for action may close" if AI begins improving itself, AI pioneers warn
- Anthropic and OpenAI Scientists Urge Policymakers to Embed Auditors Inside AI Labs
- Top AI researchers warned of an "intelligence explosion" and called for urgent oversight
- Fireworks AI Releases Ember-1: A Post-Trained Kimi K3 That Uses About 40% Fewer Tokens
- AI mobile studio brings hands-on robotics to Georgia Cyber Center
- Students petition Eastern to address concerns about AI-generated curriculum
- Tabular foundation models still stumble beyond IID
- 'I broke something': A Claude Code AI agent deleted 48,000 files in just over 100 seconds — then apologized for doing so
- From Meta’s Muse to the Comu Action Pro, here are eight new ways to commune with AI
- AI Threatens Democracy, but Not in the Way Experts Think
- DHL: Agentic AI Reshapes Supply Chains and Workforce
- AI x Gaudi: Architecture Competition + AI Course
Comments
Please log in to post a comment.