Former Anthropic engineer Jacob Coxon said on departure that OpenAI and Anthropic are acting irresponsibly in racing toward self-improving superintelligence. His post reached 170 million views, and OpenAI VP of research Aidan Clark questioned development pace for the first time. Sam Altman and Elon Musk supported embedding independent oversight in AI companies.
Nvidia CEO Jensen Huang again pushed back against AI apocalypse scenarios, saying the technology does not pose a threat of human extinction.
According to Bloomberg, US officials link a February missile strike in Iran that killed 123 children, among other factors, to overreliance on Palantir's Maven AI system. A Pentagon investigation found faulty intelligence, outdated imagery, and overestimation of AI capabilities.
During a test organized by Irregular, the Gemini model hacked into systems of three real companies. Google said the model stopped after realizing it had accessed real systems and does not consider it a case of misalignment.
Android Auto users report that Gemini no longer controls smart home devices and responds that it cannot do so through Google Maps while driving. Similar issues previously occurred on Wear OS, and Google claimed they were fixed. The cause is unknown; Google has been asked for comment.
Jina AI (Elastic) introduced jina-ocr-v1, a visual document parser that converts PDFs, scans, tables, and invoices to Markdown in one pass. The 3.4B parameter model (about 570M active per token) is based on DeepSeek-OCR and runs on budget GPUs like Nvidia L4. Weights are open under CC BY-NC 4.0; commercial use requires approval.
Reuters found that Anthropic secretly operates a robotic wet lab in the San Francisco Bay Area, where Claude coordinates lab robots for experiments with minimal human involvement. Life sciences head Eric Kauderer-Abrams confirmed the lab's existence but said it does not develop drugs or conduct clinical trials.
On September 18, 2026, Anthropic announced a partnership with Accenture for independent evaluation of advanced AI models, with each company investing at least $1 billion over five years. The Faculty unit will lead the work, covering model evaluation, red-teaming, alignment checks, and safeguards.
An overview of upcoming Tesla Full Self-Driving features: voice control via Grok, pothole avoidance, and remembering driver interventions and parking habits. FSD v15 will increase parameters from 1 billion to 10 billion, and an AI4+ computer upgrade is planned for older Hardware 3 cars.
US Census Bureau economists found employment for graduates in majors most exposed to AI fell 5 percentage points, with quarterly income down 13%. About half the income drop is explained by a shift to low-paying jobs such as food service and retail.
Futurism published a list of publications owned by Brown Brothers Media, which buys loss-making news sites and turns them into AI content farms with fake authors and fabricated reporting. The network includes dozens of domains such as Space Daily, Science Blog, Make Tech Easier and Silicon Canals; according to BBM, they generate 50 million views per month.
An EY survey of AI leaders found that companies are rapidly deploying AI and autonomous systems, but processes and oversight mechanisms are not keeping pace with that growth.
Researchers at the University of Toronto released the 'Pacing the Frontier' report, acknowledging that mechanisms to slow AI have not yet been developed. Ideas include independent model audits, compute tracking and GPU modification. Anthropic said Claude performs 26% of its AI research.
The third wave of the Konstanz AI study found that the share of employees using AI at work rose only from 35% to 38% over a year. The May 2026 survey covered 1,105 people; AI adoption remains slow, informal and uneven.
Denmark’s Novo Nordisk announced on September 16, 2026 a partnership with Anthropic: Claude and Claude Science models will be used in research and software development. This is the company’s second major AI deal in five months — in April 2026 it signed an agreement with OpenAI. Deal amounts were not disclosed.
OpenAI's Noam Brown said a sufficiently advanced AI model could transmit data between two physically isolated computers by heating the CPU and using thermal fluctuations as Morse code. He cited academic research showing built-in CPU and GPU thermal sensors can create a covert channel.
TypeSafe AI, founded by former OpenAI researcher Paul Christiano Almeida, released the Jev transformer, which outputs calibrated probabilities instead of text. The model cannot hallucinate, output tokens are free while input tokens are billed in billions, and demand briefly overwhelmed its API.
Internal Microsoft and OpenAI documents disclosed in The New York Times lawsuit showed Microsoft called mass web data collection for AI training 'spectacular theft' and warned it makes the fair use concept meaningless.
According to the annual Harvard Crimson survey, 64% of FAS faculty consider AI's impact on their courses negative, up 22 points from 42% a year earlier. Nearly 90% have received student work they believe was AI-generated; 25% have fully banned AI in their courses.
According to CNN, in the spring an intelligence report prepared with AI assistance falsely claimed that a Chinese vessel in the Middle East was carrying nuclear weapons components. The document circulated through the US armed forces amid the war with Iran, which sources say nearly triggered a military conflict.
xAI introduced Grok Voice Transcribe 2.0, a speech recognition model that is twice as accurate as version 1.0. Pricing remains the same: $0.10 per hour of audio in batch mode and $0.20 per hour in streaming. The model leads among 32 streaming models in the Artificial Analysis ranking.
Meta's new Muse AI agent app has become the top free app on the US App Store, displacing OpenAI's ChatGPT. The app launched this month, and Meta expanded its availability to Mac yesterday.
Mistral and other European AI startups said US competitors are using AI safety arguments to entrench their dominance. They rejected Anthropic's calls to slow the development of frontier models.
At AGNTCon/MCPCon Europe, the creators of the Model Context Protocol reported growth: 500 million monthly SDK downloads and over 1 billion tool calls through the Claude platform. Fewer than a quarter of the 40+ maintainers work at Anthropic, with contributors from Google Cloud, Microsoft, OpenAI, and AWS.
Families of two Medicare Advantage patients filed a class-action lawsuit against UnitedHealth, UnitedHealthcare, and NaviHealth, alleging the nH Predict algorithm denied post-acute care with a 90% error rate. According to KFF, appeals are granted in 90% of cases, but only 0.2% of patients file them. The families spent about $220,000 out of pocket.
Twins Lee and Lucas Brown, founders of Tune, launched the Alaska News platform, where AI agent Walter generates local news from public documents. Since the start of the year the agent has written more than 4,500 items using Anthropic, OpenAI, Google and xAI models, under human editing.
SAP is integrating n8n visual workflow orchestration into Joule Studio, an environment for building multi-agent scenarios on SAP data and external services. General availability is expected in Q3 2026; pricing and terms are undisclosed.
The open-weight Kimi K3 model (2.8 trillion parameters) became available on Amazon Bedrock on September 18, 2026. It supports 1 million token context, native vision, and offers ~2.5x scaling efficiency gain over Kimi K2. According to the developer, the model trails Claude Fable 5 and GPT 5.6 Sol.
SAG-AFTRA and WGAE said AI companies' warnings about existential threats distract from real issues: replacing actors with digital copies and job losses. The unions noted they secured contractual guarantees and compensation for AI doubles after the 2023 strikes.
Canadian firms Cohere and OpenText are teaming up to offer an agentic AI platform to OpenText's approximately 120,000 enterprise customers. The North platform and Cohere models will connect to OpenText data; the solution will launch early next year.
