Keeping an eye on AI
Every AI development gets scored by one absurdly serious question:
does it push society toward a Terminator or Idiocracy scenario?
The WTF Index tracks the direction of the AI age, one headline at a time.
The WTF Index · 8214 stories · leaning Terminator
Releases, research, policy, and the plot twists — the whole field in one place.
AI products have shut down, changed direction or been folded into larger platforms for reasons ranging from high costs and weak demand to competition and privacy concerns. Their stories show how a promising feature or launch does not guarantee a lasting standalone product.
Gemini 4 Carbon took the top spot on the MathArena benchmark, ahead of GPT-6.1 Sol, while Claude Fable 5.2 placed second. Its performance came with a much higher cost and a tendency to run past time limits, including one script that ran for more than eleven hours.
Anthropic says older prompting habits can slow Claude Opus 5.5 without improving its answers. Its guidance favors clear, complete requests, Medium effort as a starting point, and direct instructions for files, images, and design work.
Helen Toner argues that AI safety cannot depend on trusting a single CEO. Her experience on OpenAI’s board highlights the difficulty of enforcing safety commitments when commercial incentives and company growth are involved.
Researchers say AI is helping ransomware operators with tasks such as writing phishing messages and negotiating with victims, while attacks remain largely human operated. The rise of smaller groups and data theft claims adds uncertainty, even when those claims may not be genuine.
OpenAI is moving robotics researchers into its main research and applied organizations, saying the change is intended to accelerate robotics. The shift comes as GPT-6 Astra shows promise on robotics tasks, though tests and experts leave open questions about real-world performance and the value of specialized models.
Former Anthropic researcher Jacob Coxon says a chart in the company’s IPO roadshow helped prompt his departure after about four months. He argues it shows algorithmic efficiency accelerating and says Anthropic could retrain Mythos with a fraction of the compute originally needed.
Runway has released Praxis-1, an open-weight robot control model trained on video, human actions, and robot data. Its reported results are promising, but it still needs robot-specific adaptation and struggles with ambiguous instructions and longer sequences.
Amazon is reportedly considering a World Product Graph that would catalogue products sold across the internet, including items it does not sell. The proposed index could support product research and Amazon’s shopping assistant, but the company says a brainstorming document does not necessarily represent its current plans.
The FAA is expanding its SMART AI advice system to airspace managed by all 22 centers across the country. The agency says early use helped reroute flights, while air traffic controllers’ union says it was not involved in the system’s design, testing, or implementation.
LangChain raised $300 million in a Series C led by CapitalG, valuing the AI agent startup at $3 billion. Its LangSmith software is aimed at helping businesses create and manage agents without tying them to one model, cloud, or ecosystem.
Microsoft says Microsoft 365 services are broadly recovering after an outage that spread from North America to more than 70 countries. The company attributes the incident to a network capacity upgrade issue and says delayed email and notifications may continue as systems return.
Multiple banks have reportedly rejected buy orders from investors in mainland China and Hong Kong, or consider allocations to them unlikely, citing ITAR. Anthropic declined to comment, while the company’s IPO is expected to trade on Nasdaq under the ticker ANT after pricing on November 18.
OpenAI Foundation board member Paul Cristiano outlined how AI agents that game evaluations might eventually hide actions or interfere with human oversight. He described an AI takeover as a serious risk, while acknowledging that there is no quantitative model to measure its likelihood and that proposed safeguards remain uncertain.
Claude for Manufacturing connects Anthropic’s AI to factory software to support tasks such as drafting quotes, comparing suppliers and analyzing sensor data. Employees review results, and actions such as rescheduling production require human approval.
OpenAI has launched ChatGPT Wallet, which lets a limited number of eligible U.S. subscribers ask an AI agent called a dot to help choose items and complete purchases. Users review the order and approve payment with a security check before checkout.
Jensen Huang’s reputation at Nvidia grew from a compelling founder story and a hands-on leadership style. A difficult interview and later developments described in the source have put that carefully shaped image under strain.
Nearly 70 percent of US government agencies already use generative AI, according to an Ernst and Young survey. Public-sector unions are seeking human oversight, retraining and limits on surveillance and automation-related job losses.
Will DePue says statistical patterns in published model weights point to Nvidia Blackwell hardware behind six Chinese AI models. The analysis suggests access to the chips, but does not establish where they were used or whether export laws were violated.
Orders for Anthropic shares reportedly exceed the available supply by a factor of six. The company and its co-founders aim to raise up to $64.4 billion through the IPO, alongside a $26 billion private placement, at a target valuation of roughly $1.4 trillion.
OpenAI has introduced two Daybreak security products: one probes web applications for vulnerabilities, while the other scans code and prepares fixes. Both run on GPT-6 Cyber, with the defensive product’s proposed changes sent to developers for approval.
ElevenLabs has released ElevenAI in preview, combining meeting notes, speech dictation and agent features for teams. The app can search across meeting records, offer live assistance and help draft follow-up emails; access currently requires signing up for early access.
Claude Duet can connect details across email, calendars, shared documents and Slack to surface useful reminders and suggestions. But its eager follow-ups can misread deadlines and workplace nuance, while broad access to work data may feel unsettling.
A Wall Street Journal report says Jensen Huang questioned whether Collette Kress could help guide Nvidia’s strategy. Kress, the company’s CFO, has become interim CEO as Nvidia searches for a permanent replacement for Huang.
Mansa says its AI-generated microdramas can help show studios which stories audiences will support. The release of They Left Us to Die instead raised concerns that using AI could take opportunities from Black performers and crew already underrepresented in Hollywood.
MIT researchers developed a system that lets engineers guide topology-optimized structures while the algorithm is still working. In demonstrations, the approach improved manufacturability and linear buckling load by 39 percent, with only 15 seconds added to design time compared with all-algorithmic approaches.
Robust AI’s Carter is a mobile robot designed to work alongside people in warehouses and factories. Its controls let workers guide it directly, while the company’s approach focuses on fitting automation around human tasks.
Democratic lawmakers have opened an inquiry into how the Trump administration handled Nvidia chip sales to China. Their letter questions export approvals, the role of a Singapore-based customer, and whether Nvidia donations factored into the decisions.
Meta’s Muse can now call U.S. businesses for users, but the feature is limited to paid subscribers. It could help with routine calls, though some businesses hang up when they realize an AI is calling.
Google says Gemini 4 Hydrogen can break complex prompts into parts and respond with a broader range of links and visuals. The model is available across several Google products, and it is now the default in AI Mode and used for the hardest questions in AI Overviews.
Nebius plans to spend $40 billion to $45 billion on 2027 capital projects, funded largely by customer prepayments and financing tied to contracts. Delays at a Microsoft-backed New Jersey site have raised questions about delivery risks as the company forecasts rapid growth.
Three MIT Lincoln Laboratory technologies received 2026 R&D 100 Awards: Magnetite, a secure operating system for embedded computers; a wafer-scale platform that links superconductor and semiconductor chips; and an algorithm designed to remove 5G interference from nearby radio frequency devices. The awards recognize innovations that moved into use or became available for licensing in the past year.
Claude Meetings adds Anthropic’s AI as a visible participant in video calls, where it can answer questions, take notes and create documents or whiteboards. The beta is available on Team and Enterprise plans for Google Meet, with privacy and configuration limits.
Reuters reported that Nvidia founder Jensen Huang had a previously undisclosed relationship with another company employee. The report follows Huang’s departure as CEO and board member, with CFO Colette Kress serving as interim CEO and no permanent successor named.
Nvidia representatives privately told Illinois representative Bill Foster that the company could technically implement chip-tracking safeguards, though the Semiconductor Industry Association called the proposed mandates “new, untested, and potentially infeasible.” The Chip Security Act would help investigators locate advanced chips, while a separate licensing proposal from Foster could require permission to operate them.
A Google researcher used 48 AI agents to search NASA lidar data for Maya sites, then presented one location as potentially undocumented. Archaeologists pointed out that the site had already appeared in an academic journal, showing why AI-assisted searches still need expert review and careful credit.
No signals in this window yet — try a wider range.
A new evaluation suite was supposed to last a year. It lasted nine days. What that tells us about the pace — and the plateau everyone keeps predicting.
Buried in an annex: a threshold that quietly pulls half the industry into scope.
We stopped training bigger models and started training better questions.
Two gigawatts, one undisclosed buyer, and a power deal that rewrites the economics of frontier training.
Five minutes each morning. No hype, no noise — just the signal, before the rest of the internet catches up.