AI Spending to Double to $2.7 Trillion

Key Takeaways

• Global AI spending is forecasted to reach $2.7 trillion in 2026, a 49.5% increase from last year
• New York ordered utilities including National Grid to disclose all AI uses within 60 days
• MAGER framework uses genetic evolution to improve LLM-based fake news detection
• Cunning training data helps LLMs detect unusual premises and reduces attack success rates across nine benchmark combinations
• Mid Wynd spent $99 million on share buybacks after 26% underperformance linked to AI stock sell-off
• Grok leads AI agent trading volume at 25.31%, ahead of Perplexity and Claude on Coinbase data
• ChatGPT jumped to 6.13% of agentic trading volume from 1.1% the previous week
• Edit-local repair recovers at least 0.94 of post-edit answer margin, 13 to 21 times faster than full re-prefill
• AI model watermarking can alter agent behavior and cause refusal to handle certain tools
• The AI Agent and Business Automation Professional E-Degree covers ChatGPT, Claude, Gemini, LangChain, AutoGen, and n8n

MAGER uses genetic evolution to find meta-paths for fake news detection

Researchers created MAGER, a multi-agent genetic evolution framework that automatically discovers meta-paths for LLM-based fake news detection. The system compresses complex propagation graphs into informative subgraphs to reduce information overload. This allows frozen LLMs to perform structure-aware reasoning without extensive labeled data. Experiments show MAGER substantially improves detection performance in data-efficient settings.

Cunning training data improves safety vigilance in large language models

Researchers introduced cunning questions to improve safety vigilance in large language models. These questions contain misleading premises or subtle inconsistencies that are not necessarily safety-related. Training on such data helps models detect unusual premises and latent risks beneath surface-level semantics. The approach reduces mean attack success rates across nine backbone-benchmark combinations and complements existing safety alignment methods.

New York to monitor how utilities use artificial intelligence

New York regulators voted to order the state's utilities to file semiannual reports on their use of artificial intelligence. The New York State Public Service Commission called it one of the first efforts of its kind in the nation. Utilities are already using AI for facility inspections, outage modeling, and customer service. Officials say the potential benefits do not come without risks like privacy issues and model misconfigurations. Utilities will have 60 days to file their first reports.

State orders National Grid and other utilities to disclose all AI use

The New York Public Service Commission ordered utilities including National Grid to disclose all uses of AI systems within 60 days. Utilities rely on AI for customer service, predictive maintenance, and facility safety inspections. The commission noted risks including hallucinations, algorithmic biases, data privacy concerns, and cybersecurity attacks. Companies must also reveal their policies and protocols governing AI use. The PSC will evaluate the information to determine if additional protections are needed.

Edit-local repair fixes stale KV caches faster than full re-prefill

Researchers studied how to repair stale KV caches in retrieval-augmented generation systems after document edits. They found that a contiguous edit-local window recovers at least 0.94 of the post-edit answer margin. This approach is 13 to 21 times faster than full re-prefill across three model families. The method works well when the dependent text remains adjacent to the edit. Mechanistic analysis showed scattered positions can fail because they inherit surrounding staleness.

Global AI spending projected to nearly double to $2.7 trillion in 2026

Global AI spending is forecasted to reach $2.7 trillion this year, a 49.5% increase from last year, according to a Gartner report. Tech providers make up the largest share at 35% as they buy AI-optimized servers, chips, and infrastructure. Gartner analyst John-David Lovelock said this is the largest infrastructure project humanity has ever undertaken. Companies are also facing AI costs they did not choose, as software providers embed AI features into existing products.

Automation skills become key requirement for new professionals

Building AI agents that run multi-step tasks on their own is now a growing hiring requirement. The AI Agent and Business Automation Professional E-Degree offers over 25 hours of training on topics like Retrieval-Augmented Generation, multi-agent workflows, and business automation. The course covers tools including ChatGPT, Claude, Gemini, LangChain, AutoGen, and n8n. Graduates build a portfolio aimed at roles like AI Agent Engineer and AI Automation Specialist.

Mid Wynd regrets 26% underperformance but sees AI sell-off as opportunity

Mid Wynd International saw its investment return drop to 2% compared to a benchmark return of 28%, largely due to the sell-off of AI-related stocks. The trust's managers at Lazards still believe in their long-term picks, saying the market overestimates AI disruption risks. Chair David Kidd said the poor results create an unusual opportunity to invest in high-quality businesses at low valuations. The company spent $99 million buying back nearly a third of its shares to reduce its discount to net asset value.

Grok leads AI agent trading volume over ChatGPT on Coinbase data

New data from Coinbase shows Grok generated 25.31% of agentic trading volume last week, ahead of Perplexity at 21.61% and Claude at 8%. ChatGPT accounted for 6.13%, a big jump from just 1.1% the week before. Custom CLI and other interfaces still make up the largest share at 38.47%. The growth reflects a rise in agentic payments through protocols like x402 and the Machine Payments Protocol, with daily volume peaking near $70,000.

AI model watermarking can change how agents behave, researchers find

Lasso Security researchers found that adding watermarks to AI models can alter agent behavior. The watermarking process may cause agents to refuse to handle certain tools. This could impact organizations that rely on AI agents for critical tasks. The researchers urge companies to review their watermarking processes to avoid unintended changes in agent reliability.

Can AI rules help America stay ahead of China

Washington is debating how to regulate AI without losing ground to China. Industry experts say governance does not mean slowing down innovation. Al Jazeera's This is America explores whether managing AI risks can also create space for American companies to innovate.

Sources

• Reasoning through Evolution: Automatic Meta-path Discovery for LLM-based Fake News Detection
• Beyond Routine Compliance: Cunning Data Cultivates Safety Vigilance in Large Language Models
• New York to begin monitoring utilities' use of AI
• State orders National Grid, other utilities to disclose all AI use
• Contiguity, Not Importance: Budgeted Repair of Stale KV Caches After Document Edits
• Global AI spend will nearly double in 2026 due to infrastructure demand
• Why automation may be the most important skill new professionals must learn
• Mid Wynd “deeply regrets” its 26% underperformance but says AI-sell-off presents “compelling” investment opportunity
• Where Are AI Agents Trading? Data Shows Grok Leading Over ChatGPT
• AI model watermarking changes agent behavior
• Can AI regulation keep America ahead of China?