Good morning. It's Monday, May 18, and we're covering he first formal U.S.-China AI safety talks, a startup chasing self-improving AI systems, and ChatGPT’s new push into personal finance tools.

Know someone who'd love this? Forward it their way. First time reading? Sign up here to keep up with the future of tech.

Rather than ban AI-assisted writing, arXiv will suspend authors for one year if their preprints show "incontrovertible evidence" that they didn't check LLM output — hallucinated references or leftover comments to and from the chatbot — with all subsequent submissions then requiring acceptance at a peer-reviewed venue first. It's the first formal slop penalty from a major preprint platform, and lands as a May 2026 Lancet audit found fabricated citations in biomedical papers rose 10x since 2023 — from 1 in 2,828 papers to 1 in 277 in early 2026.

The rollout starts small: only U.S. Pro users can test ChatGPT’s new finance tools, despite more than 200 million monthly users already seeking budgeting and investing help. OpenAI says users can securely connect bank accounts, view spending dashboards, and ask GPT-5.5 questions grounded in their financial data and personal goals. The preview aims to refine privacy controls and real-world performance before a broader expansion.

Google’s upcoming “Gemini Spark” agent may handle inboxes, meeting prep, and scheduled tasks — while warning users it could share sensitive data or make purchases without approval. The feature, spotted in Google app beta 17.23 ahead of I/O 2026, replaces the earlier “Gemini Agent” branding inside the Gemini app. Spark reportedly uses chats, connected apps, location, and browsing context to personalize actions, with Google labeling the system “experimental.”

Some wealthy investors now see fish farms and tractor dealerships as safer bets than AI startups, betting physical assets will outlast software cycles. Equity Group Investments, backed by the late billionaire Sam Zell’s family, says it targets “old-economy” businesses with stable cash flow and lower AI disruption risk. Tax incentives, tariff uncertainty, and the rise of the “HALO” trade — heavy assets, low obsolescence — are driving interest in sectors like agriculture and auto dealerships.

Higgsfield is the world's first AI-powered streaming platform, dropping original series made entirely by artificial intelligence. No studios, no gatekeepers, no $200M budgets. Just stories that shouldn't exist yet, streaming right now for free.

Stream for Free.

A breakdown of Anthropic’s 2028 AI warning, covering the U.S.-China race, export controls, open-source AI, and self-improving models.

Treasury Secretary Scott Bessent said that the United States and China will begin discussions on artificial intelligence safety during or following the Trump-Xi summit in Beijing. The talks are expected to focus on guardrails for advanced AI systems, including protocols aimed at preventing powerful models from falling into the hands of nonstate actors such as hackers or terrorist groups.

Bessent did not provide a timeline or structure for the discussions, but described them as cooperation between the world’s “two AI superpowers.” The announcement comes as both countries accelerate AI development while facing growing concerns over security risks, biological weaponization, and the broader geopolitical race for technological dominance. → Read the full article here. (Paywall)

A new San Francisco startup called Recursive Superintelligence emerged from stealth with $650 million in funding and a team of prominent AI researchers led by Richard Socher, founder of You.com and former Salesforce chief scientist. The company aims to build a recursively self-improving AI system capable of identifying its own weaknesses and redesigning itself without human intervention — a concept researchers have pursued for years but not yet achieved.

Socher said the company’s approach relies on “open-endedness,” allowing AI systems to continuously evolve through autonomous ideation, implementation, and validation cycles. The startup also plans to commercialize its research, with Socher saying products could arrive “in quarters, not years,” even as the broader industry debates the risks and implications of increasingly autonomous AI systems. → Read the full article here.

Developer Hermann Björgvin created Clawdmeter, a handheld ESP32-powered companion device that monitors Claude Code token usage in real time using animated pixel-art feedback. Built on a Waveshare ESP32-S3 AMOLED development board, the device connects to a Linux host over Bluetooth while a background daemon polls Anthropic’s API every 60 seconds to retrieve session and weekly utilization data.

Usage levels are displayed through increasingly frantic “Clawd” animations, alongside detailed stats and reset timers accessible via touchscreen controls. The device also doubles as a Bluetooth HID controller, allowing its buttons to trigger Claude Code voice mode and interface shortcuts system wide. → Read the full article here.

Detroit Automakers Slash 20,000 Jobs: GM, Ford, and Stellantis cut 19% of U.S. salaried roles as AI and software-driven vehicles reshape hiring needs.

OpenAI Weighs Apple Lawsuit: OpenAI may pursue legal action over Apple’s ChatGPT integration, citing weak visibility and fewer subscriptions.

AI Robot Automates Tire Changes: Automated Tire’s SmartBay changes tires in 30 minutes without removing wheels, reducing labor needs in service bays.

Runway Bets AI Beyond Language: Runway says video-trained AI “world models” could outperform text-based systems in filmmaking and science.

NVIDIA CEO Pushes Skilled Trades: Jensen Huang says AI’s growth is boosting demand for electricians, welders, and technicians faster than desk jobs.

Use this when: you're preparing for a salary, contract, or vendor negotiation and need a clear strategy.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team