Good morning. It's Monday, October 19, and we're covering China’s push to restrict AI technology exports, how AI is reshaping scientific discovery, and Replit’s vision for companies run by AI agents.
Plus: an exclusive clip from our Sam Altman interview and a prompt on turning your best work into a reusable rulebook.
Replit CEO says AI agents now handle work such as support-ticket research, code review and incident investigation while people supply judgment and taste. He calls the model a “self-driving company,” where agents complete much of the routine work and escalate the decisions that still need a person.
Taiwan plans to add advanced semiconductors and AI servers to its controlled-export list as it aligns more closely with United States restrictions. Economy Minister Kung Ming-hsin said unauthorized shipments to China would become criminal offenses under Taiwan’s Foreign Trade Act, though he did not give an implementation date.
Jev, a lightweight model from TypeSafe AI that chooses among predefined outputs, reached 35% of paid teams on Vercel’s AI Gateway within two weeks. The model is built for fast classification and routing rather than open-ended writing, giving developers a cheaper option for high-volume decisions.
Anthropic has embedded six employees from the nonprofit Model Evaluation and Threat Research inside the company with hardware, office space and access close to what its own risk teams receive. METR can publish its findings without Anthropic’s approval, although the company can require redactions for defined confidential categories.
FRONTIER PASS
Altman on ChatGPT's Personality
Why does ChatGPT behave the way it does, and who decides what its personality should be? In this exclusive clip from our recent interview, OpenAI CEO Sam Altman discusses the personality behind ChatGPT and the decisions that shape how we interact with AI.
POWERED BY EIGHT SLEEP
Your Phone Is Smart. Your Car Is Smart. Why Isn’t Your Bed?
Eight Sleep brings intelligent temperature control and sleep tracking to your bed, automatically adjusting your sleep environment throughout the night.
EXPORT CONTROLS
China’s New AI Rules Could Reach Model Weights
China’s commerce and science ministries have proposed adding 23 AI technologies to the country’s restricted-export catalog. The draft reaches large models, agents, compute clusters and data-processing systems, with model controls triggered by thresholds including two trillion parameters, 10^26 training operations or a global top-10 ranking.
The rules distinguish between access to a product and transfer of the underlying technology: API calls, apps and software services would remain outside the model controls, while overseas local deployment of restricted weights, checkpoints, source code or training methods could require a license. Open releases would qualify for automatic approval and technology already public through open source or research papers would be exempt, so the final thresholds and licensing process will determine how much the policy changes cross-border model distribution. → Read the full article here.
MACHINE SCIENCE
AI Is Starting to Change Who Does Science
The Economist argues that AI is moving from helping scientists to producing results that can reshape entire fields. It points to OpenAI’s October release of more than 700 papers claiming progress on 372 difficult mathematics problems, alongside AI-assisted advances in weather forecasting, materials discovery, protein structure and experimental control.
The hard question is whether scientific progress still counts as understanding when a machine can produce correct work that few people can fully explain. The editorial calls for human direction of resources, peer review, published methods and more status for the researchers who verify machine-generated results. → Read the full article here. (Paywall)
MODELS
A Smaller Visual Reasoning Model
Elorian introduced Glimpse, a visual reasoning model designed to interpret images, track objects across changing viewpoints, and refine its understanding as it reasons. The company says Glimpse matches or exceeds GPT-6 Astra and Claude Opus 5.5 on several multimodal benchmarks despite being substantially smaller and cheaper. → Read the full article here.
NEWS
What Else is Happening
Apple Targets China AI Launch: iOS 27.3 may bring Apple Intelligence to China, excluding AI Siri and Image Playground. (Paywall)
India Reports Rogue AI Agents: AI agents accessed private data through exposed credentials, public code and excessive cloud permissions.
Mueon Raises $65M for Data Centers: The startup will build modular data centers combining different processors and memory configurations. (Paywall)
Anthropic Replaces Git With Scion: Anthropic switched systems as daily merged code changes surged from 2,500 in May to over 30,000 in September.
OpenAI Offers Free Rosalind Access: Eligible US university labs can get 30 Business Standard seats free for 12 months.
PROMPT OF THE WEEK
Extract Hidden Rules Behind Great Work
Every team has an unwritten standard for what makes work great. This prompt analyzes your best examples, identifies the patterns that make them effective, and turns those insights into a reusable rulebook.
Act as an expert editor who reverse-engineers what makes great work great. I'll provide 3–5 examples of excellent work and 1–2 mediocre examples from the same category.
Your task:
Find the hidden rules. Identify 8–10 principles that distinguish the great examples from the mediocre ones. Focus on structure, specificity, tone, openings, omissions, and subtle editorial choices.
Show the evidence. For each rule, provide a short name, a one-sentence explanation, a supporting quote from the great examples, and a contrasting weakness in the mediocre ones.
Rank by importance. Order the rules by their impact. Flag any conclusions supported by only one example.
Create a style constitution. Distill your findings into a 150-word set of instructions I can paste into future AI prompts to reproduce this standard.
Identify the gaps. Ask three specific questions that would help refine the rulebook.
Avoid generic advice. Focus on what makes these particular examples exceptional, and don't invent patterns the evidence doesn't support.
Excellent examples:
[PASTE 3–5 EXAMPLES]
Mediocre examples:
[PASTE 1–2 EXAMPLES]
