AI Safety Pact or Cartel? What the Slowdown Really Means
September 15, 2026The AI industry is doing two contradictory things at once this week: building faster and promising to slow down. The contradiction isn't just cosmetic — it suggests that nobody at the top of this industry actually believes safety and competitive advantage can be separated.
The top AI labs just agreed to pump the brakes. Or did they?
Sam Altman, Dario Amodei, Demis Hassabis, and Elon Musk reached a loose agreement over the weekend to slow the pace of frontier AI development — "frontier" meaning the most powerful, cutting-edge models these companies are racing to build. The stated reason is safety. Four companies that collectively dominate the market just found a polite way to limit competition — and the fact that it might also be a genuine safety measure doesn't make the antitrust problem disappear. When the biggest players in any industry agree to slow down together, consumers and regulators should ask who benefits. The structure of this deal — voluntary, informal, among direct competitors — looks a lot like the kind of coordination that draws regulatory attention regardless of intent. My read: safety concern and market protection aren't equally weighted here. The timing is too convenient for the incumbents.
Salesforce built a sales AI on top of Nvidia's free model. The big labs should pay attention.
Salesforce and Nvidia released Koa, an AI model trained specifically for sales, marketing, and customer support work, built on Nvidia's Nemotron — an "open-weight" model, meaning the underlying code is freely available for anyone to use and modify. Salesforce didn't need OpenAI or Anthropic to build something powerful and task-specific. That's the story. If companies can take free, capable base models and train them to do exactly what their business needs, OpenAI's $20-per-seat subscription starts looking like a rental fee for something you could own. The awkward chatbot grafted onto your CRM is the old model — what replaces it is purpose-built, trained on your industry's data, and doesn't require cutting a check to San Francisco every month.
Paying for premium AI gets you four months of advantage. Then the free models catch up.
Mozilla released a report, previewed by Ars Technica, finding that cheaper open models — many of them Chinese — consistently close the capability gap with Silicon Valley's most expensive frontier models within about four months, at roughly one-fifth the cost. "Capability gap" means the difference in how well a model performs on real tasks like writing, reasoning, and coding. For businesses that have justified premium AI subscriptions on the grounds that they're getting meaningfully better output: you are paying for a head start that disappears before your next budget cycle. The question worth asking before your next renewal isn't whether the AI is good — it's whether it's four months of good.
A startup thinks it can insure you against your AI agent going rogue
A new company called AIUC — Artificial Intelligence Underwriting Company — raised $40 million to build what amounts to liability coverage for AI agents. An "AI agent" is an AI system that doesn't just answer questions but takes actions on your behalf: booking things, sending emails, executing transactions, making decisions inside software. The founders, from Anthropic and METR (an AI safety research organization), are betting that as more businesses deploy agents to do real work, the question of who's responsible when one does something wrong becomes legally and financially urgent. Boardrooms are still asking whether the AI works. The more important question — what happens when it doesn't, and who pays — is arriving faster than most legal and finance teams are prepared for.
The trillion-dollar AI buildout might be a bubble. Or it might not. That's actually the problem.
MIT Technology Review worked with Wharton finance professor Jessica Wachter to assess the economic risk of the current AI infrastructure boom — the massive spending on data centers, chips, and power that is happening right now on the assumption that AI will generate returns to justify it. Her conclusion is not that it's definitely a bubble, but that the uncertainty itself is the risk: unlike previous tech investments, the range of possible outcomes here is so wide that normal financial modeling breaks down. For anyone whose organization is making significant AI investment decisions, this is worth sitting with. The honest answer to "will this pay off?" is not yes or no — it's that nobody has a reliable framework for knowing, and the people spending the most money are also the most motivated to project confidence.
https://www.technologyreview.com/2026/09/15/1144028/ai-infrastructure-boom-investment-bubble-risk/
The most clarifying question right now isn't whether AI will change your industry — it's whether the companies selling you AI actually know what it's worth.
Stop staring at the blank ChatGPT box
The Plain-English AI Prompt Pack: 100 copy-paste prompts for real work — emails, meetings, marketing, managing people. Works with any chatbot.
Get the Prompt Pack →Get the daily brief
Plain-English AI news for people with real jobs. Free, three minutes a day.