A pretraining researcher who worked at both OpenAI and Anthropic resigned, warning both labs are recklessly racing toward self improving superintelligence that could kill everyone by decade’s end, calling for pacing agreements or a capability freeze
Apple introduced Apple Reference Image, a tool to help users determine if photos have been edited, including AI alterations
OpenAI may pause new Pro subscriptions as demand for Astra surges to unprecedented levels
Firecrawl launched pay-as-you-go for paid plans, with automatic $5 top-ups and user-controlled monthly spending caps.
Salesforce in talks to acquire Listen Labs, an AI-powered customer research platform, for $2B, up from a $500M valuation earlier this year.
Clay raised $115M led by Wellington Management at a $7.1B valuation, up from $3.1B when it raised $100M in August 2025.
Paul Christiano is joining the OpenAI Foundation Board and its Safety and Security Committee, and will also serve as a non-voting observer on the OpenAI Group PBC Board.
Grok Bot added the ability to draft messages inline for user approval before sending.
Lovable launched a Partner Program letting its 800,000+ builders get certified and reach new clients through a partner directory.
Suno lunched v6 and v6-wild models, built with artists and producers v6 for precision, v6-wild for more experimental output.
Meta is acquiring Stilla, a multiplayer AI teammate for companies, to strengthen its business AI products.
ChatGPT Voice Mode now supports GPT 5.6 Sol and GPT 6 Astra, with Astra limited to Pro accounts.
Straitly launched an inference gateway offering 5% cashback on closed-source model usage and 10% on open-source, with access to 177 models and automatic provider failover.
Lightfield raised a $47M Series A led by a16z to build a CRM alternative modeled as a business world model; 5,000+ companies have signed up since its November launch.
Anthropic built an economic scenario explorer projecting AI could push US GDP to $44.4T by 2030 in an extreme scenario, alongside significantly higher unemployment among knowledge workers
The Lightfield item is the one I’d keep watching. If agents are going to do meaningful customer-facing work, a CRM can’t remain just rows, fields and workflow triggers, it has to expose enough context to model how the business actually operates. The interesting investment question is whether incumbents can evolve their system-of-record advantage quickly enough, or whether an AI-native context layer becomes a new control point above them.
The juxtaposition in this edition is worth sitting with. Anthropic released an economic scenario model projecting GDP 8.6% above a no-AI baseline by 2030 in its "substantial" scenario — and in the same week, one of their alignment researchers resigned citing existential risk. Five days from now, by September 15, Anthropic and every other GPAI provider above the 10^25 FLOPs training threshold must submit formal systemic risk evaluations to the European AI Office. The economic model and the risk filing go to very different audiences in the same week.