Source: Apple's budget for training AI is now millions of dollars per day; its conversational AI team Foundational Models has ~16 people led by Ruoming Pang
here's how (report) Asma Hussain / iThinkDifferent : Apple to make Siri smarter using language models in iOS 18 Ed Hardy / Cult of Mac : Apple's conversational AI efforts began 4 years ago X: @wintermoat : Apple touting the accomplishment of hiring some Google engineers. https://www.theinformation.com/ ... [image] Wayne Ma / @waynema : https://www.theinformation.com/ ... Some Apple employees believe its most advanced large language model is better than GPT-3.5. The 16-person team that trained it also quietly open sourced the tools they developed to train these models in July. Check out my latest scoop for @theinformation Forums: MacRumors Forums : iOS 18 Rumored to Feature Major Siri Improvements
Context & Ripple Effects
Apple’s conversational-AI work reportedly began years before later evidence of Ajax model testing for Siri and Messages, indicating an internal model program rather than a newly assembled response to the market.
The reported training spend and compact Foundational Models team help explain why generative AI later became a tent-pole Apple project and why the company paired its own work with planned ChatGPT access in Siri and other apps.
First-order effects
- Apple is committing substantial recurring compute spending to train proprietary models while concentrating the work in a small team led by Ruoming Pang.
- The team’s open-sourced training tools can broaden external use of Apple-developed infrastructure, even as Apple keeps its most advanced model work in-house.
Second-order effects
- A stronger Apple model program raises the pressure on Google and other mobile-platform rivals to match AI capabilities where users interact with apps and assistants.
- High daily training costs make model quality and deployment efficiency central to whether Apple can turn its installed-base reach into a durable AI advantage rather than rely chiefly on outside providers.
Third-order effects
- The story points to a split between companies that can fund proprietary foundation-model development and those that must differentiate through distribution, product integration, or partnerships.
- If this pattern persists, AI competition in consumer devices will be shaped less by a standalone chatbot and more by who can embed models into native workflows at sustainable cost.
The trend: Consumer-platform companies are moving from experimental AI teams toward costly, vertically integrated model programs tied to their operating-system distribution.