Apple boosts spending on creating ChatGPT-like technology

Apple has boosted its budget for developing artificial intelligence, emphasizing creating conversational chatbot features for Siri — allegedly spending millions of dollars daily on research and development.

In May, it was learned that Apple had been recruiting more engineers to work on generative AI projects. While the company didn't make an official forward-looking statement, CEO Tim Cook said that generative AI is "very interesting."

But, the story of Apple's foray into generative AI starts much earlier than May. Four years ago, Apple's head of AI, John Giannandrea, formed a team to work on large-language models (LLMs), the basis of generative AI chatbots like ChatGPT.

Apple's conversational AI team, the Foundational Models team, is led by Ruoming Pang, who previously worked at Google for 15 years. The team has a significant budget, and trains advanced LLMs using millions of dollars daily. Despite having only 16 members, their advancements rival those of OpenAI, which spent over $100 million to train a similar LLM.

According to The Information, at least two other teams at Apple are working on language and image models. One group focuses on Visual Intelligence, generating images, videos, and 3D scenes, while another works on multimodal AI, which can handle text, images, and videos.

Currently, Apple is planning to integrate LLMs into Siri, its voice assistant. This would allow users to automate complex tasks using natural language, similar to Google's efforts to improve their voice assistant. Apple believes its advanced language model, Ajax GPT, is better than OpenAI's GPT 3.5.

Ultimately, incorporating LLMs into Apple products has its challenges. Unlike some competitors who use a cloud-based approach, Apple prefers running software on-device for better privacy and performance. However, Apple's LLMs, including Ajax GPT, are quite large, which makes it difficult to fit them onto the iPhone due to their size and complexity.

There are precedents for shrinking large models, such as Google's PaLM2, which comes in different sizes, including one suitable for devices and offline use. While it's unclear what Apple's plans are, the company could opt for smaller LLMs for privacy reasons.

In May, Internal documents and anonymous sources leaked details of Apple's internal ban on ChatGPT-like technology and the plans for its own LLM.

Latest Exclusives

Latest comparisons

13 Comments

hexclock 10 Years · 1316 comments

About 1 year ago

I read a couple of articles about AJAX. The first one said it is run on Google cloud and built with Google’s JAX frame work.

The other said that AJAX is Apple proprietary. Which is correct? Or can they both be correct at the same time?

Also, does this mean if this is ever implemented into Siri or other Apple software, iPhone users will have their data routed through Google whether we like it or not?

And why would Google let Apple develop a competing technology on their own servers? Unless that’s where the millions of dollars are going.

mpantone 18 Years · 2254 comments

hexclock said:

I read a couple of articles about AJAX. The first one said it is run on Google cloud and built with Google’s JAX frame work.
The other said that AJAX is Apple proprietary. Which is correct? Or can they both be correct at the same time?
Also, does this mean if this is ever implemented into Siri or other Apple software, iPhone users will have their data routed through Google whether we like it or not?
And why would Google let Apple develop a competing technology on their own servers? Unless that’s where the millions of dollars are going.

Apple's generative AI activities have not been disclosed by them (unsurprisingly). So anything you read about Apple's efforts are entirely speculative.

At this point, it's pointless to guess how data will be routed. That said, Apple's commitment to user privacy is very, Very, VERY strong. iCloud, iMessage, etc. My best guess is that Apple's production service will run on their own servers in their own datacenters. Remember that datacenters aren't free. You need to put hardware in them, get people to run them, etc.

Over time, Apple has made a concerted effort to wean themselves from Google's grasp. It didn't happen overnight either.

In any case Apple will never disclose the development process to the public.

mayfly 1 Year · 385 comments

I want to see Apple develop an AI powered music generator. It would be great to see it write better songs than 99% of the crap being performed by Disney "Pop Tarts" and Boy Bands using auto-tune. And who or what couldn't write better tune than a revenge song about ex-boyfriends like Taylor Swift? (Who are these tools willing to hook up with her, anyway?)

mayfly said:

I want to see Apple develop an AI powered music generator. It would be great to see it write better songs than 99% of the crap being performed by Disney "Pop Tarts" and Boy Bands using auto-tune. And who or what couldn't write better tune than a revenge song about ex-boyfriends like Taylor Swift? (Who are these tools willing to hook up with her, anyway?)

Remember Microsoft Songsmith? That was good for laugh.
Unfortunately, the vast majority of music listeners are perfectly fine with simplistic melodies, repetitive hooks, and third grade lyrics in 4/4 time signature. Still, there’s plenty of well thought out music out there to find.

FileMakerFeller 6 Years · 1561 comments

hexclock said:

I read a couple of articles about AJAX. The first one said it is run on Google cloud and built with Google’s JAX frame work.
The other said that AJAX is Apple proprietary. Which is correct? Or can they both be correct at the same time?
Also, does this mean if this is ever implemented into Siri or other Apple software, iPhone users will have their data routed through Google whether we like it or not?
And why would Google let Apple develop a competing technology on their own servers? Unless that’s where the millions of dollars are going.

Apple rents compute and storage capacity from multiple cloud vendors including Google, Amazon (AWS) and Microsoft (Azure). That's where the overwhelming majority of the project's money is being spent - building a large language model based on data from a billion+ devices does not come cheap.

But that doesn't mean the data is being "routed through Google" - it's a private cloud with built-in encryption; the cloud vendor does not have access to the encryption keys and cannot access the plaintext version of the data.

JAX is an open source framework, I would guess that "Apple JAX" involves some extension code that is customised based on observations and analysis of the existing data set and perhaps knowledge of the devices on which (portions of) the code will be run in the future.