Apple's LLM reportedly outperforms GPT-4, but Sam Altman already admitted it "kind of sucks" as OpenAI gets ready to unveil a new model that's "really good, like materially better"

Apple's ReALM beats OpenAI's GPT-4 model despite using fewer parameters.

What you need to know

• Apple is reportedly developing a LLM dubbed Reference Resolution As Language Modeling (ReALM) that will enhance Siri's capabilities and allow it to understand context in a conversation and even process onscreen content.
• According to benchmarks shared by researchers, Apple's smallest model shares similar performance stats as OpenAI's GPT-4, whereas its largest models outperform it entirely.
• It's unclear when the models will ship, whether with the iOS 18 update or new lineup of iPhones.

Apple has been relatively quiet and rather slow hopping onto the AI bandwagon compared to its competitors like Microsoft, which was quick to make a multi-billion dollar investment in the technology further strengthening its ties with OpenAI. This has proven to be beneficial as Microsoft is the world's most valuable company with over $3 trillion in market capitalization. 13 market analysts unanimously voted that Microsoft will be the world's most valuable company in the next five years ahead of Apple, owing to its early lead and commitment to AI.

Apple's iPhone sales have taken a major hit, specifically in China as other competitors are starting to dominate the market. As you might be aware, Apple is slated to hold WWDC later this year in June. It's expected to make a host of announcements revolving around its new lineup of phones, We also expect the company to make several AI-related announcements. 

And as it now seems, we might be getting an early look at what the company intends to present to the world in the next few months. While details remain slim, Apple AI researchers recently published a research paper highlighting the company's plans to supercharge Siri with nifty AI capabilities. 

Is ReALm better than OpenAI's GPT-4?

(Image credit: Daniel Rubino)

The research paper is based on a Reference Resolution As Language Modeling (ReALM). It seems that Apple is looking to build on Siri's capabilities with this new model, as it'll enhance its capability to understand context in a conversation and even process onscreen content. 

Additionally, ReaLM will reportedly ship with the capability of converting conversational, onscreen, and more into a text-format that can be processed by large language models (LLMs). Per the research paper, the ReALM model will be available in four sizes: ReALM-80M, ReALM-250M, ReALM-1B, and ReALM-3B. The “M” and “B” at the end of the ReALM models represent the number of parameters in millions and billions respectively.

Per the research paper, the researchers ran benchmarks of Apple's new AI tools against OpenAI's GPT-3.5 and GPT-4 models. Based on their findings, Apple's smallest AI model showcased a similar performance when compared to OpenAI's GPT-4.

According to the researchers:

“We demonstrate large improvements over an existing system with similar functionality across different types of references, with our smallest model obtaining absolute gains of over 5% for onscreen references. We also benchmark against GPT-3.5 and GPT-4, with our smallest model achieving performance comparable to that of GPT-4, and our larger models substantially outperforming it.”

The researchers indicated its ReaLM outperforms OpenAI's GPT-4 model despite having “fewer parameters.” Rumors indicate GPT-4 has 1.76 trillion parameters. Even Sam Altman recently admitted GPT-4 “kind of sucks” with rumors looming in the air indicating OpenAI is getting ready to release a new model that will reportedly make ChatGPT “really good, like materially better.”

It's unclear if these models will ship with Apple's iOS 18 update or with the soon-to-launch lineup of iPhones.

Source