LLMs, AI Expansion in Apple’s Ecosystem, and the New M4 Processor in Macs

Hi everyone! Today, we’ll discuss large language models and their reasoning limitations, Apple’s expanding AI ecosystem, and the new M4 chip. I’m Roman Mishchenko, an engineer at CleanMyMac, and let’s dive into the latest updates in iOS and macOS development!

M4 Macs

Apple has unveiled a new generation of Macs featuring the updated M4 processor lineup. Performance has improved, though Apple still compares these chips to the M1 lineup. This makes sense, as most users upgrade their devices every few years, so comparisons to earlier models are logical.

The new M4 processors power the iMac, MacBook Pro, and Mac mini. While the first two have only received hardware updates, the Mac mini also features a brand-new design. It’s now even smaller, resembling a tiny box while retaining the full power of the new chip. Interestingly, the SSD in the new Mac mini can be replaced! However, since the memory controller chip is embedded in the M4 processor, only Apple-manufactured SSDs will be compatible. Furthermore, the SSDs in the M4 and M4 Pro versions of the Mac mini differ and are not interchangeable.

The starting price for the Mac mini is just $599, including 16 GB of RAM, gigabit Ethernet, and 256 GB of SSD storage.

Users have already begun testing the new Mac mini, and one of the most exciting applications is clustering multiple machines for machine learning. It’ll be interesting to see benchmarks and comparisons with Nvidia solutions in the same price range.

GitHub Copilot for Xcode

At the GitHub Universe conference, the company announced a significant update for iOS/macOS developers: Copilot for Xcode.

GitHub Copilot is an AI-powered programming assistant that helps write code. The Xcode extension integrates directly into the IDE, offering real-time code suggestions as you type. This extension builds upon an existing project by Intitni, which has been around for over a year and received rave reviews.

The new version remains open-source under the MIT license, but full functionality requires a GitHub Copilot subscription.

It’s worth mentioning that Apple also introduced its own coding assistant this summer—Xcode Predictive Code Completion Model—with Xcode 16 and macOS Sequoia. It will be fascinating to see how these tools compete.

LLM Can’t Reason

Models like GPT-4, LLaMA, and Mistral have shown impressive capabilities in natural language processing, Q&A, and creative tasks. However, whether they truly possess reasoning skills remains a hot topic in AI research.

Six Apple scientists recently explored this question, analyzing whether improved scores on the GSM8K benchmark reflect better reasoning abilities. Their findings revealed the fragility of mathematical reasoning in LLMs. As the complexity of questions increases, performance drops significantly.

They concluded that current LLMs don’t engage in true logical reasoning. Instead, they replicate observed reasoning patterns from training data, potentially due to benchmark test data leaking into training datasets.

For example, adding a seemingly relevant sentence to a question caused performance to drop by up to 65%, even though it didn’t affect the required reasoning steps. Similarly, altering question details also degraded model accuracy.

This highlights the limitations of current LLM technology and its gap from achieving General Artificial Intelligence. Check out the full study via the link in the description.

Pixelmator Acquisition by Apple

Apple has acquired Pixelmator, the company behind popular image-editing apps like Pixelmator Pro, Pixelmator for iOS, and Photomator.

In their announcement, the Pixelmator team expressed excitement about joining Apple, saying they’ve been inspired by the company since day one. They believe this acquisition will help them reach a broader audience and have a greater impact on creative professionals worldwide.

For existing Pixelmator users, the company assured that no significant changes to the current apps are planned for now.

This acquisition strengthens Apple’s ecosystem by integrating advanced image-editing tools, creating stronger competition for Adobe. Additionally, some Pixelmator features might become APIs, similar to how Apple integrated Shazam after acquiring it.

That’s all for now! Don’t forget to subscribe and stay tuned for the next digest. See you soon!