Revolutionizing AI: How Pocket-Sized Models Are Shaping the Future of Computing

To look back at this article, go to My Profile and then click on Saved stories.

Knight Will

Compact AI Innovations May Herald a Computing Revolution

The launch of ChatGPT in November 2023 marked a moment where it was initially available exclusively via cloud services, due to the sheer size of its underlying model.

Today, I'm operating an AI program of comparable ability on a Macbook Air, and it's not generating any notable heat. This reduction highlights the swift progress scientists are making in streamlining AI models, making them more compact and energy-efficient. It also illustrates that scaling up isn't the sole strategy for enhancing the intelligence of machines significantly.

My laptop is now powered by a clever AI model known as Phi-3-mini, which is designed to mimic the intelligence and humor found in ChatGPT. This model is among the latest series of scaled-down AI models unveiled by Microsoft's team. Despite its size, which is small enough to be operational on a mobile phone, I chose to experiment with it on my laptop while using an iPhone app named Enchanted for communication, which offers a chatting platform akin to the one found in the official ChatGPT app.

In a study discussing the Phi-3 series of models, Microsoft's team of scientists state that the model I employed compares positively with GPT-3.5, the OpenAI model that powered the initial version of ChatGPT. This assertion is grounded in evaluating its efficacy across various conventional AI tests aimed at assessing reasoning and general understanding. From my personal experiments, it appears to be equally proficient.

At its yearly developer event, Build, this week, Microsoft unveiled the innovative Phi-3 model, termed "multimodal," designed to process audio, video, and text. This announcement arrived shortly after OpenAI and Google each highlighted their groundbreaking AI assistants, which are grounded in multimodal models and available through the cloud.

Microsoft's diminutive series of AI technologies indicates that it's becoming feasible to develop a variety of useful AI applications that do not rely on cloud computing. This advancement could lead to the emergence of new applications by making them quicker to respond or more secure in terms of privacy. (An essential component of the newly introduced Recall feature by Microsoft, which employs AI to allow searches of any action ever performed on your PC, are these offline algorithms.) However, the Phi series also sheds light on the characteristics of contemporary AI and potentially how to enhance it. Sébastien Bubeck, a Microsoft researcher involved in the initiative, shared with me that the purpose behind constructing these models was to investigate if a more selective approach to the training data of an AI system could serve as a method to refine its capabilities.

Massive linguistic algorithms such as OpenAI's GPT-4 or Google's Gemini, which fuel chatbots and various applications, are generally nurtured with vast amounts of textual data extracted from books, web pages, and virtually any obtainable medium. While this practice has sparked legal debates, OpenAI and similar entities have discovered that enhancing the volume of data input into these systems, along with the computing resources allocated for their development, can lead to the emergence of novel functionalities.

Authored by Carlton

Authored by Emily Mullin

Authored by Andy

Authored by Scott Gilbertson

Bubeck, intrigued by the characteristics of "intelligence" displayed by language models, opted to explore whether meticulously selecting the data inputted into a model could enhance its performance without the need to significantly expand its training dataset.

In September last year, his group worked with a model that was approximately 6% the scale of OpenAI's GPT-3.5. They trained this smaller model on high-quality synthetic data, which was produced by a more significant AI model and contained information from various specialized fields, such as programming. The performance of this scaled-down model exceeded expectations. "To our astonishment, we found that our model outperformed GPT-3.5 in programming tasks through this method," he mentions. "This outcome was truly unexpected for us."

The team led by Bubeck at Microsoft has uncovered additional insights through this methodology. Their research found that by training a particularly small model with children's tales, it was capable of generating consistently logical responses. This is noteworthy because AI systems of similar scale often generate nonsensical results when trained using standard methods. This further indicates that AI applications, which appear to lack sufficient power, can be made effective by teaching them with appropriate content.

Bubeck suggests the findings imply that enhancing the intelligence of future AI technologies will involve more than merely increasing their size. Furthermore, it appears that smaller-scale models, like Phi-3, will play a crucial role in the evolution of computing. Operating AI models directly on devices like smartphones, laptops, or desktops minimizes delays or disruptions that can happen when data is processed in the cloud. This ensures data remains on the user's device and could pave the way for innovative AI applications that are not feasible with a cloud-based approach, including AI features that are seamlessly integrated into the device's operating system.

There is widespread anticipation that Apple will reveal its much-anticipated artificial intelligence plan at its upcoming WWDC conference. The company has often highlighted its unique combination of in-house hardware and software, which enables AI processing directly on its devices. Instead of competing directly with giants like OpenAI and Google in the development of vast cloud-based AI systems, Apple may adopt a unique approach by concentrating on miniaturizing AI technologies to be conveniently used on its consumer devices.

Recommended for You …

Delivered to your email: Dive into Will Knight's Fast Forward for the latest developments in artificial intelligence.

Indian electorate faces a deluge of deepfake content

Incarcerated individuals purchased tablets, only to discover an unfulfilled commitment

The single factor impeding the progress of heat pumps

Eternal Sunshine: The Top Sunglasses Picks for Any Adventure

Dave Paresh

Steven Levy

Lauren Goode

Name Unchanged

Knight Will

Lauren Goode

Rogers Reece

Knight Will

Additional Insights from WIRED

Evaluations and Manuals

© 2024 Condé Nast. All rights reserved. Purchases made through our website may yield commission for WIRED due to our Affiliate Agreements with retail partners. Content from this site must not be copied, disseminated, broadcast, stored, or used in any way without explicit consent from Condé Nast. Advertising Options

Choose a global website