The Age of Pervasive Intelligence

A diverse ecosystem of commodity AI models, embedded as AI features across applications and augmenting devices in an AI of Things will lead to Pervasive Intelligence.

Figure 1. AI Art - Ant Intelligence, an ecosystem of many small intelligences. Will pervasive embedded AI be like that?

The New Wave of AI

As AI keeps developing, each turn in the road has offered surprises that forces us to update our understanding of what AI can be. This month’s releases of OpenAI’s o1, Microsoft Copilot Wave 2, efficient new GPT-4-level smaller LLMs, and video generation AI model APIs are a part of a new wave.

This new wave is a sign of AI maturing as a technology, as well as a confirmation of trends we’ve seen in AI development: Embedding AI features across software applications; better and more efficient smaller LLMs; AI as an API. All of these add up to a wave of Pervasive Intelligence.

Copilot's Second Wave

With so much released last week, we only briefly mentioned the Microsoft Copilot Wave 2 announcement, which introduced embedded AI as features in Microsoft Excel, PowerPoint, and Outlook.

I asked Copilot to remind me who developed the first spreadsheet for PCs:

VisiCalc was the trailblazer—the one that kicked off the whole spreadsheet revolution. Developed by Dan Bricklin and Bob Frankston, it made its debut in 1979. 

The spreadsheet application offloaded the drudgery of numerical calculations (adding, averaging, etc.) in financial and other kinds of data analysis. Now Microsoft is embedding AI in their Excel spreadsheet application, giving it Python programming powers and enabling AI-driven data analysis. The ‘AI data scientist’ is now an Excel feature.

Figure 2. Excel with copilot, you can chat, have Copilot write Python data analysis functions, visualize results, and more.

It’s the same theme across their applications. Their Copilot-based “narrative builder” generative AI feature helps in slide deck ideation and creation in PowerPoint. Outlook uses Copilot AI to manage your inbox. They introduced a Pages feature for AI and team collaboration around information, as well as Copilot Agents to automate workflows.

Copilot Wave 2 presents AI as an embedded feature across applications, working in the background. What’s striking about their announcement is how practical and mundane it is, rather than how “new” or viral. It’s a sign that AI is starting to mature, ‘settle down,’ and solve real-world tasks.

The AI Blind Spot

Recently, Peter Theil on the Joe Rogan Experience gave his view on AI. He mentioned these perspectives that were dominant in the 2010s AI debates:

• Nick Bostrom’s Superintelligence thesis, envisioning AI as developing into a God-like Superintelligence that could challenge human dominance.

• The vision expressed by Kai-Fu Lee in The AI Superpowers, that treats AI as a pre-transformers machine intelligence tool of Great Powers, a struggle of China versus Silicon Valley. In this scenario, China wins because their government has no ethical qualms about exploiting AI as a surveillance-state tool for measuring and controlling their population.

The Turing Test, suggested by Alan Turing in 1950, is a long-standing traditional definition of AI. As Theil describes it, that concept got lost in the 2010s discussions on AI, but it came back when ChatGPT passed the Turing test. ChatGPT was the pivotal moment that validated the view that AI is about getting machines to think like humans.

AI can now generate ("write your AI novel") and repurpose (Google NotebookLM podcasts) content in text, audio, and images, often well enough to pass for human creation.  However, we’ve noted that the Turing Test is itself flawed and incomplete, noting that “the Turing Test may be less about AI intelligence, and more about how competent AI is in deception.” Moreover, taking comparisons between human and machine intelligence too literally leads to the Anthropomorphic error.

Sci-Fi has treated AI as a singular Oracle super-mind run on “Big Iron” computers, in movies such as 2001: A Space Odyssey, Wargames, and I, Robot. In “I, Robot” in particular, the massive AI brain is smart and powerful enough to enlist an army of robots in an AI-led revolution. Shades of AI Superintelligence doomsday scenarios.

AI skeptics, AI doomers, and AI optimists each have their perspectives, priors, and different visions of AI. While all perspectives have validity, we are like blind men feeling a part of an elephant and declaring the animal based on limited experience. We don’t know enough to be sure, and it is in the blind spots where we end up missing the whole story of AI.

One blind spot many have is they miss the value of small, embedded AI models everywhere.

Pervasive AI Models

“AI is the new electricity.” – Andrew Ng

Our belief that “AI Changes Everything”, as our masthead says, is not about AI replacing humans in jobs - it will displace (not replace) most jobs -  or AI becoming a huge AI Oracle super-mind overlord. Although AGI and the Singularity are inevitable, AI won’t have to get to Superintelligence or even AGI to change the world.

I said in March 2023 that “Current state-of-the-art AI will change the world,” because GPT-4-level AI was powerful enough to be used across a huge swath of useful tasks.  AI changes everything because it will be everywhere, interacting with all aspects of our work, entertainment, and personal lives.

Now GPT-4-level AI has come to your desktop. AI releases last week delivered GPT-4-level open-source AI models you can run locally:

• Mistral released 22B Mistral Small.

• Qwen released a family of AI models that are all SOTA for their size. Qwen 2.5 32B, which gets a score of 83 on MMLU, 49 on GPQA, and 88 on HumanEval outperforms GPT-4o-mini.

• The Qwen 2.5-Coder 7B model is a top-notch coding model that outperforms GPT-4o and o1-preview on Livebench coding, as shown in its Technical Report.

Alibaba and Mistral have delivered yet more great AI models to choose from. The story this tells is that AI is not defined by one massive AI model, but by a diverse ecosystem of many AI models, some with specialized purposes and roles, and some small enough to run locally.

Embedded Intelligence and the AI of Things

One trend we saw with the internet was how it began on the desktop, tethered to a modem and wire, then extended to the laptop and wireless, then spread to the smartphone, and finally became a fabric with the Internet of Things.

AI will follow a similar trend, but it may happen more quickly. AI was initially exclusive to a few OpenAI APIs and playgrounds, then it spread to other endpoints. Now you can download GPT-4-level AI models to your desktop, and Apple can run AI models on your phone.

You don’t even have to download your own AI mode. GPT-4o and even o1-level AI is on ChatGPT phone app. Most AI models, from LLMs to video generation AI models, are accessible via API.

Offering AGI as an API means that a creative software developer can embed intelligence into any piece of software. It leads to the AI of Things, which like the Internet of Thing will be a pervasive fabric of intelligence embedded into all manner of tools, devices, and products.

Any product that has embedded internet connectivity could also have embedded intelligence, via on-board AI or API-connected AI. It could mean fridges that tell us when we run low on milk and security cameras that report on intruders in real-time. It will mean autonomous vehicles, from cars and trucks to golf carts and lawnmowers. “Smart” products will get smarter.

The Shape of the Intelligence Age

Technology brought us from the Stone Age to the Agricultural Age and then to the Industrial Age. From here, the path to the Intelligence Age is paved with compute, energy, and human will. – Sam Altman, in “The Intelligence Age”

Sam Altman published an essay titled “The Intelligence Age,” sharing his latest thoughts on AI. It’s a brief, high-level essay, which places the AI revolution as the start of a new age in history, the Intelligence Age. That puts AI on par with the invention of agriculture and the steam engine.

We agree with his broad stroke description, including his assertions that AGI is upon us in possibly the next 1000 days (that’s 3 years) and that “the upside is so tremendous that we owe it to ourselves, and the future, to figure out how to navigate the risks in front of us.”

The goal of “AI Changes Everything” is to describe and explain the shape of this AI revolution we are in, to help you navigate those challenges and risks. Here’s what the flood of new AI releases and fast AI progress is telling us:

• AI is becoming an embedded feature across most software applications.

• AI is defined not by one massive AI model, but by a diverse ecosystem of many AI models, many with specialized purposes and roles.

• LLMs are getting commoditized, as Satya Nadella stated in his Copilot Wave 2 intro. LLMs are becoming cheap and pervasive.

• A consequence of commoditized LLMs is that the value is flowing upwards to the AI-first applications and agents that pull AI models together.

• AI will be on tap via APIs that will make it available and embedded across many devices – the AI of Things.

All of this adds up to Pervasive Intelligence. Pervasive Intelligence will change the way we live in hundreds of small ways that will add up to a new world of possibility in the Intelligence Age. AI is like electricity, as Andrew Ng says - a utility, a service, and a feature embedded into our digital lives.

One last point: Technology adoption curves for AI are faster than any prior technology. Still, there is a gap from research to production AI model to embedded AI application, and then to widespread productive use of it. The ripple effects to our economy and culture will not play out instantly but may take many years. We are still in the early innings of the AI revolution.