Happy 40th Birthday, Macintosh. Here’s to the next 40.

This week marked the 40th anniversary of the Macintosh, an innovation that changed my life. I had spent the previous two years learning to code Basic, Pascal, Logo, and Fortran (yep, you read that correctly) on an Apple II. I loved the ability to create minimal interactivity and especially graphics, crude though they were. The Mac changed everything.

The Mac allowed us to see and create things that felt human. The mouse allowed us to “touch” and the GUI allowed us to see. We could create and consume things that weren’t possible before with computers. The Macintosh became part of our creative toolset to express our humanity in digital form. Gone were the days that using computers was limited to those who could program. The Mac opened the computing world to anyone to create and share their message with the world.

On the 15th anniversary of the Macintosh, Steve Jobs wrote, “While I am not normally one to look back, today is a good day to remember Apple’s legacy, which is to bridge the gap between sophisticated technology and ‘the rest of us’ who make up most of humanity. It’s our job to make complex technology easy to use and fun to use.”

Over the past 40 years, computers have made many more things easy and fun. We listen to music, watch movies, communicate with anyone. Computers have become so small that we carry them in our pockets. And the initial “touch” of the mouse has now become a literal “touch” of the screen. But, despite these amazing changes, they have remained the same in one important way: they remain tools that we control. No matter how sophisticated or complicated the task, today’s computers are designed to follow our directions and, if they fail, the failure is considered a bug and sent back to engineering to fix. This is all changing.

I believe we are at the beginning of the next great shift in our relationships with machines. And I believe that this shift may have a more profound effect than any of those before for two reasons: computers have control of the mouse, and they are writing messages of their own.

One of our research obsessions is agentic AI—the idea that AI has agency to make decisions and perform tasks on its own. For the past 40 years, computers have been able to do remarkable things, but all of those things followed specific instructions from people. We are now heading into a world where AI may be given a broad objective and allowed to find its path. You might say, “I want to go on a vacation in New Zealand. Take care of it.” Or you might say, “My company needs to cut costs. Decide the optimal headcount reduction, which people to let go, and handle the process.” Or you might say, “We need to solve climate change. Figure it out.”

There is promising potential in using AI to solve these complex problems. But that potential requires handing the mouse to the machine itself. To tap the expansive intelligence of AI, we will have to relinquish control. How will we know if we are ready to hand over control of the mouse to the machine? For 40 years, we’ve decided what to click. In what situations are we ready to let the machine click itself?

The second major change is the ability for a machine to create content itself. For 40 years, we have used computers to communicate with each other, but the machines are now communicating themselves.

Marshall McLuhan coined the phrase “the medium is the message” to describe what he saw as the effect of technology on the messages we share. The printing press allowed for long-form text, while the telegram shortened every message. Radio introduced audio messages, while television encouraged entertaining messages—even in news broadcasts. This pattern has continued as the internet and social media shape our conversations, creating clickbait and extreme content to appeal to the pressures of the attention economy.

Generative AI introduces a profound shift as our machines are no longer just shaping the messages we share but are creating messages themselves. Human history has been dominated by an echo chamber of humanity talking with ourselves. Today, however, we are talking with machines—and they are talking back. Yes, they have learned language from us and their ideas are derived from our previous messages. But they are stringing together words and concepts in novel ways that are theirs alone. This means that computers are no longer an extension of our human expression—they are now something entirely different.

A few years before the launch of the Macintosh, Steve described the computer a 'bicycle for our minds,' the greatest tool ever created because it would increase the efficiency of our minds as a bicycle does our bodies. He proved to be correct, as the computer, led by the Macintosh, has dramatically increased human efficiency in changing the world around us—for both good and bad. We believe that AI changes the nature of computing, requiring a new metaphor because computers will no longer be human-directed tools like bicycles. Computers are no longer just connecting human minds, they are becoming minds themselves with the agency to make decisions without specific direction from a person. How those minds operate and for whom may be the most important question of our time. Today, we have a glimpse of what might come, and while the timelines are uncertain, the disruptive impact is not.

As the Macintosh made a huge leap following years of early PC developments, we bet that the Macintosh of the AI age has yet to be designed. Our vision for AI follows in the legacy of the Macintosh—technology that is equally science and art, humanities and engineering, and, above all, designed 'for the rest of us.' Our vision is to design AI that is a mind for our minds—an agentic intelligence that is in service to all of us.

This Week from Artificiality:

Whole-Mindedness

Current AI resembles left-brained reasoning - optimized, logical but decontextualized. Humans play the right-brained role anchored in real world connections.

Key Points:

• Our minds synthesize both left and right brained cognition into unified consciousness.
• Two divergent AI paths emerge—either keeping AI subservient and left-brained, or pursuing artificial general intelligence with balanced hemispheres.
• Pursuing AGI risks unintended consequences from uncontrolled synthetic intelligence. Contextual relevance may remain a human duty.
• Advancing language models offer a third path—enhancing holistic human cognition through closer human-AI integration versus external AI synthesis.
• Humans may prefer complementing innate socially-oriented abilities over replicating general intelligence artificially.
• AI designed to elevate collective wisdom over individual optimization aligns with humanity's values and evolution.

Pro Obsession: Agentic AI

Defined as “the degree to which a system can adaptably achieve complex goals in complex environments with limited direct supervision", agentic AI promises to transform how we make decisions.

Key Points:

• Agentic AI is defined as AI that can achieve complex goals in diverse environments with limited supervision. It promises to transform decision-making by acting as personal agents.
• We provide a framework for different levels of agentic AI, from basic assistance to highly adaptive systems that can respond to novel situations. Higher levels require greater goal complexity, environmental complexity, independence, and adaptability.
• We explore how an agentic AI could transform enterprise search and analysis by autonomously navigating systems, collating relevant data, and providing actionable insights. This demonstrates its potential to enhance productivity.
• Our research priorities around agentic AI include studying human-AI collaboration dynamics, oversight for accountability, user adoption patterns, and more.
• Designing fully agentic AI involves many challenges around coordinating sub-tasks, managing unpredictability, and emergent human-machine interactions. But with advances in generative AI, agentic systems will fundamentally reshape decision-making.

World of Workflows: The Economic Reality of AI Means Scale Matters

A recent study from MIT has been grabbing attention with its unconventional take: don't worry about AI snatching your job, it's not cost-effective.

Key Points:

• A recent MIT study found that only 23% of jobs with tasks that can be automated by AI are actually cost-effective to automate. The key factor is the scale required to make AI economically viable.
• For most companies, 77% of automatable tasks are not cost-effective to automate. Significant scale through market consolidation or AI-as-a-service platforms is required.
• Cost factors like engineering, data acquisition, and accuracy requirements determine if AI automation makes economic sense, more so than computational costs.
• For investors and business leaders, the value of AI multiplies as a platform.

Ed Sim: AI Venture Capital

An interview with Boldstart Founder, Ed Sim, about venture investing and AI.

Key Points:

• Ed Sim started Boldstart Ventures in 2010 to provide early stage funding for enterprise startups, writing smaller checks than typical VC firms. The firm now manages a nearly $200 million main fund and a $175 million opportunity fund.
• Generative AI is an exciting new technology, but the key is backing founders who are solving real problems for end users in a unique way that is 10x better than current solutions. AI is just the underlying technology.
• AI security is critical for enterprise adoption. Ed invested early in Protect AI, which helps monitor AI models for security, privacy, and compliance issues. AI security will be key to scale adoption.
• There are still open questions around data governance with large language models that access sensitive company data. Approaches that check governance policies before providing answers are the safest for now.
• Factors like inference cost, subscription fatigue, and proving ROI will impact how quickly some of the consumer generative AI applications gain traction. Creative solutions around caching, pricing models, and hybrid human+AI loops can help.
• There will be opportunities related to embedding expertise into systems to empower junior and senior employees. Tools like GitHub Copilot show potential to augment technical skills.

Read/Write: Steps of Prompting

Chain of thought, tree of thoughts, and now graph of thoughts—a progression that may lead to agentic AI.

Key Points:

• Researchers are exploring methods to enable large language models to display more human-like, emergent reasoning capabilities.
• Chain of Thought and Tree of Thoughts approaches improved reasoning but were still fundamentally linear.
• Human reasoning operates more like a graph, interconnecting thoughts in parallel, dynamic networks.
• Graph of Thoughts (GoT) models allow language models to reason in an interconnected, non-linear graph structure.
• In an initial test, GoT outperformed other techniques in a sorting task, increasing quality while reducing costs.
• For complex planning, GoT could enable simultaneously considering multiple options and constraints.
• This approach may allow more sophisticated reasoning and decision-making, closer to human cognition, though whether it enables emergent capabilities remains to be seen.