|
Hello, and thank you for reading the inaugural Control Plane newsletter.
AI has seen extremely rapid progress over the last few years, and through a confluence of personal and professional factors, I've been lucky enough to observe its advances from a front-row seat. I say "lucky" because I often feel like I have a better sense of the technological leaps and profound changes that are occurring in this space than most. It is very difficult to keep up with this field—not just for the average person, but even for professional journalists—and that means that most people don't realize the scale of the transformation that is underway.
That's why I want to spend some time in this first edition of Control Plane emphasizing the rates of progress. This publication is focused on AI security, infrastructure, and geopolitics, and all of those areas are very strongly impacted by advancements in the fundamental AI technologies.
The most obvious advancement is in what you might call "intelligence", or perhaps "capabilities". AI has gotten smarter. Very quickly. A good lens for this is the Human-Centered AI Institute at Stanford University's annual AI Index Report, which tracks various AI models against standard benchmarks like "MMMU (Massive Multitask Multimodal Understanding)"—a benchmark that assesses whether an AI can reason across text and images together, the way a college student would when working through diagrams, charts, and written questions—and "GPQA (Graduate-Level Google-Proof Q&A)", which includes questions written by PhD scientists in biology, chemistry, and physics that are designed so that smart non-experts cannot Google their way to the answer. There's also "SWE-bench", which tracks coding ability. And the numbers paint a pretty clear picture:
|
|
Beyond inference, algorithmic efficiency has also been improving, so that fewer of those increasingly cheap tokens are needed to perform complex tasks. An Epoch AI study that looked at AI systems designed to classify images found that they could reduce the compute needed by half roughly every nine months in order to accomplish certain tasks.
Zooming out, it's very important, I think, to understand that AI progress has been exponential. And if you chart it out on a log scale, it looks like a straight line—steady yet radical progress, pointing to a future that is astonishing and stupefying.
|
|
We haven't yet seen any solid reason to doubt that this line will keep going as compute scales up and R&D continues. It could even accelerate now that AI is starting to train itself in a recursive loop. I think it's likely that it will.
Why am I spending so much time emphasizing these rates of progress? Because there seems to be a significant disconnect between the perceptions of everyday, casual users and those who are very active in the AI space. And that totally makes sense—the frontier capabilities very rarely intersect with people organically in everyday life. Many people use ChatGPT without noticing which underlying model it's running on. The latter can actually make a huge difference in the quality of its answers to certain kinds of questions, but if someone is just using it as a Google search replacement, asking things like, "Who were the Luddites?" or "What is recursion?", then they're not really going to notice. The worst version of ChatGPT can very competently answer those questions. But if you ask ChatGPT, "Can you reorganize all the files on my computer and delete things that seem like malware?" it can't do that. Other products from OpenAI, such as Codex, can. But everyday users aren't playing around with Codex because it assumes a certain familiarity with coding and it isn't marketed to everyday users.
Similarly, a lot of white collar workers are getting their first introduction to enterprise-ready AI tools through Microsoft Copilot, which a lot of people agree is not very good, at least with respect to the true frontier models. It just happens to be the case that Microsoft has a sprawling network of preexisting business partnerships with large enterprises, and therefore has gotten itself in pole position for making its AI tools available. And, naturally, workers come away unimpressed with AI.
Meanwhile, the frontier labs are racing ahead, and a comparatively small group of enthusiasts are watching. Among those enthusiasts are AI professionals, early adopters, theoretical physicists, finance and intelligence professionals, and hackers. And every day, they are all trying to figure out where the frontier is, and how to leverage state-of-the-art AI for their ends. That has profound implications for how the world is going to transform in the near, medium, and long term.
The transformation will touch almost every aspect of society. At Control Plane, we'll be watching what happens in a few key areas, all related to security.
One of these areas is infrastructure: As the need for ever-more compute goes up, there is a corresponding need to secure massive infrastructure, such as data centers and chip fabrication facilities. And as state power increasingly looks to leverage AI, that infrastructure will be considered critical infrastructure, and it will be targeted by adversaries.
That's where geopolitics comes in. At the time of writing, a major conflict is underway centered on Iran, and it has already led to serious damage to AI data centers in nearby Bahrain and the UAE. This is also probably making a lot of people anxious about Taiwan: The Taiwan Semiconductor Manufacturing Company makes about 90 percent of the world's most advanced AI chips, and it's a metaphorical stone's throw from China, whose authoritarian government has never dropped its claim to the island, and never renounced the idea of reunification through force.
Meanwhile, the AI tools themselves are already being exploited by hackers and fraudsters to infiltrate governments, businesses, and all kinds of other organizations and access sensitive assets. They're doing that through deepfakes, malicious code, semi-autonomous agents, and sundry other tricks and conjurings. This is the security area I know best, having spent most of my journalistic career studying the identity security space, and so I'll be spending a lot of time here from the beginning.
But it's all connected, and every node in this matrix is moving quickly. Stay tuned to Control Plane, and we'll do our best to keep you up to speed.
|
